AI Literacy for Citizen Participation Report

Analysis of 1,355 AI literacy sources this week reveals a discourse focused on teaching citizens to defend themselves against AI outputs — scams, deepfakes, hallucinated facts, electoral manipulation — while neglecting the question of how citizens might shape the systems being deployed on them. The citizen-as-participant framing appears in roughly one in ten sources; the dominant frame, by a wide margin, is citizen-as-target: someone to be inoculated, warned, or trained to spot a fake.

1. The Landscape

“AI literacy” in this week’s corpus is overwhelmingly a defensive vocabulary. UNESCO frames it as a response to a “crisis of knowing” produced by synthetic media Deepfakes and the crisis of knowing; NewsGuard’s latest audit reports that the rate at which chatbots repeat false claims has roughly doubled Le taux de fausses informations répétées par les chatbots d’IA a …; the BBC traces anti-immigration AI videos to overseas operators seeding domestic feeds Anti-immigration AI videos traced to overseas fakers, BBC finds. The implicit citizen here is a consumer of dubious content, not a participant in the decisions that produced the content pipeline. Even the more constructive corner of the literature — prompt engineering as a “21st-century skill” Frontiers | Prompt engineering as a new 21st century skill — defines competence as the ability to extract better outputs from a vendor’s system, not to question whether the system should be there at all.

2. Whose Literacy

The teachers, in this corpus, are mostly the same actors who built the problem. Microsoft publishes the prompting curriculum Create effective prompts for generative AI training tools - Training; OpenAI explains, on its own site, why its models hallucinate Por qué los modelos de lenguaje alucinan - OpenAI. Intergovernmental bodies — UNESCO above all — issue the anti-disinformation frameworks Inteligencia artificial y desinformación - UNESCO. Independent civic voices exist — Renaissance Numérique’s October 2025 report on deploying AI literacy is a notable case — but they sit alongside a much louder vendor-and-platform chorus. The asymmetry matters: when the company selling the system also writes the literacy curriculum, “critical thinking about AI” gets defined as informed use, not informed refusal.

3. What’s Being Taught

Three thematic clusters carry the week. The largest is detection: spotting AI-generated scams aimed at older citizens Estafas y fraudes generados con IA, recognising synthetic financial-aid fraud How scammers are using AI to steal college financial aid, evaluating electoral content Elecciones: Inteligencia artificial y desinformación electoral. The second is operational fluency — prompting, summarising, translating. The third, smaller, is structural awareness: a UCSD analysis showing that governments shape chatbot speech by shaping the training web Governments May Shape What AI Chatbots Say; a Swiss legal case on algorithmic injustice in Gothenburg’s school-assignment system Un cas pratique d’injustice algorithmique; the Cyberbullying Research Center’s national finding that nearly half of US teens report harm from conversational chatbots Conversational AI Chatbots and US Teens. That third cluster is where citizen agency actually lives — and it is the smallest.

4. What’s Missing

Almost entirely absent: how a citizen contests an algorithmic decision made about them, how data-protection rights translate into a workable refusal, how municipal procurement of AI systems is debated before deployment, and how communities surveilled by EdTech monitoring Public Schools, Private Eyes or by workplace tools might organise against it. Also missing: the literacy needs of populations that vendor curricula do not address — older adults beyond scam-detection, disabled users beyond accessibility marketing L’IA sort les personnes sourdes du monde du silence, non-English speakers, and citizens whose first encounter with AI is not a chatbot but a denied benefit. A literacy fit for participation would treat governance, not detection, as the core competency.

Core Tensions

The phrase “AI literacy” conceals genuine tensions about what citizens need to know and why. Survey the week’s evidence and the contradictions are not gaps to be filled by more curriculum — they are contested terrain about who gets to define competence, whose risks count, and whether the citizen’s task is to use these systems well or to judge them well. The most fundamental tension running through everything else: prompt fluency versus epistemic sovereignty — is a literate citizen one who can extract value from a chatbot, or one who can refuse it on informed grounds?

That tension is not abstract. Microsoft’s own training modules define literacy operationally as the capacity to “create effective prompts” and steer model outputs toward desired ends Create effective prompts for generative AI training tools - Training, and a strand of pedagogical research has elevated prompt engineering to the status of a “21st century skill” Frontiers | Prompt engineering as a new 21st century skill. The implicit anthropology here is the citizen-as-operator: someone whose civic competence is measured by throughput. Set this against NewsGuard’s audit finding that the rate of false claims repeated by major chatbots has roughly doubled Le taux de fausses informations répétées par les chatbots d’IA a …, and OpenAI’s own concession that hallucination is structural to how language models are trained Por qué los modelos de lenguaje alucinan - OpenAI. A citizen who has mastered prompts but cannot recognize when the machine is confidently wrong is, in civic terms, less literate, not more.

The second tension: individual vigilance versus collective governance. AARP teaches retirees to spot AI-generated scams Estafas y fraudes generados con IA: ¿Cómo detectarlos? - AARP; AP reports on synthetic identities draining federal financial aid How scammers are using AI to steal college financial aid; the BBC traces anti-immigration AI videos to overseas operators seeding domestic feeds Anti-immigration AI videos traced to overseas fakers, BBC finds. The reflex response is to train sharper individual eyes. But UCSD researchers point out that governments can shape what chatbots say by shaping the web the chatbots learn from Governments May Shape What AI Chatbots Say by Shaping the Web They Learn From, and UNESCO frames deepfakes as a crisis of knowing rather than a crisis of spotting Deepfakes and the crisis of knowing - UNESCO. Asking each citizen to forensically authenticate the information environment is a category error — it offloads onto individuals a problem of infrastructure and procurement.

The third tension: AI use versus AI refusal as legitimate civic positions. The Göteborg case — where automated school assignment in Sweden produced documented algorithmic injustice — illustrates what is at stake when “AI literacy” excludes the option of saying no Un cas pratique d’injustice algorithmique : l’attribution automatisée …. Cyberbullying.org’s national study of US teens, with nearly half reporting harm from conversational chatbots Conversational AI Chatbots and US Teens: Nearly Half …, and New America’s mapping of EdTech monitoring infrastructure Public Schools, Private Eyes: How EdTech Monitoring Is Reshaping Public …, both suggest that competent participation sometimes means competent withdrawal. A literacy that cannot articulate refusal is consumer training in civic disguise.

These tensions are held in place by metaphor. Across the week’s corpus, AI is figured overwhelmingly as a tool (some 304 instances in our mapping), occasionally as a threat (52), and almost never as a partner (7). The tool metaphor implies a user with intentions and a thing that does what it is told — which obscures the fact that the “tool” is trained on a contested commons, optimized by a vendor, and shaped by procurement decisions citizens never voted on. The threat metaphor inverts agency but keeps the citizen passive: now the thing acts and you defend. Only the partner metaphor — minoritarian, almost absent — forces the question of terms: on what conditions, with what recourse, accountable to whom? UNESCO’s framing of misinformation explicitly invokes shared responsibility across platforms, states, and publics Inteligencia artificial y desinformación - UNESCO, and Québec’s OBVIA framework treats education against AI-amplified disinformation as a collective project, not a personal skill PDF Éduquer contre la désinformation amplifiée par l’IA et l’hypertrucage ….

The citizen’s first literacy task, then, is to notice which metaphor is doing the work in any given pitch — and to ask what it is permitting the speaker to leave unsaid.

Power & Agency

Power & Agency Analysis

Power in AI literacy operates through definition: who decides what citizens “need to know” shapes what remains invisible. The dominant framings this week distribute agency unevenly — chatbots “hallucinate,” algorithms “decide,” deepfakes “deceive” — while the humans who built, deployed, and profited from these systems recede into grammatical passivity. The framing matters because it tells citizens where to look for accountability, and where not to.

How AI is portrayed

Read the week’s coverage closely and a pattern emerges in the verbs. OpenAI’s own explanation of why language models invent facts treats hallucination as a structural property of the model rather than a product decision — the system “hallucinates” the way weather “happens” Por qué los modelos de lenguaje alucinan - OpenAI. NewsGuard’s audit reporting that chatbots’ rate of repeating false claims has doubled uses similar grammar: the chatbots repeat, the rate climbs Le taux de fausses informations répétées par les chatbots d’IA a …. Even a careful Harvard framework on AI inaccuracy classifies failure modes of “the system” rather than the labor practices, training-data choices, and release schedules behind them.

When a Swedish municipality’s algorithm assigned children to schools in ways that systematically disadvantaged some families, the press initially called it an “algorithmic injustice” — as if the algorithm did it Un cas pratique d’injustice algorithmique : l’attribution automatisée …. It did not. A procurement committee chose it; civil servants deployed it; elected officials declined to audit it. Citizen literacy starts with translating those sentences back into the active voice.

Who defines literacy

The producers of AI are also the largest producers of AI-literacy curriculum. Microsoft Learn modules on prompting and on “inclusive learning environments” present competence with generative tools as the substance of literacy itself Create effective prompts for generative AI training tools - Training, Utiliser des outils IA pour créer un environnement d’apprentissage …. Peer-reviewed work then ratifies the move, recasting prompt engineering as a “new 21st-century skill” Frontiers | Prompt engineering as a new 21st century skill. Notice the silent substitution: literacy as the right to interrogate a technology becomes literacy as fluency in its interface. A citizen who knows how to refuse a system — to demand an explanation, to opt out, to organize against deployment — is not certifiable by any vendor.

What metaphors teach

The “tool” metaphor dominates this week’s evidence base by a wide margin, with “threat” a distant second and “oracle” hovering underneath both. Each metaphor is a syllabus.

Tool suggests neutrality and user control — pick it up, set it down. The framing obscures that the tool listens, logs, and routes value to its maker. EdTech monitoring software now installed across American schools is sold as a safety tool, while functioning as a surveillance infrastructure on minors Public Schools, Private Eyes: How EdTech Monitoring Is Reshaping Public …. Threat concentrates attention on bad actors — deepfake scammers, foreign election interference, fraudulent aid applicants How scammers are using AI to steal college financial aid, Anti-immigration AI videos traced to overseas fakers, BBC finds, Estafas y fraudes generados con IA: ¿Cómo detectarlos? - AARP. It is true and it is useful, but it routes anger toward the criminal margin while the legal core of the industry sets the conditions. Oracle — the chatbot as something to ask — naturalizes the idea that authority can be queried, and that the answer counts. UNESCO names the consequence directly: a “crisis of knowing” in which the public’s epistemic floor is being remodeled by whoever owns the training pipeline Deepfakes and the crisis of knowing - UNESCO. Researchers at UC San Diego push the point further: governments can shape what chatbots say by shaping the web those chatbots ingest Governments May Shape What AI Chatbots Say by Shaping the Web They Learn From. The oracle has an upstream.

Citizen agency

What can a non-specialist citizen actually do? More than the vendor curriculum implies, less than the panic narrative suggests. Individual capacities — lateral reading, asking who paid for the deployment, refusing to trust a confident sentence — are necessary and insufficient What Makes Students (and the Rest of Us) Fall for AI Misinformation?, Lutter contre les fake news générées par IA : entretien avec Chine Labbe. The harder agency is collective: procurement decisions in your municipality, audit requirements in your electoral rules Elecciones: Inteligencia artificial y desinformación electoral, the right to challenge an automated assignment that affects your family Un cas pratique d’injustice algorithmique : l’attribution automatisée …. Literacy that stops at the individual screen is a literacy of resignation. The week’s evidence — drawn from a pool of 6,327 sources — points at the same conclusion from many angles: knowing how AI works is the entry fee, not the seat.

Failure Genealogy

Literacy failures differ from technical failures: they occur when citizens misunderstand what AI is, what it’s doing, or how to evaluate it. Across this week’s evidence, five recurring patterns emerge — over-trust, detection blindness, unwitting disclosure, uncritical acceptance, and the missing refusal — and they cluster not around exotic edge cases but around the most ordinary encounters: a chatbot answer, a video on a feed, a financial-aid form, a confident summary.

Where understanding fails

The dominant failure is over-trust calibrated to fluency rather than accuracy. NewsGuard’s latest audit found that the rate of false claims repeated by major AI chatbots has roughly doubled year over year, with assistants now confidently asserting fabricated political and health claims they previously declined to answer Le taux de fausses informations répétées par les chatbots d’IA a …. OpenAI’s own explanation of why models hallucinate concedes a structural point most users miss — these systems are rewarded for sounding right, not for admitting ignorance Por qué los modelos de lenguaje alucinan - OpenAI. Citizens reading a polished paragraph have no surface cue to distinguish retrieval from invention.

Detection failures track the same fluency trap. The BBC traced a network of anti-immigration AI videos circulating as authentic footage to overseas operators, sometimes monetized on platforms whose moderation lagged behind the generation Anti-immigration AI videos traced to overseas fakers, BBC finds. UNESCO frames this as a “crisis of knowing” rather than a crisis of fakery — the issue is not that fakes exist but that the cost of verifying any single image has risen above the time most readers will spend Deepfakes and the crisis of knowing - UNESCO. Edweek’s reporting adds a less flattering finding: it is not only teenagers who fall for AI misinformation; adults with strong prior beliefs fall harder, because confirmation does the work that scrutiny would have done What Makes Students (and the Rest of Us) Fall for AI Misinformation?.

What assumptions mislead

Three assumptions recur. First, that a chatbot is a search engine with manners — when in fact its outputs reflect what governments and large publishers shape the training corpus to contain, a point made forcefully by UC San Diego’s recent work on how state actors can tilt model outputs by tilting the web Governments May Shape What AI Chatbots Say by Shaping the Web They Learn From. Second, that AI-generated scams will look amateurish; the AARP and Associated Press both document fraud rings using cloned voices and synthetic identities to drain retirement accounts and steal federal financial aid disbursements at scale Estafas y fraudes generados con IA, How scammers are using AI to steal college financial aid. Third, that “the algorithm” is neutral arbitration. Göteborg’s automated school-assignment system, examined this month, produced systematically unjust placements that families had no procedural language to contest Un cas pratique d’injustice algorithmique.

Consequences of gaps

The costs are not evenly distributed. UNESCO’s reporting on AI and disinformation, and the Freiheit Foundation’s election analysis across the Southern Cone, both find that the populations least equipped to verify — older voters, lower-literacy publics, language minorities — absorb the largest share of the manipulation Inteligencia artificial y desinformación - UNESCO, Elecciones: Inteligencia artificial y desinformación electoral. A Cyberbullying Research Center national study of US teens reports that nearly half have experienced harm from conversational AI — a generational data point that maps onto a wider pattern of unwitting disclosure and parasocial over-attachment Conversational AI Chatbots and US Teens. The collective cost is harder to itemize: a citizenry that cannot tell whether an electoral video is real underwrites every other failure downstream.

What would help

Honest literacy work would start from the failures, not from the tools. Québec’s Obvia consortium argues for teaching against disinformation by having citizens produce small fakes themselves, on the grounds that nothing dispels the magic faster Éduquer contre la désinformation amplifiée par l’IA. Chine Labbé’s interview with Synthmedia stresses provenance over detection — teach where claims come from, not how to spot pixels Lutter contre les fake news générées par IA. Both are honest about the ceiling: when the model is rewarded for confidence and the platform is rewarded for engagement, individual literacy is a seatbelt, not a brake.

Evidence Synthesis

Synthesizing roughly thirty analyses across the AI literacy beat, the evidence converges on an uncomfortable finding: the skills citizens most need are not the ones most programs teach. Prompt-craft and tool-fluency dominate the curriculum; the harder capacities — knowing when a system is lying, knowing who shaped what it says, knowing when to disengage — remain underdeveloped. This goes beyond technical skill: AI literacy, as the evidence describes it, is closer to a civic competence under conditions of contested reality than to a software proficiency.

What the evidence shows

The strongest convergence sits around three findings. First, generative models confabulate by design — OpenAI’s own account frames hallucination as a statistical inevitability of next-token prediction under reward structures that penalize abstention Por qué los modelos de lenguaje alucinan - OpenAI, and a Harvard Kennedy School framework treats AI inaccuracy as a structural rather than incidental property New sources of inaccuracy? A conceptual framework for studying AI …. Second, the falsehood rate is rising in production: NewsGuard’s audit found chatbots’ propagation of false claims has roughly doubled year over year Le taux de fausses informations répétées par les chatbots d’IA a …. Third, the most effective pedagogical responses are constructive rather than defensive — Université Laval researchers report durable gains when learners build deepfakes themselves before being asked to detect them Deepfakes: créer du faux contenu pour comprendre la désinformation et …, a finding echoed in the OBVIA framework for educating against AI-amplified disinformation PDF Éduquer contre la désinformation amplifiée par l’IA et l’hypertrucage … and a Harvard interview on critical AI pedagogy Teaching Students to Think Critically About AI.

Contested terrain

Evidence diverges sharply on what counts as literacy at all. Microsoft’s training catalogue treats prompt engineering as the core competence Create effective prompts for generative AI training tools - Training, and a Frontiers in Education paper argues prompt engineering is itself a 21st-century skill Frontiers | Prompt engineering as a new 21st century skill. UNESCO’s framing pulls the opposite direction, treating literacy primarily as epistemological resilience under deepfake conditions Deepfakes and the crisis of knowing - UNESCO. The split matters: vendor-defined literacy makes citizens more productive users of specific products; civic-defined literacy makes them harder to manipulate by anyone, vendors included. A scoping review of generative-AI misinformation finds the two are rarely taught together GenAI and misinformation in education: a systematic scoping … - Springer.

Across domains

Tool-specific literacy now has to extend to provenance: a UC San Diego analysis shows governments can shape chatbot outputs by shaping the training web itself Governments May Shape What AI Chatbots Say by Shaping the Web They Learn From, meaning “what the model says” is downstream of upstream political choices most users never see. The social-aspects dimension is sharper still: BBC tracing of anti-immigration AI video farms Anti-immigration AI videos traced to overseas fakers, BBC finds, Freiheit’s documentation of synthetic electoral disinformation across the Southern Cone Elecciones: Inteligencia artificial y desinformación electoral, AARP’s catalogue of AI-generated fraud against older adults Estafas y fraudes generados con IA: ¿Cómo detectarlos? - AARP, and AP reporting on synthetic identities draining college financial-aid systems How scammers are using AI to steal college financial aid all describe the same shift: literacy is now a personal-security competence, not only a cognitive one.

Gaps and uncertainty

The evidence is thinner than it looks. Almost no longitudinal data exists on whether literacy interventions hold beyond a semester. Cyberbullying.org’s national study finds nearly half of US teens have already experienced harm from conversational chatbots Conversational AI Chatbots and US Teens: Nearly Half …, but we do not know which protective competencies reduce that harm. Most curricula are evaluated by their designers. And the question of whether literacy can scale faster than the models themselves remains open — a question UNESCO raises but does not answer Inteligencia artificial y desinformación - UNESCO.

For citizens

Two evidence-based moves are within individual reach: treat every consequential AI output as a claim requiring a second source, and learn the production side — even briefly — to inoculate against the consumption side Les « deepfakes » : Comment donner aux jeunes les moyens de lutter …. What individuals cannot do alone is govern the upstream choices — training data, deployment incentives, platform liability — that determine what the systems say in the first place. That part requires collective action, and the evidence is unambiguous that no amount of personal vigilance substitutes for it.

AI Literacy for Citizen Participation Report

1. The Landscape

2. Whose Literacy

3. What’s Being Taught

4. What’s Missing

Core Tensions

Power & Agency

Power & Agency Analysis

How AI is portrayed

Who defines literacy

What metaphors teach

Citizen agency

Failure Genealogy

Failure Genealogy

Where understanding fails

What assumptions mislead

Consequences of gaps

What would help

Evidence Synthesis

Evidence Synthesis

What the evidence shows

Contested terrain

Across domains

Gaps and uncertainty

For citizens

References