AI Literacy for Citizen Participation Report

State of the Discourse

Analysis of 1,331 AI literacy sources surfaced this week, drawn from a corpus of 6,252, reveals a discourse focused on equipping individuals to use AI products while neglecting their capacity to contest them. The citizen-as-participant framing — literacy as the civic equipment for a person whose tax records, job application, medical triage, or news feed will be processed by a model they did not choose — appears in fewer than one in seven sources. Most treat literacy as a productivity skill, a child-protection problem, or, increasingly, a vendor-supplied tutorial.

1. The Landscape

What “AI literacy” means in 2026 is being defined, in large part, by the companies whose products are the subject of the literacy. GitHub publishes prompt-engineering primers as a literacy artifact Prompt engineering for GitHub Copilot Chat; OpenAI defines “skills in ChatGPT” through its own help center Compétences dans ChatGPT - OpenAI Help Center; Microsoft frames generative AI competency as accessibility training Inteligencia artificial generativa y accesibilidad | Microsoft Learn and content moderation as a literacy primitive What is Azure AI Content Safety? - Azure AI services. The vocabulary — “skills,” “prompting,” “guardrails” — is borrowed wholesale from product marketing. A citizen reading these documents learns how to operate a tool. They do not learn how the tool was trained, who profits, or what happens when it is wrong about them. UNESCO’s parallel discourse pulls in a different direction, treating literacy as a defense against information collapse Inteligencia artificial y desinformación - UNESCO, but it is the smaller signal.

2. Whose Literacy

The asymmetry of voice is stark. Vendors publish for users; ministries publish for schools; researchers publish for peers. Citizen organizations — consumer advocates, labor unions, neighborhood associations, disability rights groups outside the educational pipeline — are nearly absent from the corpus. When ordinary people appear, they appear as victims (of deepfakes, of grooming, of harassment Una violencia que nunca acaba: ciberacoso, ‘grooming’, contenido sexual …) or as patients (youth mental health under algorithmic influence New National Research Reveals How Context Shapes AI’s Impact on Youth …). They rarely appear as deliberating agents with rights to assert. The implicit pedagogical relationship is consistent: someone with the model is teaching someone without the model how to behave around it.

3. What’s Being Taught

The thematic weight sits on three clusters: prompt craft, child safety, and content authentication. The first is operational — produce better outputs from a black box. The second is protective — a recent industry report frames children primarily as a risk surface to be managed PDF AI Child Safety Final Report. The third is forensic — teaching adults to spot deepfakes We Need Laws to Stop AI-Generated Deepfakes | Scientific American Deepfakes and the crisis of knowing - UNESCO and to discount fabricated citations, which one study found in more than half of the references ChatGPT generates ChatGPT’s Hallucination Problem: Study Finds More Than Half Of AI’s …. All three are useful. None of them teach a citizen how to read an algorithmic-impact assessment, file a data-access request, or evaluate a municipal procurement contract for facial recognition. Critical understanding is reduced to skepticism of outputs, not scrutiny of systems.

4. What’s Missing

The largest gap is governance literacy. A recent ETH Zurich analysis argues that democracies are now governing disinformation through AI without citizens understanding the trade-offs being made on their behalf When AI Governs (Dis)information: Five Lessons for …. Almost nothing in the corpus addresses data rights, consent regimes, or the procedural mechanisms by which a non-expert can object to automated decisions. Underserved populations — older adults, low-income workers exposed to algorithmic management, non-English speakers facing translation-mediated public services — are largely missing. A human-rights framing exists PDF Derechos Humanos E Inteligencia Artificial but remains marginal. The literacy being built is competent. It is not, yet, civic.

Core Tensions

The phrase “AI literacy” conceals genuine tensions about what citizens need to know and why. The most fundamental: literacy is being defined, almost everywhere, as competence with vendor products rather than capacity to judge the systems that increasingly govern public life. When OpenAI publishes its guide to Compétences dans ChatGPT and GitHub publishes its primer on Prompt engineering for GitHub Copilot Chat, they are not merely documenting tools — they are setting the implicit curriculum for what counts as a competent citizen of the AI era. This isn’t a knowledge gap to fill. It’s contested terrain.

Technical fluency vs. critical understanding. The dominant literacy materials in circulation this week teach citizens how to address the machine, not how to evaluate it. A competent prompt engineer can extract better outputs; that competence does nothing to detect that more than half of the references ChatGPT produces in academic contexts are fabricated, as documented in ChatGPT’s Hallucination Problem. The skill of prompting and the skill of disbelieving the response are different cognitive operations, and only the first is being systematically taught. UNESCO’s analysis of Deepfakes and the crisis of knowing frames this as an epistemic emergency: when synthetic media becomes indistinguishable from record, prompt fluency is not a defense — it is, if anything, a distraction from one.

Individual competency vs. collective governance. The second tension runs through nearly every literacy framework: whether the citizen’s job is to adapt or to govern. The Department of Education’s recently finalized AI priority, dissected by AEI, treats literacy as workforce preparation — a private good. Compare this with When AI Governs (Dis)information, which argues that the urgent literacy question is whether democracies can hold AI systems accountable when those systems mediate public speech. One framing produces a workforce; the other produces an electorate. They are not the same project, and conflating them — as most national strategies do — quietly resolves the politics in favor of the workforce frame.

Protection from vs. empowerment with. A third tension surfaces wherever AI meets vulnerability. Child-safety reports such as the AI Child Safety Final Report and Maldita’s reporting on digital violence in Spain document harms — grooming, non-consensual synthetic imagery, harassment — that demand protective infrastructure: content classifiers like Azure AI Content Safety, legal regimes of the kind argued for in We Need Laws to Stop AI-Generated Deepfakes. Yet the same population is sold “empowerment” curricula premised on confident use. Citizens are told simultaneously that AI is so dangerous it requires platform moderation and legislation, and so benign that primary-school children should be writing prompts. Both can be true; rarely is the contradiction named.

Consumer literacy vs. citizen literacy. The dominant metaphor across the corpus is tool — by an enormous margin over threat, and by orders of magnitude over partner or infrastructure. The tool framing locates agency in the user: a hammer is neither good nor bad. This is convenient for vendors. It is incoherent for systems that, per Towards responsible artificial intelligence in education, reshape what counts as knowledge, or that, per UNESCO’s Inteligencia artificial y desinformación, reshape what counts as fact. A “partner” framing — the rarer one — would demand reciprocity: rights of refusal, contestation, audit. An “infrastructure” framing would demand the politics we apply to roads and water. The tool metaphor forecloses both.

The citizen’s task is not to choose a side in each tension but to notice that the side has usually been chosen for them — by whoever wrote the curriculum, sold the platform, or drafted the priority.

Power & Agency Analysis

Power in AI literacy operates through definition: who decides what citizens “need to know” shapes what remains invisible. Across this week’s 1,331 AI literacy items (within 6,252 total sources), the dominant pattern is striking — AI systems are repeatedly described as agents that “decide,” “detect,” “recommend,” and “govern,” while the humans who built, trained, deployed, and profited from them recede into the grammatical background. The framing matters because it teaches citizens to negotiate with a machine rather than with the people behind it.

How AI Is Portrayed

Read the documentation in citizens’ own hands and the agency assignments are everywhere. Microsoft’s content safety service “detects” harmful material and “evaluates” risk What is Azure AI Content Safety? - Azure AI services. OpenAI describes “skills” the model itself possesses Compétences dans ChatGPT - OpenAI Help Center. Even the critical literature that warns against AI tends to ascribe action to the system: deepfakes “deceive,” chatbots “hallucinate” — more than half of ChatGPT’s academic references in one audit were fabricated, but the verb attached is the model’s, not its makers’ ChatGPT’s Hallucination Problem: Study Finds More Than Half Of AI’s …. The grammar of agency teaches a lesson: when something goes wrong, the system did it. When something goes right, the system did that too. The vendor, in either case, is a stage manager invisible behind the curtain. UNESCO’s analysis of AI-driven disinformation is one of the few documents that consistently re-attributes — naming who deploys synthetic media against whom Inteligencia artificial y desinformación - UNESCO. The implication for citizen understanding is severe: a citizenry trained to argue with an “it” cannot hold a “they” accountable.

Who Defines Literacy

The institutional voices defining what counts as AI literacy are remarkably narrow. Vendor help pages set the operational vocabulary Prompt engineering for GitHub Copilot Chat. Federal agencies set the policy frame — the US Department of Education’s finalized AI priority shapes what literacy “officially” means downstream of any classroom The US Department of Education Just Finalized Its AI in Education … - AEI. Think tanks and policy shops fill in the rest AI’s future for students is in our hands - Brookings. What is conspicuously thin is democratic deliberation: citizens, workers, parents, and the targets of automated decisions appear as the population to be made literate, rarely as co-authors of what literacy should mean. A Spanish human-rights review of AI is among the few framings that puts the citizen as rights-holder rather than learner PDF Derechos Humanos E Inteligencia Artificial: Una Mirada Desde Los ….

What Metaphors Teach

The reigning metaphor is tool — neutral, hand-held, the user in command. The runner-up is threat — autonomous, looming, requiring defense. Both metaphors do political work. “Tool” obscures the supply chain: the data labor, the energy bill, the pricing leverage, the terms of service that change after you’ve already migrated your workflow. It encourages the question “how do I use this?” and discourages “who is using this on me?” “Threat” obscures the opposite end: it concentrates anxiety on the artifact (the deepfake) rather than the system that made fabrication cheap, leading to legislative proposals that chase outputs rather than incentives We Need Laws to Stop AI-Generated Deepfakes | Scientific American. A more useful citizen metaphor — barely present in the corpus — is infrastructure: something one is enrolled in, billed for, and dependent on, whose governance is a public question. UNESCO’s framing of synthetic media as a “crisis of knowing” gestures in this direction Deepfakes and the crisis of knowing - UNESCO, as does the ETH Zürich study on AI-governed information ecosystems When AI Governs (Dis)information: Five Lessons for ….

Citizen Agency

What power, then, does a citizen actually have? Individually, less than the literacy literature implies. No amount of prompt skill changes a model’s training data, a vendor’s pricing, or a regulator’s posture. The honest inventory is collective: pressing legislators on disclosure and liability, joining the small number of bodies that audit systems, refusing the framings — “the AI decided” — that launder responsibility. Knowledge here is not a shield against bad outputs; it is a vocabulary for naming the actors behind them. The literacy worth building is the one that lets a citizen rewrite the sentence with a human subject.

Failure Genealogy

Literacy failures differ from technical failures: they occur when citizens misunderstand what AI is, what it’s doing, or how to evaluate its output. The patterns are now well-documented enough to map — not as isolated mistakes, but as a genealogy of predictable breakdowns that begin with misplaced trust and end with concrete harm.

Where understanding fails

The most stubborn failure is treating fluent text as evidence of competence. A study of ChatGPT’s bibliographic output found that more than half of the references the system generated were fabricated or incorrect ChatGPT’s Hallucination Problem: Study Finds More Than Half Of AI’s … — and yet the citations look real, with plausible authors, journals, and dates. The detection problem compounds it. UNESCO’s work on synthetic media argues that the volume and quality of generated content has outpaced the public’s ability to identify it, eroding the baseline assumption that a video or recording corresponds to something that happened Deepfakes and the crisis of knowing - UNESCO. Citizens do not lack intelligence here; they lack the relevant heuristics. The cues we evolved for detecting deception — eye contact, vocal hesitation, internal contradiction in a story — were calibrated for humans, not for systems that produce confident prose without any underlying knowledge state.

The mirror failure is reflexive distrust: dismissing AI output wholesale, including the cases where it is reliable, which leaves citizens without working judgment in either direction. Both failures share a root: no mental model of what the system is actually doing.

What assumptions mislead

Three assumptions recur. First, that the interface implies the architecture — that a conversational system is reasoning conversationally, when it is performing statistical continuation. Second, that personalization implies neutrality — that because output is tailored to you, it represents you fairly. The Spanish-language human rights analysis from Unetxea documents how this assumption obscures the encoded interests of the model’s training and deployment Derechos Humanos E Inteligencia Artificial. Third, that visible safety features imply substantive safety. Vendor content-moderation products are marketed as guardrails — Microsoft’s Azure AI Content Safety frames itself as a system that “detects harmful user-generated and AI-generated content” What is Azure AI Content Safety? — and citizens reasonably read this as protection. What it actually is: a configurable filter, tuned by whoever deploys it, with whatever thresholds they choose. The protection is real; it is also bounded in ways the marketing does not advertise.

Consequences of gaps

The costs are not evenly distributed. Spanish reporting on digital violence documents how AI-generated sexual content, grooming, and synthetic harassment fall most heavily on young women and girls, who bear the harm of detection failures they had no role in creating Una violencia que nunca acaba. At the political scale, ETH Zürich’s Center for Security Studies argues that AI-mediated information environments shift the burden of verification onto individuals while removing the institutional scaffolding — editorial process, source provenance — that previously did that work When AI Governs (Dis)information. Scientific American’s argument for legal intervention against deepfakes rests on the same observation: individual literacy cannot scale to defend against industrial-grade synthetic content We Need Laws to Stop AI-Generated Deepfakes. The gap between what a citizen can verify and what a system can fabricate is the harm.

What would help

The literacy that prevents these failures is not technical fluency but something closer to source-criticism with new objects: knowing that fluency is not knowledge, that personalization is not neutrality, that filters are configured by someone. UNESCO’s framing of disinformation literacy emphasizes provenance and verification habits over tool-specific skills Inteligencia artificial y desinformación. Honestly, though: no individual literacy curriculum will close the asymmetry between a citizen and a generative system trained on the open web. Literacy reduces the failure rate. Regulation, provenance standards, and liability — none of which are the citizen’s job — close the rest.

Evidence Synthesis

Synthesizing the week’s analyses, the evidence on AI literacy for citizen participation points to a stubborn finding: comprehension of how these systems fail is now a precondition for ordinary civic life, not a specialty interest. This goes beyond technical skill — it is the capacity to refuse the frame a vendor or institution hands you, and to do so with reasons.

What evidence shows

The convergent finding across responsible-AI reviews is that literacy interventions work when they target system behavior, not user behavior. A systematic review in Nature synthesizing the responsible-AI-in-education literature finds that effective programs teach learners to interrogate outputs, provenance, and incentives rather than to “use AI better” Towards responsible artificial intelligence in education: a systematic …. Stanford’s SCALE Initiative, reviewing the K–12 evidence base, reaches a parallel conclusion: most measurable gains come from critical-evaluation tasks, not prompt fluency Understanding the Evidence Base on AI in K-12 Education | SCALE Initiative. UNESCO’s framing of the deepfake problem — that synthetic media erodes the shared epistemic floor democracies require — pushes literacy upstream into civic infrastructure Deepfakes and the crisis of knowing - UNESCO and Inteligencia artificial y desinformación - UNESCO. The base rate that justifies all this: a study reviewed at Study Finds documents that more than half of ChatGPT’s references were fabricated ChatGPT’s Hallucination Problem: Study Finds More Than Half Of AI’s …. Citizens who do not know this are not making informed choices; they are deferring.

Contested terrain

Where evidence diverges is on whether literacy can scale through vendor-supplied curricula. OpenAI’s own help documentation advertises “skills in ChatGPT” as a literacy primitive Compétences dans ChatGPT - OpenAI Help Center; Microsoft’s accessibility materials frame generative AI as inclusion infrastructure Inteligencia artificial generativa y accesibilidad | Microsoft Learn. These are useful documents and self-interested ones. ETH Zürich’s Center for Security Studies, reviewing AI’s role in disinformation governance, warns that letting platforms define the literacy agenda reproduces the asymmetry literacy was meant to correct When AI Governs (Dis)information: Five Lessons for …. Scientific American makes the harder argument: some harms — non-consensual deepfakes among them — require law, not pedagogy We Need Laws to Stop AI-Generated Deepfakes | Scientific American.

Across domains

Tool-specific literacy is real but narrow: knowing what Azure’s content-safety filters do and don’t catch What is Azure AI Content Safety? - Azure AI services, or what prompt engineering can and cannot rescue Prompt engineering for GitHub Copilot Chat, is operational knowledge, not civic understanding. The social-aspects dimension is where literacy bites: Maldita’s Spanish reporting on AI-enabled cyberharassment, grooming, and synthetic sexual imagery shows the harms landing disproportionately on women, minors, and the digitally precarious Una violencia que nunca acaba: ciberacoso, ‘grooming’, contenido sexual …. A child-safety review of generative systems documents the same pattern at scale PDF AI Child Safety Final Report. Brookings’ position — that the trajectory is decided by present policy choices, not technical inevitability — is the connective tissue AI’s future for students is in our hands - Brookings.

Gaps and uncertainty

We do not know how durable any of this is. The K–12 evidence base reviewed by SCALE is thin on long-term outcomes; cross-cultural data is sparser still. Whether literacy training transfers from one model generation to the next — when interfaces, refusals, and failure modes mutate — is unmeasured. Spanish-language human-rights frameworks gesture at the gap without filling it PDF Derechos Humanos E Inteligencia Artificial. And the youth-mental-health research underscores how context-dependent effects are New National Research Reveals How Context Shapes AI’s Impact on Youth ….

For citizens

Three evidence-based moves. First, individually: assume fabrication until you have checked, particularly with citations and attributions. Second, collectively: press for disclosure laws and liability regimes for synthetic media We Need Laws to Stop AI-Generated Deepfakes | Scientific American — pedagogy will not absorb harm that statute should be preventing. Third, structurally: refuse curricula written by the firms whose products they evaluate. Literacy that cannot name its vendors is marketing.