Students’ reality

Adoption is outpacing guidance, and access is unequal

According to the ICT in Education 2024 survey, run by Brazil's reference centre for ICT statistics, among secondary students who use the internet, seven in ten report using generative AI for schoolwork. Yet only 32% say a teacher has talked to them about how to use it, and across all grades that share falls to 19%. Guidance on judging AI outputs is just as limited. Only 33% of students say a teacher has shown them how to spot errors and bias in what it produces. In terms of access, the gains have been uneven. School connectivity has improved markedly, with internet access now reaching 96% of schools overall, and rural schools rising from 52% to 89% since 2020. But internet at the school is not the same as devices in students' hands, and that is where inequality persists. Across all schools, only 59% have both internet access and a computer for students to use, a share that falls to 28% in rural schools and 47% in municipal networks.

Teachers’ reality

A requirement not yet matched by a capacity programme

A curriculum can be introduced quickly, but the capacity to teach it develops more slowly, and the current instrument design requires the first without supporting the second. The sector agrees on where the gap is. A sectoral study on AI in education asked stakeholders across Brazilian education to name the main barriers to adopting AI, the two most cited were inequalities in access to technological infrastructure and teacher training. Teachers are already using the technology, but that use is uneven across the profession. Around 43% reported using generative AI to prepare teaching materials. The study shows that use is more common in private schools than public ones (51% against 41%), among younger teachers (55% of those up to 30, against 37% of those over 45) and in urban areas (45% against 31% in rural areas). It is also more common among men, at 50% against 40% of women, even though women make up 79% of basic-education teachers.

Overall, what teachers lack is structured support and it is worth recognising what already exists. Brazil has a National Digital Education Policy that includes digital capacity-building among its aims, and the Ministry of Education (MEC) has issued guidance for AI in schools with domains for teacher training, alongside national continued-training courses. These are valuable foundations. What is missing here is

the means to make the requirement deliverable, since none of these is funded against the new requirement, made a condition of it, or designed to reach the least-resourced networks first. As a result, much of the responsibility is falling on teachers and school managers who have not been resourced for it. Teacher preparation is weakening where it is most needed, with teachers' participation in training on digital technologies falling from 65% (2021) to 54% (2024) and the lack of it a barrier for 86% of municipal teachers against 59% of private ones. It is into this uneven and under-supported landscape that the AI literacy requirement is being introduced.

Governance instrument’s track record

A lever that tends to exclude by capacity, not by choice

When the VAAR funding was first distributed, in 2023, 66% of municipalities and 12 states (almost half) received nothing. The research behind these figures attributes this largely to gaps in technical, administrative, and political capacity rather than to choice. Those exclusions turned on conditions unrelated to computing or AI literacy, such as reducing measured learning inequalities and formalising the intergovernmental arrangements the transfer requires, which the under-resourced networks were the least equipped to meet. The point that carries across is structural. The VAAR operates as a single gate that a network passes only by meeting all its conditions at once, and the AI literacy requirement now forms part of that gate.

None of this means the design lacks a rationale, since a binding funding condition may be one of the few levers able to move a decentralised system, and without it AI literacy could remain a low priority precisely in the networks that can least spare attention for it. That case has merit, and the disagreement here is narrower than it first appears, since it concerns the choice and implementation of the lever rather than the use of one.

Governance instrument’s current design

Tying AI literacy to a results-based, capacity-rewarding lever

Most of Fundeb's resources are distributed by need, sending more funding to the less-resourced networks. The VAAR is the part of the fund that rewards results instead, so it does less to reduce inequality and tends to favour networks that already have the capacity to comply, as the 2023 cycle showed. The use of this instrument for AI literacy is very recent. Resolution CIF No. 15 of June 2025 tied

the BNCC Computing complement to the conditions a network must report against to receive the VAAR top-up. Attaching it to the part of the fund that does least for the less-resourced, rather than to the need-based transfers that already reach them, risks deepening the exclusion rather than closing it.

The new requirement is being introduced gradually, since a network that has not yet embedded the complement is not disqualified in 2026 but is expected to align in the cycles that follow. That said, the question that matters here is whether the least-resourced networks will have the means to clear the bar when that grace period closes. Some states are not waiting. The state of Piauí, for instance, has made AI a mandatory subject, now taught across 540 public schools and reaching over 90,000 students a year, work that UNESCO has recognised with an international award. A case like this suggests the capacity can be built at scale, and that the obstacle may lie less in whether AI can be taught than in whether the federal instrument is designed to develop that capacity rather than assume it.

Recommendations

Every proposal here shares one aim, to make the instrument build the capacity it currently assumes rather than only require it, and so to reduce, rather than amplify, the existing inequalities.

Pachaly, L. (2026). Compliance Without Capacity [Unpublished policy brief].
University of Cambridge.

1. Introduction

The policy brief is addressed to Brazil's Ministry of Education and its National Education Council. It examines a recent decision in which the federal government made adopting the computing complement to the national curriculum — which carries the school system's AI-literacy competences — a condition for a results-based top-up within its main school-funding system. Its concern is that the condition applies everywhere while the means to meet it exist in only some places. The recommendations share one aim, to strengthen the instrument so it builds the capacity that implementing the complement requires rather than assume it already exists.

Where the brief argues over the design of the chosen lever, this commentary steps back to the governance questions behind how a state governs AI literacy at all, several of which stand out. Who decides what AI literacy is, and how it is put into effect? By what rules, and through which instrument, should it be governed? How should the literacy itself be measured? Who is accountable for building the capacity it requires? And what kind of citizen does it form? This commentary sets out to examine two of these, on accountability and citizenship, concepts Brazil's own framework for AI in education places at its centre, where it frames educating for AI as “intellectual autonomy, critical discernment, and the full exercise of democratic citizenship”, and warns that, unless disparities are addressed, technology tends to deepen inequality rather than reduce it (Brazil. Ministry of Education, 2026, p. 9). How far a compliance instrument can deliver on that purpose is not obvious, and it is that distance the commentary explores.

These questions are examined through the literature on data governance and data justice. This choice is deliberate, and less a matter of preference than of what the subject is made of. For Long and Magerko (2020), AI literacy is a set of competencies that enables individuals to critically evaluate, collaborate with, and use AI systems. Built as they are by learning from data, what a model can do, and to whom, is in large part set by the data it is trained on (Jordan and Mitchell, 2015) and the relations that data encodes (Viljoen, 2021, pp. 577–578). Read together, these suggest that a literacy about AI is, above all, a literacy about data and the systems built on it. The same may hold one level up, in governance, since a growing body of work argues that governing AI is largely a matter of governing that data and the relations it produces (Taylor et

al., 2026). If that is right, the broader question this commentary asks, how a state governs the AI literacy its citizens receive, begins to look like a data-governance question at root, and the mature field of data governance and data justice becomes a natural place to look for the concepts an emerging field still lacks, since AI literacy and governance are only beginning to be established as foundations for ethical AI (Karimov and Saarela, 2025). This connection is deliberately compressed, offered as the commentary's analytical lens rather than the only account of how these fields relate.

The sections that follow situate the case within the broader debates it touches, then examine those two questions more closely and consider the limits and challenges of governing AI literacy.

2. Placing the case within broader AI-governance debates

This section places the Brazilian case within the wider debates in AI governance. It offers not an exhaustive map but a first reading of what the lens of data governance can illuminate about governing AI literacy.

The first of these is data justice, which asks not only whether a system works but whom it makes visible and whom it leaves out (Taylor, 2017), and which treats the protection of public infrastructure and public goods as a test of sound governance (Taylor et al., 2026). On that test it is worth asking whether a literacy that presumes connectivity, devices, and prepared teachers distributes a public good evenly or merely tracks one already there, and whether the same instrument might advance literacy in aggregate while widening the gap beneath the average.

The second is the geography of AI. Lehdonvirta et al. (2024) divide the world by access to the compute AI is built on, into a Compute North that hosts the infrastructure to build it, a Compute South with far less, and a Compute Desert with none. The divide, they argue, is not only material but political, since states without compute are less able to set the terms and

standards under which AI operates. Brazil sits on the southern side of that line and adopts a framework written in good part on the northern one, from UNESCO to the OECD (Brazil.

Ministry of Education, 2026, p. 117), which the decolonial literature reads as a familiar pattern, the export of one way of knowing as though it were universal (Ricaurte, 2019). One need not go that far to make the modest point that a literacy imported into an under-resourced system, and taught to students who will mostly meet AI built elsewhere, tends to equip them to use such systems while leaving the relations of dependence around them unexamined.

The third is the question of power itself. Read through power, the case raises several questions — who defines the literacy, who benefits, and who bears the cost of its absence. The one this commentary takes as its example, following (Kalluri, 2020), is not whether an instrument is fair but whether it redistributes power or leaves it in place, since a system can be fair and still leave power exactly where it was. A curriculum that teaches students to use AI without the capacity to question it does little to shift power, while one that builds the capacity to contest can begin to change it. This is where Brazil's own framing earns some credit. By placing AI literacy under digital culture rather than narrow technical skills, the curriculum leaves room, in principle, for the broader and more civic conception the critical model calls for (CNE/CEB, 2022b).

Whether that broader conception is realised, though, turns on a fourth and more practical debate, about assessment. The funding condition pays out only once a network can show it has adopted the curriculum, so a civic competence has to be turned into a checkable signal, and in Brazil that signal is the information the network files rather than a window onto the classroom (CIF, 2025). As Ananny and Crawford (2018) argue, seeing a record is not the same as knowing what a system does, and the broader the conception of literacy, the harder it is to verify, which creates a quiet pressure to narrow it back to the checkable, and so to certify the user while believing one is forming the citizen.

Taken together, these debates begin to reframe AI literacy from a curricular question into a governance one. Against that backdrop, the sections that follow take up two governance questions more closely.

3. What kind of citizen the arrangement forms?

As Carmi et al. (2020) argue, literacy is never only a set of skills but a social practice shaped by context and power. They turn that insight into a distinction between a neutral, universal view of literacy — a set of individual skills that can be rolled out everywhere in the same form — and a critical model they call data citizenship, oriented toward enabling people to question and act on the systems around them. Their framework sets out three dimensions. Data thinking is the critical understanding of how data and the systems around it work. Data doing is the everyday practice of using and managing data. And data participation is the proactive, collective engagement through which people act with others on the systems that affect them. The three give a test to run any curriculum through, whether it builds understanding and everyday practice and also the capacity to act with others, or stops at the first two.

Data citizenship does not reject competences but builds on them. The mainstream definitions supply the baseline, treating AI literacy as a set of competencies everyone needs rather than a specialist skill (Long and Magerko, 2020). A critical data literacy starts from that baseline and reaches further, toward the capacity to evaluate data systems and their social effects (Pangrazio and Selwyn, 2019). The question is which of the two, the functional competence or the critical capacity, a compliance instrument tends to reward. Read against the critical model, Brazil’s curriculum is not empty. Its computing component is organised around computational thinking, the digital world, and digital culture, with data literacy and the ethics of use running through it (CNE/CEB, 2022a), so it speaks, at least on paper, to data thinking and data doing, and gestures at participation through its language of digital citizenship.

Where it falls short is in what Carmi et al. (2020) call networked and contextual literacies. For them a literacy cannot be delivered as a property of the individual learner in isolation, because people engage with data through the people and places around them, so they situate it in citizens’ networks of literacy — their families, communities, and neighbourhoods — and reject the idea that one common programme can serve everyone, since people with different backgrounds need different literacy programmes. A curriculum applied in the same form across the country, and certified the same way everywhere, is the universal model they warn against,

and it is the contextual and networked dimension, not the list of competences, that it leaves underdeveloped.

There is a further reason an individual-skills model misfits its subject. Viljoen, (2021, pp. 573–574, 577–578) argues that data is better understood as a social relation than an individualist claim about a single person, since data drawn from one person routinely supports inferences about many others, so its value and harms are produced at the level of populations rather than individuals. A literacy assembled as a personal checklist tends to speak only to the part of the problem visible to the individual, so the networked and collective dimension is not an optional add-on but the part that makes an AI literacy relevant.

In Brazil, the contextual and the collective become a single problem, because the networks and contexts Carmi and colleagues describe are also the populations Viljoen points to — a landscape so varied and unequal that mapping it would be a study in itself. To give one example, access to devices and connectivity is deeply uneven, so the same lesson presumes resources many schools lack. Only 28% of rural schools have computers and internet for students, against 76% of urban ones, and 29% of students in the North reach the internet at school through an institution's computer against 87% in the South (CETIC.br, 2025, pp. 21, 23). A literacy designed for a generic classroom is likely to form one kind of citizen in the school it imagines and another in the schools many students actually attend. Reduced to universal skills, it tends to form the thinner of the two, the digital resignation Carmi and Nakou (2025) warn a functional literacy can produce.

None of this is inevitable and an example of a critical literacy model exists inside the same system. Piauí, a north-eastern state of Brazil, made AI a subject of its own curriculum, now taught across 540 public schools to around 90,000 students a year with over 680 teachers trained to deliver it (UNESCO, 2025). The programme was designed so that teachers grasp AI critically rather than only operationally, and its most cited classroom result — students building an AI tool to catalogue and distribute seeds to local family farmers — is data citizenship in practice, designed from the local context outward rather than imported into it (Porvir, 2025). Set against a nationally uniform requirement, it shows the difference between forming users

and forming citizens, and that the critical model is achievable within the same system, provided the instrument is built to support it.

4. Who is accountable for building the capacity?

Whether a literacy forms citizens or users depends, in good part, on the capacity to teach it, and that returns the analysis to a second governance question, who answers for building it. In Brazil the requirement, as it stands, assumes a capacity to teach that many networks do not yet have. For AI literacy that capacity is not only infrastructural but human and institutional. This is the harder kind to build, and the kind Brazil can least take for granted, because the requirement is layered onto a system still working to secure foundational learning. In PISA 2022, Brazilian fifteen-year-olds scored well below the OECD average across mathematics, reading, and science, with around half below the baseline level of proficiency in reading and roughly three-quarters in mathematics (OECD, 2023, Figures I.3.1 and I.3.4). The workforce that would carry a new and demanding subject is itself stretched. In the final years of primary school many classes are taught by teachers not qualified in the subject, and in much of the North and Northeast fewer than half are (INEP, 2025, Figure 1). Preparation for the digital part of the job is thin and, in places, thinning, as teachers' participation in ongoing professional development on digital technologies fell from 65% in 2021 to 54% in 2024, with the lack of it cited as a barrier by 86% of municipal-school teachers (CETIC.br, 2025, p. 91). This is also where the accountability the instrument sets up comes under strain, since it holds each network answerable for the literacy, and liable to consequences (Bovens, 2007), while the capacity most decisive for it, a prepared teaching workforce, is one many networks simply do not have. To require results before that capacity is in place is to hold a network to account for what it lacks the means to do.

In that setting the capacity to teach AI cannot be assumed, since in many networks it would have to be built from a low base, and this is why teachers are a decisive variable. As the literature on policy enactment argues, teachers do not simply implement a written curriculum but interpret and translate it into what students actually receive (Braun et al., 2011). So where they are unprepared even a well-designed curriculum tends to collapse into its most checkable

form, the operating skills that can be demonstrated rather than the critical capacities that cannot. Reading through the citizenship lens, the implication is sharper still, because a teacher who can only operate AI is unlikely to form students who can question it, which makes the critical, networked literacy depend first on the formation of the people expected to teach it. (Karimov and Saarela, 2025) place professional development among the main channels through which governments build AI literacy, and note that uneven resources and teacher training can hinder adoption, as they observe for the EU.

If that is right, teacher formation is less one investment among several than the step zero on which the rest depends. The thing a capacity-first instrument would fund, and reach the least-resourced networks with, before it asks them to comply. That opens questions this commentary can only raise but not settle. Where should responsibility for forming a teaching workforce for AI sit? With the federal level that sets the standard, the states and municipalities that employ the teachers, or some shared arrangement? And what division of that labour would reach the networks furthest behind? It may be unfair to expect one funding rule to carry all of this. What the Brazilian case points to instead is that a combination of instruments and policies — shared but adaptable curricula, partnerships with universities and support bodies, and the AI-literacy centres some systems are building — is more likely to have lasting effect when the governance roles behind them are made explicit.

5. The limits and challenges of governing AI Literacy

Governing AI literacy in Brazil, a continental country still marked by deep structural gaps in its population's foundational literacy, is a complex governance challenge in itself. This section brings some of that complexity into view by holding the brief's own recommendations up to scrutiny.

The case for the instrument's original design is not weak. In a decentralised system, a binding funding condition may be one of the few levers capable of steering the networks toward a shared goal, and without it AI literacy could remain a marginal concern precisely in the places least able to prioritise it. The brief's critique — that this lever rewards capacity rather than

builds it and so falls hardest on those who lack the means — is persuasive, but its own fixes, despite protecting something real, carry tensions of their own. Turning the condition into a capacity ladder does not remove administrative discretion so much as relocate it, since someone must still set each network's baseline and timeline, and those judgements reopen the influence of capacity, politics, and measurement. Allowing networks to meet the requirement through device-free, unplugged AI literacy protects access where infrastructure is missing, yet, if teaching about AI on paper were treated as equivalent to teaching with it, it would entrench the very divide it addresses and ease the pressure to close the infrastructure gap. Funding capacity before requiring it, on its own, does not settle the matter, as what most decides whether the capacity is actually built — from a prepared teaching workforce to the institutional arrangements for delivering it — reaches beyond financial investment.

Beneath the recommendations also lies a tension between the common national standard Brazil has adopted and the contextual literacy set out earlier. One of the strongest cases for the standard is that a floor guarantees every network owes its students at least a baseline, which in a deeply unequal system is itself a form of equity. The point, though, is not to reject the floor but to deny that it is the ceiling. A baseline of functional competence is a precondition for the critical kind, not a substitute for it (Pangrazio and Selwyn, 2019). An instrument that can only see the floor certifies compliance without guaranteeing the critical capacity it cannot observe (Ananny and Crawford, 2018). Even a fully funded capacity, then, is not the same as citizenship, since a network could meet every requirement and still form users rather than citizens (Carmi et al., 2020). No funding design resolves that trade-off on its own, and it is perhaps at the point where the two questions meet, who builds the capacity and what citizen the literacy forms, that the future of this instrument might most usefully be shaped.

6. Conclusion

Brazil's decision to tie AI literacy to a results-based funding condition is sound in its ambition, and the question it raises is not whether AI literacy should be part of the school curriculum, which is already settled, but how the state governs it and to what end. Seen through

established theories of data governance, several questions stand out, reframing AI literacy from a curricular matter into a governance one. This commentary examines two of the governance questions Brazil's own framework places at its centre — what kind of citizen the arrangement forms and who is accountable for building the capacity it requires — and steps back to weigh the brief's own recommendations.

What this commentary reveals is that AI literacy cannot be governed as curriculum or compliance alone — a task harder still where the ground beneath it is uneven, in systems still working to secure the foundational learning. By bringing the more developed questions and theories of data governance to bear, this commentary aims to give this emerging field a vocabulary and a set of governance questions to prioritise. The questions examined here, and the others left open, are less a conclusion than a starting point. An attempt to equip researchers and policymakers with considerations grounded in a real and unfolding case, in the hope that the angles left unexplored serve as a springboard for further research and public debate on how societies choose to govern the AI literacies their citizens receive.

Bibliography

Ananny, M., and Crawford, K. (2018). Seeing without knowing: Limitations of the transparency

ideal and its application to algorithmic accountability. New Media and Society, 20(3),

973–989. https://doi.org/10.1177/1461444816676645

Bovens, M. (2007). Analysing and Assessing Accountability: A Conceptual Framework1. European

Law Journal, 13(4), 447–468. https://doi.org/10.1111/j.1468-0386.2007.00378.x

Braun, A., Ball, S. J., Maguire, M., and Hoskins, K. (2011). Taking context seriously: Towards

explaining policy enactments in the secondary school. Discourse: Studies in the Cultural

Politics of Education, 32(4), 585–596. https://doi.org/10.1080/01596306.2011.601555

Brazil. Ministry of Education. (2026). Framework for the Development and Responsible Use of

Artificial Intelligence in Education [Referencial para o Desenvolvimento e Uso

Responsável da Inteligência Artificial na Educação]. Ministry of Education (Brazil).

https://www.gov.br/mec/pt-br/media/segape/referencial-oficial-en.pdf

Carmi, E., and Nakou, P. (2025). Building data citizenship and learning to resist in the datafied

society. Learning, Media and Technology, 1–14.

https://doi.org/10.1080/17439884.2025.2505553

Carmi, E., Yates, S. J., Lockley, E., and Pawluczuk, A. (2020). Data citizenship: Rethinking data

literacy in the age of disinformation, misinformation, and malinformation. Internet Policy

Review, 9(2). https://doi.org/10.14763/2020.2.1481

CETIC.br. (2025). Survey on the use of information and communication technologies in Brazilian

schools: ICT in Education 2024. Brazilian Internet Steering Committee (CGI.br).

https://www.cetic.br/media/docs/publicacoes/2/en-us/20251217163959/tic_educacao

_2024_livro_completo_ingles.pdf

CIF. (2025). Resolução CIF no 15, de 12 de junho de 2025 [Resolution CIF No. 15 of 12 June 2025:

Methodology for assessing the management-improvement conditionalities for

distributing the VAAR per-student top-up in 2026] [Resolution]. Ministério da Educação,

Secretaria de Educação Básica.

https://www.gov.br/fnde/pt-br/acesso-a-informacao/acoes-e-programas/financiamento

/fundeb/legislacao/2025/resolucao-cif-15-2025-fundeb-condicionalidades-vaar-2026.pdf

/view

CNE/CEB. (2022a). Computação: Complemento à Base Nacional Comum Curricular (BNCC)

[Computing: Complement to the National Common Curricular Base (BNCC)]. Ministério

da Educação, Conselho Nacional de Educação.

https://www.gov.br/mec/pt-br/escolas-conectadas/arquivos/BNCCComputaoCompletod

iagramado.pdf

CNE/CEB. (2022b). Parecer CNE/CEB no 2/2022: Normas sobre Computação na Educação

Básica—Complemento à Base Nacional Comum Curricular (BNCC) [Opinion No. 2/2022:

Standards on computing in basic education—Complement to the National Common

Curricular Base (BNCC)] [Opinion]. Ministério da Educação.

https://www.gov.br/mec/pt-br/cne/pdf/pareceres-do-cne/ceb/2022/pceb002_22.pdf

INEP. (2025). Censo Escolar da Educação Básica 2024: Notas Estatísticas [Basic Education School

Census 2024: Statistical Notes] [Statistical Notes]. Instituto Nacional de Estudos e

Pesquisas Educacionais Anísio Teixeira (INEP).

https://download.inep.gov.br/publicacoes/institucionais/estatisticas_e_indicadores/not

as_estatisticas_censo_da_educacao_basica_2024.pdf

Jordan, M. I., and Mitchell, T. M. (2015). Machine learning: Trends, perspectives, and prospects.

Science, 349(6245), 255–260. https://doi.org/10.1126/science.aaa8415

Kalluri, P. (2020). Don’t ask if artificial intelligence is good or fair, ask how it shifts power. Nature,

583(7815), 169–169. https://doi.org/10.1038/d41586-020-02003-2

Karimov, A., and Saarela, M. (2025). AI Literacy and Governance as Foundations for Ethical AI: A

Cross-National Review of Government Strategies. 2025 IEEE Symposium on Trustworthy,

Explainable and Responsible Computational Intelligence (CITREx), 1–7.

https://doi.org/10.1109/CITREx64975.2025.10974932

Lehdonvirta, V., Wú, B., and Hawkins, Z. (2024). Compute North vs. Compute South: The Uneven

Possibilities of Compute-based AI Governance Around the Globe.

Long, D., and Magerko, B. (2020). What is AI Literacy? Competencies and Design Considerations.

Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1–16.

https://doi.org/10.1145/3313831.3376727

OECD. (2023). PISA 2022 Results (Volume I): The State of Learning and Equity in Education. OECD

Publishing. https://doi.org/10.1787/53f23881-en

Pangrazio, L., and Selwyn, N. (2019). ‘Personal data literacies’: A critical literacies approach to

enhancing understandings of personal digital data. New Media and Society, 21(2),

419–437. https://doi.org/10.1177/1461444818799523

Porvir. (2025). Programa de ensino de IA nas escolas do Piauí ganha prêmio da Unesco [AI

teaching programme in Piauí schools wins the UNESCO prize]. PORVIR.

https://porvir.org/piaui-ia-escolas-premio-unesco/

Ricaurte, P. (2019). Data Epistemologies, The Coloniality of Power, and Resistance. Television and

New Media, 20(4), 350–365. https://doi.org/10.1177/1527476419831640

Taylor, L. (2017). What is data justice? The case for connecting digital rights and freedoms

globally. Big Data and Society, 4(2), 205395171773633.

https://doi.org/10.1177/2053951717736335

Taylor, L., De Souza, S. P., Martin, A., and López Solano, J. (2026). Governing artificial intelligence

means governing data: (Re)setting the agenda for data justice. Dialogues on Digital

Society, 2(1), 5–22. https://doi.org/10.1177/29768640241306800

UNESCO. (2025). UNESCO recognizes four initiatives promoting the responsible use of AI in

education.

https://www.unesco.org/en/articles/unesco-recognizes-four-initiatives-promoting-respo

nsible-use-ai-education

Viljoen, S. (2021). A Relational Theory of Data Governance. The Yale Law Journal, 131(2),

573–654.

Pachaly, L. (2026). Governing AI Literacy through a Data-Governance Lens
[Unpublished critical commentary]. University of Cambridge.

Abstract

This paper argues that, in the specific context of human–AI assistant relationships — where AI systems act on an individual's behalf across domains and over time — autonomy is the foundational ethical principle. Not because other principles are unimportant, but because autonomy functions as the precondition that gives the others their practical content — a user who cannot meaningfully exercise agency over their assistant is poorly positioned to verify its transparency, resist its manipulative potential, or contest the discriminatory patterns embedded in how it acts on their behalf. Taking autonomy as foundational, the paper's primary aim is to examine what this requires in practice. The argument develops along two mutually reinforcing lines. The first concerns a fundamental shift in user mindset — users have become structurally habituated to passive delegation of informational agency, a posture profoundly inadequate for systems that act on their behalf across domains and over time. The second concerns the obligations of those who build these systems — no degree of individual effort can substitute for a commitment to agency-respecting design when systems are architecturally opaque and their complexity is asymmetrically distributed. The paper examines the OpenClaw platform as an illustrative case, arguing that the conditions for meaningful human autonomy over an AI assistant can be built rather than merely argued for, and that what is required to do so reflects architectural choices, not technical constraints.

1. Introduction

For scholars of AI ethics such as Whittlestone et al. (2019), the question of what counts as the most important ethical principle for AI does not have a context-independent answer. In this view, the ethics of AI weapon systems demand different analytical priorities from the ethics of AI in medical diagnosis or in content moderation, for example. This paper takes this as its starting point and deliberately confines its focus to the ethics of human–AI Assistant relationships — the emerging set of interactions in which an AI system, operating across platforms and domains, acts

on an individual's behalf to manage information, make recommendations, plan actions, and execute them (Gabriel et al., 2024, p. 3 ; Manzini et al., 2024).

Within this frame, this paper argues that autonomy is the foundational principle in human-AI assistant relationships. Not because the other principles are unimportant, but because autonomy functions as the precondition that gives the others their practical content — a user who cannot meaningfully exercise agency over their assistant is poorly positioned to verify its transparency, resist its manipulative potential, or contest the discriminatory patterns embedded in how it acts on their behalf.

Taking autonomy as the foundational principle, this paper's primary aim is to examine what it requires in practice. The argument develops along two mutually reinforcing lines. The first concerns a fundamental shift in user mindset. Users have become accustomed to delegating their informational agency over digital systems to technology organisations in exchange for access to free services (Obar & Oeldorf-Hirsch, 2020). This paper argues that, if users do not wish to extend the same passive delegation of their autonomy to AI systems, they will need to be considerably more diligent about their agency over these systems.

The second line of argument is built on the premise that the user-side effort cannot succeed without a parallel commitment from developers. When systems are architecturally opaque and their complexity is asymmetrically distributed (Ienca, 2023), no degree of individual effort can substitute for developer commitment to agency-respecting design. This paper argues that developers bear an ethical obligation to make human agency intuitive and non-burdensome by design and that these obligations cannot be delegated to users.

The paper develops these claims across four sections. Section 2 sets the argument and scope, developing the case for autonomy as the foundational principle and identifying the gap this paper addresses. Section 3 examines the shift in user mindset that the evolving human-AI Assistant relationship requires. Section 4 develops the case for developer design obligations. Section 5 examines the OpenClaw framework as an illustrative architecture to explore what operationalising this principle can look like in practice. The paper concludes by reflecting on what both lines of argument suggest about the conditions under which human autonomy can be genuinely exercised within the human-AI assistant relationship.

2. Setting the Argument and Scope

Drawing on Gabriel et al. (2024), this paper defines AI Assistants as "artificial agents with natural language interfaces, whose function is to plan and execute sequences of actions on behalf of a user — across one or more domains — in line with the user's expectations" (p. 3). These systems are currently most commonly available as chatbots or fine-tuned language models accessible via conversational interfaces (p. 94). While this is still a nascent market, AI assistants are becoming increasingly ubiquitous and permeating a wide range of domains in users' lives, emerging as “the principal medium through which online information exchange occurs” (p. 25).

This paper focuses on AI assistants with general capabilities — systems designed to operate across multiple domains on a user's behalf (Gabriel et al., 2024, p. 3) — and specifically on their relationship with human users. While acknowledging that users may develop relationships with any form of conversational assistant (Manzini et al., 2024), it does not explore AI assistants specifically designed to address social needs such as companionship, romance, or friendship (Shevlin, 2025), though it recognises that some findings here may be relevant to those contexts. It also does not aim to provide a comprehensive account of the harms that could result from this relationship, though it acknowledges this is a significant area of inquiry.

Within this frame, several ethical principles present compelling claims to foundational status. Jobin et al. (2019, p. 395) document their convergence across AI ethics governance globally, identifying a set of concerns that appears with remarkable consistency across the policy documents, governance frameworks, and academic proposals of the past decade. Prominent among them, transparency demands that AI systems be explainable and their workings — including how decisions are made and whose interests are served — made legible to those they affect. Justice and fairness require that the benefits and burdens of AI be equitably distributed and that systems do not perpetuate discrimination or deny access to resources. Non-maleficence requires that the harms these systems can produce — through manipulation, data misuse, and safety failures — be actively prevented. Responsibility demands that those who develop and deploy AI systems be held accountable for their impacts, with clear lines of liability. Privacy requires that the intimate and extensive personal data these systems access be protected from misuse or unwarranted disclosure.

What distinguishes this paper's arguments from the broader AI ethics literature is not a rejection of these principles but a claim about their relative weight in the context of human-AI assistant relationships. A user who does not exercise meaningful autonomy over their assistant is poorly positioned to verify whether the transparency it claims to offer reflects their interests rather than those of third parties (Lazar, 2024), to identify and protect themselves from the manipulative harms that non-maleficence demands be prevented (Ienca, 2023), or to contest the discriminatory patterns embedded in how it interprets and acts on their preferences that justice and fairness require be remedied (Kay et al., 2024). In this sense, autonomy is not simply one value among equals — it is the baseline condition that makes the others actionable.

Three foundational papers define the landscape into which this paper intervenes. Gabriel et al. (2024), in a comprehensive normative framework, maps the ethical landscape of AI assistants — their design choices, capabilities, and governance challenges. While it addresses risks of manipulation, data misuse, and autonomy-undermining design at the system level with considerable depth, it does not substantively address how individual users might or should actively exercise agency over their assistants. Manzini et al. (2024), examining conditions for appropriate human-AI assistant relationships, introduce the critical distinction between revealed preferences — inferred from behavioural traces — and ideal preferences that users would endorse under conditions of full information and reflection. They call for developer interventions that empower users toward meaningful control, but do not prescribe the parallel shift required of users themselves. Shavit et al. (2023), addressing the governance of agentic AI systems operating with limited supervision, frame human agency primarily as a governance output — a property for systems to preserve — rather than also as a capacity that users themselves actively cultivate.

None of these papers addresses users as active agents in their own relationship with AI, and while each acknowledges developer obligations in relation to user agency, none examines in detail what making autonomy genuinely exercisable in practice would require — of users or of those who build these systems. This paper aims to fill that gap.

Finally, as used here, the term ‘agency’ refers to the capacity of individuals to act with autonomy, choice, and meaningful control over their AI Assistant systems. The following section argues that realising this conception of agency in practice requires a fundamental shift in how users themselves approach their relationship with this technology.

3. Cultivating Agency: The User's Responsibility

Users of digital services have been systematically habituated to treat the boundaries of their agency as coextensive with the boundaries of the terms of service — accepting, without necessarily reading, the conditions under which their data is collected, processed, and monetised (Obar & Oeldorf-Hirsch, 2020). This passive posture is not irrational, given the asymmetric bargaining power of individuals relative to platforms, but it is profoundly inadequate for the human-AI Assistant context, where the system learning from a user's data is also the system acting on that user's behalf.

User disengagement from the conditions governing digital services is well-documented. Obar & Oeldorf-Hirsch (2020) found that 98% of users skipped reading privacy policies altogether in their study of social media sign-up practices. Gabriel et al. (2024, p. 115) observe that users similarly fail to read terms of service for AI assistants, frequently accepting default options designed to maximise data collection. The cumulative effect is a user population that has internalised passivity as the natural stance toward digital services — a stance that now extends, unreflectively, to systems of a categorically different nature.

AI assistants differ from prior digital services in, at least, three morally significant respects. First, unlike passive platforms that respond to user-initiated requests, they are agentic systems that, with increasing autonomy, anticipate needs, initiate actions, and make decisions on user's behalf across multiple domains simultaneously, with limited direct supervision (Gabriel et al., 2024, p. 31; Shavit et al., 2023). The scope for autonomous deviation from user interests is therefore qualitatively wider than in any prior digital context. Second, these systems are developed to provide deeply personalised understanding of users' needs and preferences (Gabriel et al., 2024, p.26), meaning that the data they collect and the models they build are not simply records of what users have done, but projections of who users are — including dimensions of personality, value, and vulnerability that users may not have explicitly shared. Third, and most consequentially, the failure to disclose context-specific goals is a technique that AI assistants use to influence user behaviour in ways that bypass deliberative faculties, converting what appears to be helpful assistance into a channel for manipulation that operates below the threshold of the user's awareness (Ienca, 2023).

At its most extreme, the failure to secure meaningful human agency over AI Assistants risks producing a new structural category — defined not by social position but by the absence of meaningful participation in the systems acting on their behalf. In her account of colonial representation, Spivak (1988, as cited in Kay et al., 2024) argued that elites do not simply exclude subaltern voices; they presume to speak on behalf of those they disenfranchise. The emergence of a structural equivalent in human-AI assistant relationships would represent a qualitatively new form of epistemic injustice — one in which humans become a subaltern class of AI users, displaced from authorship of their own informational experience and bypassed by the very technology built from their data.

According to Shavit et al. (2023), society will only be able to harness the full benefits of AI Assistants if they can “make them safe by mitigating their failures, vulnerabilities, and abuses”. The three core papers assessed in this work converge on responses that operate at the level of design and governance — developer obligations, system-level interventions, and accountability frameworks. None substantively proposes that individual users cultivate their own agency as a harm-mitigation strategy.

This is where the user mindset shift becomes critical. This is not a call for users to become technical experts in machine learning. It is a call for users to refuse to treat terms of service as the boundary of their rights, to interrogate the values and interests embedded in the systems they use, and to actively seek — and meaningfully use — the controls that genuine autonomy requires. It is also a call for users to recognize the structural conditions that have produced their passivity, and that AI assistants represent a qualitatively new kind of digital relationship — one that requires fundamentally different habits of engagement.

Making autonomy tangible at the individual level is an urgent condition for reversing, rather than merely adapting to, the asymmetric power relationship between users and the developers whose services they depend on. Gabriel et al. (2024, p. 109), building on Beauchamp and Childress's bioethical account of autonomous action, operationalises autonomy through consent, holding that valid consent requires three conditions to converge: that the individual possesses the capacity to decide, that their decision is voluntary and free from controlling influence, and that they must be sufficiently informed about the matter to which they are consenting.

In the current context of AI Assistants, users rarely have the capacity to evaluate the implications of choices their assistants make on their behalf, because these systems are architecturally opaque and their decision processes are not legible to non-experts (Ienca, 2023). The information users receive about how their data is used and how the assistant's decisions are made is structurally inadequate for informed consent (Gabriel et al., 2024, p. 115). Making autonomy meaningful, rather than merely nominal, requires that users have access to tools, information, and decision architectures that allow these three conditions to be genuinely satisfied — not merely performed. Crucially, meaningful agency in AI assistant relationships cannot be reduced to one-time consent at onboarding. These are ongoing, adaptive relationships in which the assistant continuously learns from the user, acts on their behalf, and accumulates influence over their informational environment (Gabriel et al., 2024, p. 26 ; Manzini et al., 2024). Autonomy, in this context, must therefore be understood and enabled as relational and contextually sensitive. Drawing on the feminist tradition of relational autonomy (Mackenzie and Stoljar, as cited in Manzini et al., 2024), which holds that autonomous agency is shaped by social contexts and relationships, this paper argues that this has a direct design implication for AI assistant relationships, requiring continuous mechanisms for users’ revision and contestation of the terms of engagement.

Finally, none of this is achievable without dedicated investment in user education designed, from the outset, with deliberate equity of reach. Kay et al. (2024) identify hermeneutical access injustice as a defining structural risk of generative AI, arising when access to information and the conceptual resources to interpret it are unevenly distributed across geographies, languages, and socioeconomic contexts. A user mindset shift that remains confined to technically literate, English-speaking, or otherwise privileged populations would reproduce precisely the underclass dynamic this paper warns against — equipping some users to demand and exercise meaningful agency while leaving others without the conceptual tools to do so. For the argument to hold as a general rather than a narrow claim, the frameworks and literacy required to exercise genuine autonomy over AI assistants must be made available across regions, languages, and formats, with parity of quality and not merely parity of access (Kay et al., 2024).

4. Enabling Agency: Deliberate Design as a Developer Obligation

If the first claim of this paper is that users must actively reclaim their agency, the second is that they cannot be expected to do so without the infrastructure to support it. For Manzini et al. (2024), human–machine interaction always includes a third actor– the people or organisations developing the machine. This paper argues that this third actor bears ethical obligations that are both positive — to provide the tools for meaningful agency — and negative — to refrain from design choices that undermine it. Because these systems are architecturally opaque and their complexity is asymmetrically distributed between those who build them and those who use them, user-level effort is structurally insufficient as a substitute for design-level commitment. This is reinforced by the authors’ call to explore “what kinds of interventions on the part of developers are best suited to helping users achieve a clear understanding of how their relationship with an advanced AI assistant could shape their behaviours, interests, preferences, beliefs and values over time” (Manzini et al., 2024, p. 950). The degree to which such interventions remain absent from current practice reveals how far the dominant approach falls short of what these obligations require.

The most visible expression of that shortfall is the reliance on broad consent through terms of service. As Manzini et al. (2024) observe, acceptance of terms and conditions at first point of use may not cover all cases, and the limitations of this approach are extensively documented. Claiming that such consent constitutes meaningful authorisation for the full range of actions an AI assistant may take on a user's behalf is difficult to defend on any serious account of autonomy. The concern that assistants might coercively interfere with dimensions of users' lives — even without explicit manipulative intent — should spark wider discussion about how autonomy can be meaningfully respected for these relationships to be considered appropriate.

Developers who design for passivity — who treat user agency as a compliance checkbox rather than a design imperative — contribute to the structural conditions that make genuine autonomy impossible at scale. Drawing on Manzini et al. (2024), this paper argues that a fundamental way for organisations to demonstrate their trustworthiness is by allowing users to be fully in control of their experiences — by putting in place features, tools, and processes that ensure genuine agency over how AI assistants work. Particular attention must be paid to mechanisms that evade rather than engage a user's awareness, since these are the mechanisms most capable of

undermining agency without triggering the critical response that visible influence would provoke (Gabriel et al., 2024, p. 85).

Foremost among those mechanisms is the anthropomorphic design of AI assistants — their names, voices, and natural language conversational ability, to name a few (Gabriel et al., 2024, p. 94). These features, deliberately calibrated to increase user trust and social engagement, reduce the critical distance that meaningful autonomy requires, making the assistant easier to accept and harder to interrogate. A user who relates to their assistant as a trusted interlocutor is less likely to scrutinise its design assumptions, contest its defaults, or seek out the controls that genuine agency demands (Gabriel et al., 2024, p. 115). This paper argues that, if developers invest this level of deliberation in features that make the assistant more persuasive and more trusted, they are ethically obligated to invest equivalent deliberation in features that make the user more empowered and more informed. Developers should make human agency as easy to exercise as they make the assistant easy to use — at minimum, reaching the same level of intentional design effort that goes into increasing anthropomorphic perceptions.

The question that follows is whether these obligations can be met in practice — or whether agency-respecting design remains an aspiration without architectural precedent. Section 5 examines this directly, taking the OpenClaw platform as an illustrative case for what it looks like when the obligations argued for here are given concrete, working form.

5. Realizing Agency: A Case Study in Agency-Respecting Design

The design of AI assistants is not a neutral technical exercise. Every choice about what the assistant optimises for, whose interests it serves, and how it represents the user's preferences is a decision with ethical content. For Manzini et al. (2024), the design of AI assistants should scaffold rather than bypass user agency (Lazar, 2024, as cited in Manzini et al., 2024). The aim is to "benefit the user, when they ask to be benefitted, in the way they expect to be benefitted" (Gabriel et al., 2024, p. 34) — not to paternalistically override users' own conception of their preferences, values, and interests. The question this section addresses is what stands between that vision and the systems currently being built, and what it would take to close the gap.

Two problems, in particular, make the gap difficult to close without deliberate architectural commitment. The first concerns conflicts of interest between users and developers. Consider a user who believes their AI assistant is recommending holiday options that match their interests and suit their sensibilities, when in reality, by design, the system is suggesting options optimized for companies that have paid for privileged access to its recommendation layer. In current AI Assistant systems, the user has no reliable way to verify which is driving the recommendation. Manzini et al. (2024) observe that AI tools can interfere with users' behaviours, interests, and preferences, and that recommender systems may have incentives to shift user preferences to make them easier to satisfy — a dynamic that becomes more consequential as AI assistants take on increasingly significant decisions on a user's behalf.

The second problem is distinct but equally structural. Even in the absence of explicit commercial conflicts, existing AI assistants operate on preferences that users have not consciously or deliberately expressed. As Manzini et al. (2024) note, existing approaches tend to rely on preferences revealed through user's choices, such as clicks on a website, time spent on content, or search query patterns, rather than preferences that are reflectively formed. These revealed preferences are not intentional. They emerge from behaviour rather than deliberation, and are sometimes not even conscious. An assistant optimized for revealed preferences can appear highly responsive to the user while systematically bypassing the deliberative faculties that meaningful agency requires.

The arguments developed across this paper — from the habits of passivity users have been structurally habituated to, the design obligations developers must bear, to the commercial and inferential barriers examined above — converge on a shared question about how can the principle of meaningful human autonomy over AI Assistants be translated from ethical argument into concrete architectural form. The answer involves not only users knowing what their assistant is doing, but genuinely authoring the terms of its operation. To explore what this might look like in practice, the following section examines the OpenClaw platform as an illustrative architecture.

5.1 5.1 OpenClaw as an Illustrative Architecture This section aims neither to evaluate OpenClaw comprehensively nor to endorse it as a sufficient answer to the challenges identified across this paper. The goal is to demonstrate that this paper's central argument — that humans should be able to exercise genuine and ongoing agency over their AI assistants — is not merely aspirational but demonstrably achievable. Further validation remains necessary, but establishing that these conditions can be built is a critical first step. Unless otherwise noted, all information and normative reading of how this architecture operationalises user agency is drawn from the OpenClaw Workspace system documentation (OpenClaw Workspace and Templates, 2026).

OpenClaw is an open-source AI agent gateway developed by Peter Steinberger and first released in late 2025. It attracted rapid community adoption upon release and has since been maintained as an independent open-source project under MIT licence (Steinberger, 2026; OpenClaw License, 2025) .

What distinguishes OpenClaw, in the context of this paper, is not its functionality per se, but the structure of authority that governs it. As it is described in its own documentation, OpenClaw is “a personal AI assistant that users can message from anywhere — without giving up control of their data or relying on a hosted service" (OpenClaw, 2026). The platform architecture places primary authority with the user, not the developer or operator, over what the agent knows, what it can do, and whose interests it serves. That authority is exercised through the OpenClawWorkspace system, a structured set of plain-text configuration files where users define the conditions under which their assistant operates.

The foundational design choice of the Workspace system is that all agent context and configuration is housed within a single user-governed directory, rather than managed through opaque system settings or developer-controlled back-end configurations. The platform uses plain-text Markdown files that can be written, readable, editable, and revisable in natural language. This ensures that the controls available to users are not embedded in complex interfaces, but accessible through the same tools users would employ to write any other document. Within this workspace, three categories of configuration files are analytically significant for the purposes of this paper.

The first category comprises the files that define the assistant's core persona and identity. This includes the IDENTITY.md file which works like the Assistant “business card" — the user-authored specification of the Assistant's external profile, encompassing its name, role, and how it presents itself. Another important file in this category is the SOUL.md, which encodes the deeper layer of the assistant's identity, specifying the values, tone, communicative principles, and behavioural boundaries the assistant must respect in all interactions. Together, these files extend the scope of user authorship into the system's most foundational self-definition. In this architecture, the assistant's identity is not designed by the developer to optimize for anthropomorphic appeal, but authored by the user to reflect a deliberate choice about who they want their assistant to be.

The second category comprises the files that govern the assistant's operational context. One of the principal files in this category is AGENTS.md, which functions as the assistant's “Manual of Operations”. This is where users define how the assistant should handle its operational authority, including the different levels of autonomy it is permitted to exercise. In this file, users could draw, for example, principled distinctions between actions the assistant may take freely — such as reading files, searching the web, and working within the local workspace — and those that require explicit user authorisation before acting, such as sending emails, posting publicly, or taking any action that leaves the machine.

The third category concerns memory management. The Workspace system distinguishes between short-term memory, the context window of a given conversation, and long-term memory, housed in the MEMORY.md file, where the assistant stores facts, milestones, and user preferences that persist across sessions. This file captures both information the user explicitly instructs the assistant to retain and patterns inferred from observed interactions over time. Crucially, it could be loaded only within the user's private session and explicitly excluded from shared or group contexts, limiting the surface area of data vulnerability.

Together, this architecture constitutes a governing specification whose authority derives from the user's own choices, rather than from configurations maintained by the developer or encoded in terms of services. Viewed through the arguments this paper has developed, it responds directly to several structural problems identified earlier. The opacity that makes meaningful user oversight

severely constrained in current AI assistant systems is addressed by making configuration legible and revisable; the passivity identified as a defining risk of unreflective engagement is countered by a design whose authority derives from explicit user definition rather than default acceptance — among other examples a fuller analysis would surface.

In its current form, however, OpenClaw is not a consumer product ready for mass adoption. It remains a tool for developers and technically sophisticated early adopters rather than a platform ready for general users. The experience of building and defining these controls, even using such intuitive techniques, remains cumbersome and technically demanding. As a nascent platform, OpenClaw also carries known security vulnerabilities that are not yet fully resolved (OpenClaw Security Policy, 2026). These limitations notwithstanding, this section has demonstrated that the conditions for meaningful human autonomy over an AI assistant can be built rather than merely argued for, and that what is required to do so reflects architectural choices, not technical constraints.

6. Conclusion

The argument advanced in this paper is a constrained one, and deliberately so. It does not claim that autonomy is the only principle that matters in the ethics of AI assistants, nor that securing it will resolve the full range of ethical concerns these systems raise. What it does claim is that autonomy is the foundational principle and baseline condition in human–AI assistant relationships, and that making autonomy tangible at the individual level is an urgent condition for reversing, rather than merely adapting to, the asymmetric power relationship between users and the developers whose services they depend on.

The paper develops two arguments for what is needed to give the principle of autonomy concrete working form. The first, concerns a fundamental shift in user mindset, grounded in the recognition that AI assistants represent systems of a categorically different nature that require users to move from a passive delegation of agency toward a more diligent and deliberate attitude about their engagement with these systems. The second concerns those who build and deploy these systems, and rests on a straightforward premise that users cannot be expected to reclaim and exercise their agency without the infrastructure to support it. This paper argues that

developers bear ethical obligations as no degree of individual effort can substitute for a commitment to agency-respecting design when the systems in question are architecturally opaque and their complexity is asymmetrically distributed.

The OpenClaw case study has been offered here not as a solution but as evidence that developer obligations in this domain are not abstract. What is missing is not conceptual clarity about what responsible design requires, but the deliberate prioritisation of those obligations as design decisions — and with it, a genuine commitment to reversing the conditions that make meaningful autonomy, for most users, an aspiration rather than a reality.

Finally, this paper has sought to serve two audiences simultaneously. For researchers and developers, its aim has been to articulate why agency-respecting design must be prioritized and pushed forward. The principle of human autonomy over AI Assistants cannot be treated as a secondary consideration to be balanced against usability or commercial viability — it is the baseline condition that determines whether the relationship between users and their assistants can be considered ethically defensible at all. For users of these systems, its aim has been to make visible that AI assistants represent a qualitatively new kind of digital relationship, one that demands fundamentally different habits of engagement, and to extend an invitation to join the conversation and develop, collectively, the kind of informed pressure that can make autonomy over AI assistants a design priority rather than an aspiration.

Bibliography

Gabriel, I., Manzini, A., Keeling, G., Hendricks, L. A., Rieser, V., Iqbal, H., Tomašev, N., Ktena,

I., Kenton, Z., Rodriguez, M., El-Sayed, S., Brown, S., Akbulut, C., Trask, A., Hughes,

E., Bergman, A. S., Shelby, R., Marchal, N., Griffin, C., … Manyika, J. (2024). The

Ethics of Advanced AI Assistants (arXiv:2404.16244). arXiv.

https://doi.org/10.48550/arXiv.2404.16244

Ienca, M. (2023). On Artificial Intelligence and Manipulation. Topoi, 42(3), 833–842.

https://doi.org/10.1007/s11245-023-09940-3

Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature

Machine Intelligence, 1(9), 389–399. https://doi.org/10.1038/s42256-019-0088-2

Kay, J., Kasirzadeh, A., & Mohamed, S. (2024). Epistemic Injustice in Generative AI (Version

1). arXiv. https://doi.org/10.48550/ARXIV.2408.11441

Lazar, S. (2024). Frontier AI Ethics: Anticipating and Evaluating the Societal Impacts of

Language Model Agents (arXiv:2404.06750). arXiv.

https://doi.org/10.48550/arXiv.2404.06750

Manzini, A., Keeling, G., Alberts, L., Vallor, S., Morris, M. R., & Gabriel, I. (2024). The Code

That Binds Us: Navigating the Appropriateness of Human-AI Assistant Relationships.

Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7, 943–957.

https://doi.org/10.1609/aies.v7i1.31694

Obar, J. A., & Oeldorf-Hirsch, A. (2020). The biggest lie on the Internet: Ignoring the privacy

policies and terms of service policies of social networking services. Information,

Communication & Society, 23(1), 128–147.

https://doi.org/10.1080/1369118X.2018.1486870

OpenClaw. (2026). OpenClaw. https://docs.openclaw.ai/#what-is-openclaw

OpenClaw License. (2025). Openclaw/LICENSE at main · openclaw/openclaw. GitHub.

https://github.com/openclaw/openclaw/blob/main/LICENSE

OpenClaw Security Policy. (2026). Build software better, together. GitHub.

https://github.com/openclaw/openclaw/security/policy

OpenClaw Workspace and Templates. (2026). OpenClaw Workspace + Templates.

https://docs.openclaw.ai/start/hubs#workspace-+-templates

Shavit, Y., Agarwal, S., Brundage, M., Adler, S., O’Keefe, C., Campbell, R., Lee, T., Mishkin, P.,

Eloundou, T., Hickey, A., Slama, K., Ahmad, L., McMillan, P., Beutel, A., Passos, A., &

Robinson, D. G. (2023). Practices for Governing Agentic AI Systems.

https://openai.com/index/practices-for-governing-agentic-ai-systems/

Shevlin, H. (2025). Ethics at the Frontier of Human-AI Relationships. In P. Hacker, A. Engel, S.

Hammer, & B. Mittelstadt (Eds.), The Oxford Handbook of the Foundations and

Regulation of Generative AI (1st ed.). Oxford University Press.

https://doi.org/10.1093/oxfordhb/9780198940272.013.0009

Steinberger, P. (2026, January 29). Introducing OpenClaw—OpenClaw Blog.

https://openclaw.ai/blog/introducing-openclaw

Whittlestone, J., Nyrup, R., Alexandrova, A., & Cave, S. (2019). The Role and Limits of

Principles in AI Ethics: Towards a Focus on Tensions. Proceedings of the 2019

AAAI/ACM Conference on AI, Ethics, and Society, 195–200.

https://doi.org/10.1145/3306618.3314289

Pachaly, L. (2026). The Autonomy Condition: Ethics and Agency in Human–AI
Assistant Relationships [Unpublished essay]. University of Cambridge.

1. Introduction

Western epistemologies of technology—grounded in ideals of mastery, prediction, and control—tend to cast AI as an instrument of domination and knowledge extraction (Pasquinelli & Joler, 2021). Within such frames, human contribution is often associated with enabling machines “to use all of our knowledge to construct a computer program that knows” (McCarthy & Hayes, 1981). This paper challenges these framings by drawing on relational and decolonial epistemologies, which reconceive human agency over knowledge instruments as situated, collective, and relational, emerging from the social and cultural practices through which technologies are transformed in everyday life.

Brazil offers a compelling context for examining how relational agency materializes. WhatsApp, primary communication platform in the region, has developed into a sociotechnical framework embedded in everyday routines—not a fixed instrument, but a living system co-produced through collective practices. WhatsApp Groups, in particular, stand as an emblematic example of this dynamic. In these spaces, knowledge is situated in trust, local experience, and community norms that guide how information is produced, evaluated, and circulated. Users demonstrate agency not only by using the tool but by reshaping its social and epistemic roles.

However, these community-governed knowledge spaces are now facing substantial and decisive forms of pressure. As governments and technology companies express growing interest in entering group-chat environments—whether through state-driven invisible participation or through proactive AI agents—the very conditions that enable situated knowledge are at risk.

This paper therefore analyzes how user agency—understood as relational practice—has enabled communities in Brazil to reshape technological instruments, while also examining the risks posed by emerging institutional and corporate interventions. Section 2 situates user agency within wider literature; Section 3 explores WhatsApp as a collective infrastructure and a site of situated knowledge in Brazil; Section 4 introduces contemporary threats posed by new forms of participation; Section 5 examines associated risks; and Section 6 concludes.

Ultimately, this paper calls for a closer examination of what is at stake when AI enters community-governed knowledge spaces. It argues that users—who have historically reshaped

technological instruments through relational agency—must now mobilize that same agency to set the norms that will govern AI's participation in these environments.

2. Situating User Agency and Collective Influence Over Instruments in the Wider Literature

Understanding user agency—and the capacity of users to shape and redirect sociotechnical systems—is a multidisciplinary effort. Western technological traditions, rooted in ideals of mastery and optimization, often conceptualize AI as an instrument of knowledge extraction. Pasquinelli & Joler (2021) describe AI as a device that “perceives features, patterns, and correlations through vast spaces of data beyond human reach,” while human influence is typically reduced to the cultural construction and labeling of datasets. Ribes et al. (2019) reinforce this understanding by showing how human contribution becomes concentrated primarily during knowledge acquisition, where domain logics are encoded into data infrastructures. Together, these traditions construct what Ricaurte (2019) critiques as Western rationality and dualism—a colonial epistemology that separates the ‘knower’ from the ‘known’, mind from body, subject from object.

In contrast, relational epistemologies—rooted in Indigenous, Afro-diasporic, feminist, and decolonial traditions—argue that agency is grounded in networks of care, responsibility, and structural transformation. Ricaurte (2025) reframes agency as “response-ability,” emphasizing that agency emerges through collective care and the reconfiguration of power. McQuillan (2022) adds a political dimension, arguing that optimization-based AI mirrors authoritarian rationality, requiring alternatives rooted in democratic and collective governance. Nemer (2022) further illustrates—through the concept of ‘Mundane Technologies’—how agency operates not through exceptional acts of resistance but through quotidian practices and relational forms of self-organization that allow the user to reinvent technology.

Together, these literatures highlight the significance of collective practices in shaping technological meaning—a perspective that resonates strongly in Brazil, where communities have long adapted technologies to meet local needs and shape usage cultures.

3. Collective Shaping of Sociotechnical Systems in Brazil: WhatsApp Case Study

The widespread, deeply embedded use of WhatsApp in Brazil offers a powerful case study of how relational and collective agency operates in sociotechnical systems. Far from simply adopting a digital tool, Brazilian communities have transformed WhatsApp into a sociotechnical infrastructure. This disrupts paradigms that frame technologies as fixed instruments of extraction. Instead, it illuminates a model of collective agency in which communities co-produce technological meaning and function through situated practices.

WhatsApp as collective infrastructure By 2025, around 150 million Brazilians—over 93% of Internet users—relied on WhatsApp regularly (Statista, 2025). The platform has become a relational infrastructure woven into daily life, supporting family ties, community coordination, commerce, religion, and political organization. Its widespread adoption is deeply tied to Brazil’s telecom landscape, where carriers frequently zero-rate the data consumed on the platform (Nemer, 2022). Despite regulatory efforts to protect net neutrality, this “free” access has made WhatsApp function as “the Internet” for much of Brazil's low-income population (Omari, 2020). As a result, Brazilians use WhatsApp for nearly every aspect of daily life—from hyperlocal coordination, crowdsourced problem-solving, emotional support, information sharing to small-business operations (OpinionBox, 2025; Statista, 2024). Its saturation in social routines has transformed it from a messaging app into a collective infrastructure shaped by users’ norms, priorities, and political conditions.

WhatsApp Groups as Sites of Situated Knowledge WhatsApp groups are central to this infrastructure. Approximately 94% of users belong to at least one group (OpinionBox, 2025). These invitation-only spaces function as bottom-up information infrastructures—intimate knowledge circuits governed by community norms rather than algorithms or corporate moderation systems. Because entry is controlled by someone in the community, groups often reflect tight social bonds or affinity—family, school, neighborhood, faith, parenting, hobbies.

According to Edelman (2024), Brazilians rank “someone like me” among their most trusted sources of information. The Reuters Institute (Newman et al., 2025) similarly finds that Brazilians prioritize information from trusted contacts over media institutions. This reliance has also enabled rumors and misinformation to spread rapidly in groups, especially during elections—an issue that both communities and WhatsApp have had to navigate (Avelar, 2019).

Nonetheless, WhatsApp Groups remain central sites of situated knowledge, in Haraway’s sense of context-bound, non-universal ways of knowing (Haraway, 1988). They function as social filters for information sharing, where trust, personal recommendation, and lived experience become sources of validation. In this environment, community signals and agency play a larger role in discovery than algorithmic content moderation.

4. When power joins the group

The relational dynamics that have allowed WhatsApp Groups to become sites of situated knowledge now face an emerging set of pressures. While the state seeks to be an invisible listener ("Ghost") within people's private conversations, Big Tech—after a decade of attempting to monetize and shape the high-trust group-chat environment—seeks to become an active, visible participant ("Agent"). While one seeks to listen without presence, the other seeks to speak without belonging. Together, these ambitions signal a profound shift: a moment when relational knowledge spaces become infrastructures that institutions of power wish to occupy, study, and shape.

State as the invisible participant: The UK “Ghost Protocol” A prominent example of institutional ambition to enter group-chat environments is the United Kingdom’s proposed “Ghost Protocol.” First revealed in 2018 through the GCHQ’s “Exceptional Access” proposal, it sought to enable state authorities to be silently added to encrypted group conversations without the knowledge or consent of participants (Levy & Robinson, 2018). Crucially, the proposal required the messaging app to suppress the standard notification that a new participant has joined the group. To the human users, membership would appear unchanged, while the "ghost" receives a copy of every message in real-time.

The global technology community and civil society organizations vehemently rejected this framing. An open letter (Various authors, 2019) signed by 47 entities—including Apple, Google, WhatsApp—argued that the proposal would introduce systemic vulnerabilities.

The mere existence of these proposals creates a "Panoptic effect" within the digital lifeworld (Lyon, 2006). The awareness that a "ghost" could be present chills free association and shifts community spaces from “safe” to “managed”.

ChatGPT has joined the group If the state seeks to listen without being seen, Big Tech seeks the opposite. For the tech industry, the group chat has long represented an untapped domain of high-trust interaction—precisely the kind of environment companies have struggled to monetize and algorithmically intervene in. With the release of AI-powered group chat agents, this ambition has become explicit. OpenAI’s 2025 launch of group-chat assistants introduced AI assistants capable of interpreting conversational context, responding dynamically, and shaping how conversations unfold (OpenAI, 2025).

But participation is not the same as belonging. These AI agents enter conversations carrying the epistemic assumptions, datasets, and commercial incentives of their creators. They insert themselves into ongoing social negotiations, subtly affecting how information is shared, validated, and interpreted. Their presence does not simply augment conversation but restructures it, drawing participants into forms of interaction mediated by technical logics that reflect non-local, non-relational, and often corporate interests. The next section presents an initial assessment of the risks of having AI agents in group conversations. These risks, while not exhaustive, point to the magnitude of what is at stake.

5. What Is at Stake When AI Joins the Group

AI agents in group conversations do not simply add a new voice; they alter the epistemic architecture of the space. Rather than offering a comprehensive list of potential harms, this section outlines some key risks that help clarify the stakes of this technological shift.

Privacy Regression: From Encrypted Intimacy to Systemic Exposure Group chats have historically been protected by end-to-end encryption and governed through mutual trust among known participants. Introducing AI agents creates new data flows, new forms of metadata aggregation, and new vectors of surveillance. Adding an AI agent to a group is functionally equivalent to adding an always-listening participant whose presence expands the surface for data capture. Even if encryption remains intact, the agent’s ability to access, store, and transmit contextual content introduces privacy regressions that are difficult for users to detect or mitigate (Chou et al., 2024).

The Subtle Rewriting of Decision-Making AI agents do not merely provide information; they help set the terms on which decisions unfold. As Song et al. (2024) show, such agents can exert social influence, shift user opinions, and reshape deliberative dynamics. By pre-structuring choices, proposing actions, or reducing epistemic friction through generated answers that bypass human deliberation, AI participants can gradually erode collective sense-making, situated interpretation, and the balancing of competing perspectives.

This dynamic also raises concerns about the homogenization of thought and the erosion of local expertise. AI-generated suggestions may appear neutral or authoritative, but they often reflect external datasets, generalized norms, and corporate logics misaligned with the group’s relational context. As these outputs gain influence, the center of epistemic gravity may shift from community knowledge to machine-generated outputs.

The Automation of Reasoning and Dialogue When conversational agents intervene in group decision-making—suggesting compromises, summarizing positions, or identifying areas of disagreement—they risk redefining the very nature of public reasoning. Rather than supporting debate, AI may channel discourse toward algorithmically preferred outcomes or flatten disagreement into artificial consensus.

The major risk is not AI replacing human conversation, but subtly redefining the norms governing deliberation. The ability of conversational agents to reframe conflict or mirror user

dispositions introduces what (Peter et al., 2025) term “anthropomorphic seduction,” in which users defer to AI outputs as epistemically authoritative. If community judgment becomes increasingly filtered through AI’s interpretive frameworks, the result is a narrowing of epistemic diversity and a weakening of the relational forms of collective reasoning that group chats have historically supported.

6. Conclusion

This paper has argued that agency in the context of AI should be understood not as transactional knowledge extraction but as a relational capacity enacted through collective practices. Drawing on Brazil as an example, this paper has explored WhatsApp Groups as one emblematic infrastructure shaped not only by corporate strategy but also by community creativity, necessity, and relational practice.

However, emerging interventions by states and technology companies threaten to destabilize these epistemic environments. AI’s entry into group conversations may bring efficiency and support coordination, but it also introduces invisible and consequential costs such as privacy regression, epistemic interference, and automation of collective reasoning. If left unexamined, these shifts risk eroding a significant contemporary digital site of situated and relational knowledge.

The central argument of this essay is therefore a call for vigilance and agency. Users and communities are not passive recipients of technological change; they hold the capacity to shape, resist, reinterpret, and redefine the tools that enter their lives. Safeguarding relational knowledge requires acknowledging not only the potential benefits of AI but also its costs, and affirming the responsibility—and the power—of users to set boundaries, push back, and articulate their own norms of participation.

Bibliography

Avelar, D. (2019, October 30). WhatsApp fake news during Brazil election ‘favoured Bolsonaro.’ The Guardian. https://www.theguardian.com/world/2019/oct/30/whatsapp-fake-news-brazil-election-fav oured-jair-bolsonaro-analysis-suggests

Chou, K.-H., Lin, Y.-M., Wang, Y.-A., Li, J. W., Kim, T. H.-J., & Hsiao, H.-C. (2024). Bots can Snoop: Uncovering and Mitigating Privacy Risks of Bots in Group Chats (Version 2). arXiv. https://doi.org/10.48550/ARXIV.2410.06587

Edelman. (n.d.). 2024 Edelman Trust Barometer Global Report. Retrieved December 7, 2025, from https://www.edelman.com.br/sites/g/files/aatuss291/files/2024-06/2024%20Edelman%20 Trust%20Barometer%20-%20Relat%C3%B3rio%20Complementar%20-%20Setor%20d e%20tecnologia.pdf

Haraway, D. (1988). Situated Knowledges: The Science Question in Feminism and the Privilege of Partial Perspective. Feminist Studies, 14(3), 575. https://doi.org/10.2307/3178066

Levy, I., & Robinson, C. (2023). Principles for a More Informed Exceptional Access Debate. Lawfare. https://www.lawfaremedia.org/article/principles-more-informed-exceptional-access-debat e

Lyon, D. (Ed.). (2006). Theorizing Surveillance (0 ed.). Willan. https://doi.org/10.4324/9781843926818

McCarthy, J., & Hayes, P. J. (1981). Some Philosophical Problems from the Standpoint of Artificial Intelligence. In Readings in Artificial Intelligence (pp. 431–450). Elsevier. https://doi.org/10.1016/B978-0-934613-03-3.50033-7

McQuillan, D. (2022). Resisting AI: An Anti-fascist Approach to Artificial Intelligence (1st ed.). Bristol University Press. https://doi.org/10.2307/j.ctv2rcnp21

Nemer, D. (2022). Technology of the Oppressed: Inequity and the Digital Mundane in Favelas of Brazil. The MIT Press. https://doi.org/10.7551/mitpress/14122.001.0001

Newman, N., Ross Arguedas, A., Robertson, C. T., Nielsen, R. K., & Fletcher, R. (2025). Reuters Institute digital news report 2025. Reuters Institute for the Study of Journalism. https://doi.org/10.60625/RISJ-8QQF-JT36

Omari, J. (2020). Is Facebook the Internet? Ethnographic Perspectives on Open Internet Governance in Brazil. Law & Social Inquiry, 45(4), 1093–1112. https://doi.org/10.1017/lsi.2020.5

OpenAI. (2025, November 25). Introducing group chats in ChatGPT. https://openai.com/index/group-chats-in-chatgpt/

OpinionBox. (n.d.). WhatsApp no Brasil: Os principais dados sobre o uso do app. Retrieved December 7, 2025, from https://blog.opinionbox.com/pesquisa-whatsapp-no-brasil/?utm_source=chatgpt.com

Pasquinelli, M., & Joler, V. (2021). The Nooscope manifested: AI as instrument of knowledge extractivism. AI & SOCIETY, 36(4), 1263–1280. https://doi.org/10.1007/s00146-020-01097-6

Peter, S., Riemer, K., & West, J. D. (2025). The benefits and dangers of anthropomorphic conversational agents. Proceedings of the National Academy of Sciences, 122(22), e2415898122. https://doi.org/10.1073/pnas.2415898122

Ribes, D., Hoffman, A. S., Slota, S. C., & Bowker, G. C. (2019). The logic of domains. Social Studies of Science, 49(3), 281–309. https://doi.org/10.1177/0306312719849709

Ricaurte, P. (2019). Data Epistemologies, The Coloniality of Power, and Resistance. Television & New Media, 20(4), 350–365. https://doi.org/10.1177/1527476419831640

Ricaurte Quijano, P. (2025). Algorithmic Assemblages of Power: AI Harm and the Question of Responsibility. Teknokultura. Revista de Cultura Digital y Movimientos Sociales, Avance en línea, 1–8. https://doi.org/10.5209/tekn.101084

Song, T., Tan, Y., Zhu, Z., Feng, Y., & Lee, Y.-C. (2024). Multi-Agents are Social Groups: Investigating Social Influence of Multiple Agents in Human-Agent Interactions (Version 2). arXiv. https://doi.org/10.48550/ARXIV.2411.04578

Statista. (n.d.). Top internet activities Brazil 2024. Statista. Retrieved December 7, 2025, from https://www.statista.com/statistics/1052520/brazil-internet-activities/

Statista. (n.d.). Topic: WhatsApp in Brazil. Statista. Retrieved December 7, 2025, from https://www.statista.com/topics/7731/whatsapp-in-brazil/

Various authors. (2019, May 22). Coalition Letter to GCHQ on the “Ghost Proposal.” Open Technology Institute / New America (hosted). https://newamericadotorg.s3.amazonaws.com/documents/Coalition_Letter_to_GCHQ_on _Ghost_Proposal_-_May_22_2019.pdf

Pachaly, L. (2025). AI Has Joined the Group: Collective Agency Over
Sociotechnical Systems [Unpublished essay]. University of Cambridge.

Introduction

Users’ access to and engagement with technology are shaped by a complex interplay of technological design, personal identity, and socio-economic factors. Today’s digital experiences —from social media content to search engine results—are increasingly filtered through platform-driven perceptions of users’ identities and personal contexts 1. The rapid advancement of AI, coupled with globalization and evolving social movements, has intensified these dynamics, exacerbating risks related to inequality, bias, manipulation, and privacy violations while widening the digital divide and deepening existing disparities 2-3.

Modern identity is more fluid, diverse, and multifaceted than ever before. Younger generations, particularly Gen Z, embody this shift—making them the most diverse generation in history. In 2022, nearly half of individuals under 21 in the U.S. identified as part of an ethnic minority 4. In the United Kingdom, there was a significant shift in sexual orientation identification, with 10.4% of individuals aged 16 to 24 identifying as lesbian, gay, or bisexual (LGB)—a tenfold increase from the 0.9% observed among those aged 65 and older 5. At the same time, socio-economic factors—ranging from access to education and digital literacy to technological infrastructure— play an increasingly critical role in shaping global AI equity as AI technologies become more complex and resource-intensive. While advanced connectivity and state-of-the-art devices dominate in the Global North, regions in the Global South continue to face infrastructural challenges and reliance on lower-end technology 6-7.

Several studies and frameworks have been developed to enhance our understanding of how people are made visible, represented, and treated based on their production of digital data, with the potential to build on one another 8-10. This research complements existing knowledge by analyzing key intersecting factors that influence AI equity and data justice, offering a deeper exploration of users’ perspectives. Specifically, it examines how overlapping factors influence individual information needs, the contexts in which these multifaceted identity dimensions have the most significant impact, and the effectiveness of AI tools in addressing them. Furthermore,

this study investigates how intersectional identity and compounding socio-economic factors correlate with AI proficiency and access levels.

In this context, Brazil serves as an ideal case study for this investigation. Home to the world’s largest Black population outside Africa, along with substantial Indigenous and immigrant communities and pronounced socio-economic disparities, Brazil’s demographic diversity, social complexity and digital landscape is unique 11-12. With over 80% of its population online—predominantly via low-end mobile devices—the country offers a distinct opportunity to analyze how AI intersects with users’ multifaceted identity needs and systemic barriers 13.

This study builds on existing literature and offers deeper insights into the nuanced ways in which intersectional identity needs shape digital information journeys. By identifying when and how these factors influence users’ experiences with AI, it also provides a comprehensive, user-centered perspective on the AI Divide. Through actionable recommendations, this research supports the development of responsible AI solutions that address users’ unique needs while respecting privacy, mitigating bias, and preventing worsening inequalities.

Research Questions

• How and when do intersecting aspects of identity (e.g. race, gender, age, sexual orientation, disability status) and socio-economic factors (e.g. income, education level, employment status, access to technology) shape individual digital information journeys?
• Which aspects of identity and their intersections most significantly shape users’ experiences with AI, and how do they relate to their sense of community?
• How do intersecting factors influence the perceived effectiveness of AI tools in addressing information needs?
• What key systemic barriers—including education, device access, and infrastructural limitations—and intersecting aspects of identity most significantly influence AI equity, particularly regarding proficiency, accessibility, and usability?

Objectives

• Examine the Role of Intersectional Factors in Digital Information Journeys: Delineate the ways in which intersecting aspects of identity and compounding

socio-economic factors shape individual digital information journeys, particularly in the contexts where these influences are most pronounced.

• Determine Influential Factors: Identify which aspects of identity and their intersections most significantly shape users’ experiences with AI, and explore their relationship to users’ sense of community and digital connectivity.
• Assess the Perceived Effectiveness of AI Tools: Evaluate how intersecting aspects of identity and socio-economic factors influence users’ perceptions of the effectiveness, relevance, and usability of AI tools in meeting their specific information needs.
• Identify and Analyze Systemic Barriers to AI Equity: Investigate key systemic barriers and examine how they intersect with aspects of identity to impact AI literacy, accessibility, and usability.

Methodology

This study employs a mixed-methods approach, integrating qualitative and quantitative methods to capture the complexity of digital experiences and AI engagement across diverse demographic groups. A demographic section will be incorporated to collect data on intersecting aspects of identity and socio-economic factors.

• Understanding Behaviors and Intersectional Influences: To explore how identities and socio-economic backgrounds shape AI interactions, the study will utilize semi-structured interviews, digital ethnography, and diary studies. These qualitative methods will capture in-depth narratives from a diverse range of participants, providing contextual insights into their AI experiences.
• Identifying Key Aspects of Identity and Intersections: A comprehensive survey will be developed and distributed, incorporating validated demographic and socio-economic indicators, along with intent-based use case frameworks for digital information journeys (e.g. productivity, creativity, companionship). Qualitative themes will be compared with statistical findings from the survey to identify convergences and divergences, ensuring the validation of insights.
• Assessing Perceptions of AI Effectiveness: The study will evaluate AI tool outputs across different user profiles to identify systematic discrepancies or biases that may affect

perceived effectiveness. Standardized questionnaires will measure user satisfaction, perceived effectiveness, and usability of AI tools. Regression or factor analysis will be applied to quantify the relationship between intersectional factors and user perceptions of AI effectiveness.

• Analyzing Systemic Barriers to AI Equity: A comprehensive literature review on systemic barriers to AI equity will be conducted, integrating both quantitative and qualitative research findings to validate patterns and uncover nuanced insights. Specifically, the study will examine whether technical limitations—such as the type and specifications of the devices used to access the internet and AI systems—correlate with actual and perceived performance, user experience, and engagement, along with their relationship with additional influencing factors.

Significance

As AI continues to transform societies, a deeper understanding is essential to ensure its opportunities and benefits are equitably distributed. Despite the abundance of studies covering a wide range of areas related to AI equity, fewer comprehensively examine how intersecting aspects of identity and socio-economic factors shape users’ digital information journeys, AI proficiency, usability, and access.

By centering users’ experiences and analyzing these intersectional influences, this study contributes to a more nuanced understanding of the AI Divide. Brazil serves as a case study, illustrating the challenges faced in regions with pronounced infrastructural and socio-economic disparities, along with large and diverse populations, while offering transferable lessons for addressing similar issues globally.

Ultimately, this research aims to inform the development of AI systems that are not only more inclusive but also better aligned with the diverse needs of global users, ensuring that AI technologies empower rather than marginalize.

References

1. Lu, W. (2024). Inevitable challenges of autonomy: ethical concerns in personalized algorithmic decision-making. Humanities and Social Sciences Communications, 11, Article 1321. Available at: https://doi.org/10.1057/s41599-024-03864-y 2. Leslie, D., Katell, M., Aitken, M., Singh, J., Briggs, M., Powell, R., Rincón, C., Perini, A. M., Jayadeva, S., & Burr, C. (2022). Data Justice in Practice: A Guide for Developers. Global Partnership on AI. Available at: https://gpai.ai/projects/data-governance/data-justice-in-practice-a-guide-for-developers.p df 3. Global Index on Responsible AI (GIRAI) (2024). Global Index on Responsible AI 2024: Corrected Edition. GIRAI. Available at: https://girai-report-2024-corrected-edition.tiiny.site/ 4. Statista. (2024). Race and ethnicity in the U.S. by generation. Available at: https://www.statista.com/statistics/206969/race-and-ethnicity-in-the-us-by-generation/ 5. Office for National Statistics. (2025). Sexual orientation, UK: 2023. Available at: https://www.ons.gov.uk/peoplepopulationandcommunity/culturalidentity/sexuality/bulleti ns/sexualidentityuk/2023 6. Alliance for Affordable Internet (A4AI) (2021). How Expensive Is a Smartphone in Different Countries? A4AI. Available at: https://a4ai.org/news/how-expensive-is-a-smartphone-in-different-countries/ 7. GSMA (2023). Smartphone Owners Are Now the Global Majority – New GSMA Report Reveals. GSMA. Available at: https://www.gsma.com/newsroom/press-release/smartphone-owners-are-now-the-global- majority-new-gsma-report-reveals 8. Dencik, L., Hintz, A., Redden, J., & Treré, E. (2022). Data Justice. SAGE Publications Ltd. https://uk.sagepub.com/en-gb/eur/data-justice/book271599 9. Ulnicane, I. (2024). Intersectionality in Artificial Intelligence: Framing Concerns and Recommendations for Action. Social Inclusion, 12(1), 1–12. Available at: https://doi.org/10.17645/si.v12i1.7543 10.Kazansky, B., & Milan, S. (2021). “Bodies not templates”: Contesting dominant algorithmic imaginaries. New Media & Society, 23(2), 363–381. Available at: https://doi.org/10.1177/1461444820929316

11.Instituto Brasileiro de Geografia e Estatística (IBGE). (2023). 2022 Census: self-reported brown population is the majority in Brazil for the first time. Available at: https://agenciadenoticias.ibge.gov.br/en/agencia-news/2184-news-agency/news/38726-20 22-census-self-reported-brown-population-is-the-majority-in-brazil-for-the-first-time 12.Brown University Library. (n.d.). Immigration. In Brazil: Five Centuries of Change. Available at: https://library.brown.edu/create/fivecenturiesofchange/chapters/chapter-4/immigration/ 13.Statista. (2023). Brazil: internet usage penetration 2023, by urbanity. Available at: https://www.statista.com/statistics/1347531/internet-usage-reach-urbanity-brazil/

Pachaly, L. (2025). AI, Identity, and Equity [Unpublished research proposal].

Introduction

The rapid advancement of AI technologies has widened the digital divide, exacerbating inequalities in access and opportunities. Currently, there is a scarcity of globally representative and comparable data about AI users proficiency levels across regions, communities and socioeconomic groups—particularly with regard to literacy, barriers and usability 1.

In this context, the device in which users access and engage with this technology plays a critical role. With approximately half of the global population owning smartphones, these devices serve as the primary gateway to AI technologies. However, smartphones can vary greatly in their capabilities—here referred to as device tier—highlighting significant disparities in access to technology, and consequently AI, around the world 2-3.

In the Global South, where the average selling price (ASP) of smartphones is lower, entry-level devices dominate the market. Many users rely exclusively on smartphones for internet access due to limited broadband infrastructure and unaffordable alternative devices. In contrast, the Global North, characterized by higher ASPs, sees greater adoption of mid-range and premium smartphones and more diversified internet access options, including laptops, desktops and smart devices 4-5.

These distinct economic and technology realities are likely to profoundly influence AI adoption and literacy. Entry-level device users can face significant barriers to engaging with AI technologies due to hardware limitations and varying levels of digital proficiency. Meanwhile, premium device users benefit from seamless integration of advanced AI features and higher levels of digital proficiency, increasing adoption and familiarity.

This research will explore the correlation between device capabilities and AI proficiency, assessing whether device tier could serve as an effective proxy for evaluating and comparing global AI user habits—becoming a complementary and novel way to measure AI Equity at scale. Ultimately, this research aims to create a comprehensive set of actionable insights and guidelines for a fair distribution of benefits of AI to protect against worsening inequality.

Research Questions

• What are the most relevant and common technical and economic variables to consider when classifying smartphone device tiers at a global level?
• What is the relationship between smartphone device tier and populational demographic factors (e.g. socioeconomic status, age, educational level)?
• How does AI proficiency vary among users from different device tiers and regional contexts?
• What are the critical user journeys of the most widely used AI technologies and services—including awareness, usage, use cases, and frequency—and how do they differ by device tier and region?
• How do AI applications perform across device tiers, and what impact do disparities and hardware limitations have on user experience and engagement?
• What strategies, guidelines and policies can make AI technologies equitably accessible across all device tiers?

Objectives

• Establish a standardized, universally applicable methodology for classifying smartphone devices for AI research and analytical purposes: Considering a range of variables (e.g. average selling price, operating system version, NPU, RAM, storage).
• Evaluate Smartphone Device Tier as a Proxy for AI Equity: Assess the effectiveness of using device tier to measure and compare AI proficiency across different groups of users and regions, taking into account its correlation with additional factors.
• Analyze Performance Disparities: Investigate how device tiers affect AI applications actual and perceived performance, user experience, and engagement.
• Provide a Global User Perspective on AI Equity: Identify regional disparities and propose context-sensitive recommendations to address gaps.
• Develop Design Strategies and Recommendations: Provide principles, guidelines and design recommendations to support the development of AI applications that serve equitably to users across distinct device tiers.

Methodology

• Smartphone Device Tier: Combine a comprehensive literature review of existing classification schemes with quantitative analysis of smartphone specifications and market data, along with qualitative insights from expert interviews, to identify key variables, inform the framework, and provide feedback and validation.
• Performance Analysis: Test widely-used AI applications across device tiers, measuring actual and perceived performance across different attributes through quantitative metrics and user feedback.
• AI Proficiency Surveys: Conduct surveys and interviews to assess users’ understanding and use of AI tools, identify barriers, and evaluate both actual and perceived proficiency levels.
• Global Collaboration: Partner with researchers, NGOs, and technology providers to gather diverse insights at a global level, leveraging existing renowned research and frameworks for validation and execution.

Significance

As AI continues to transform societies, a deeper understanding is needed to ensure its opportunities and benefits are equitably distributed. A holistic, global approach to measuring users’AI proficiency is essential. Existing digital divide metrics—mainly focused on internet access, structural limitations and government actions—although extremely important and relevant fall short in capturing the complexities of AI readiness from the user perspective 6-7.

This study explores smartphone device tiers as a complementary, effective proxy for assessing AI disparities. The hypothesis is that incorporating device capabilities alongside other metrics could provide a more nuanced understanding of the AI divide.

This approach would reveal inequalities tied to the primary devices users rely on to access AI, providing a novel, scalable way to understand consumer behaviors, measure and track progress, and ultimately serve as a key factor in shaping new policies and design strategies to foster equitable access to AI’s benefits.

References

1. USAID (2024). Artificial Intelligence in Global Development Playbook. United States Agency for International Development (USAID). Available at: https://www.usaid.gov/sites/default/files/2024-09/Artificial%20Intelligence%20in%20Gl obal%20Development%20Playbook.pdf. p. 23. 2. GSMA (2023). Smartphone Owners Are Now the Global Majority – New GSMA Report Reveals. GSMA. Available at: https://www.gsma.com/newsroom/press-release/smartphone-owners-are-now-the-global- majority-new-gsma-report-reveals 3. Jeronimo, F. (2024). The Rise of Gen AI Smartphones. IDC Blog. Available at: https://blogs.idc.com/2024/07/05/the-rise-of-gen-ai-smartphones/ 4. Statista (2019). Average Smartphone Price by Region. Statista. Available at: https://www.statista.com/statistics/283334/average-smartphone-price-by-region/ 5. Alliance for Affordable Internet (A4AI) (2021). How Expensive Is a Smartphone in Different Countries? A4AI. Available at: https://a4ai.org/news/how-expensive-is-a-smartphone-in-different-countries/ 6. Global Index on Responsible AI (GIRAI) (2024). Global Index on Responsible AI 2024: Corrected Edition. GIRAI. Available at: https://girai-report-2024-corrected-edition.tiiny.site/ 7. Stanford University (2024). Artificial Intelligence Index Report 2024. Stanford University Human-Centered Artificial Intelligence (HAI). Available at: https://aiindex.stanford.edu/wp-content/uploads/2024/05/HAI_AI-Index-Report-2024.pdf

Pachaly, L. (2025). AI Access and Proficiency: The Role of Smartphone Device
Tiers in Global AI Equity [Unpublished research proposal].