AI Literacy for Citizen Participation Report

State of the Discourse: AI Literacy for Citizen Participation

Analysis of 916 AI literacy sources this week — drawn from a corpus of 4,400 — reveals a discourse still bolted to the classroom, obsessed with detection and cheating, while the citizen who is acted upon by AI barely appears. The citizen-as-participant framing surfaces in perhaps a tenth of the material; most sources treat literacy as a school-management problem — how to catch a student using ChatGPT — rather than a civic condition every adult now inhabits whether or not they ever touch a chatbot.

1. The Landscape

Watch the move that defines the field: “AI literacy” gets defined by the institutions with the most to lose from getting it wrong. Stanford’s teaching guidance frames literacy as a set of competencies for using and evaluating models Comprender la alfabetización en IA | Teaching Commons; ETS, a testing company, frames it as something assessable — and therefore sellable PDF Opportunities and Challenges for Asessing Digital and AI Literacies - ETS. The prior three iterations of this report kept litigating whether literacy should balance job-readiness against ethics. That debate is now a distraction. The delta this week: the sharpest evidence is no longer about workforces at all. It is about a public living inside a synthetic information environment — deepfaked municipal candidates Municipales 2026 : IA, deepfakes et désinformation, la démocratie …, and a “crisis of knowing” that UNESCO names bluntly Deepfakes and the crisis of knowing.

2. Whose Literacy

The discourse is overwhelmingly expert-to-citizen, top-down. Testing firms, government ministries Cadre d’usage de l’IA en éducation, and platforms narrate what ordinary people should know; ordinary people rarely narrate back. Where citizen voices do appear, they arrive as data points about harm rather than as agents — teenagers surveyed on trust PDF 2025 Teens, Trust, and Technology in the Age of AI - Common Sense Media, or children studied for their susceptibility to synthetic content Children’s susceptibility to content generated by artificial …. The asymmetry matters: a literacy defined entirely by institutions tends to produce compliant users, not skeptical citizens.

3. What’s Being Taught

Two clusters dominate, and both privilege the wrong verb. The first is detection — the fantasy that a tool can tell you what’s real. It cannot: AI detectors falsely accuse international students at documented rates AI Detection Tools Falsely Accuse International Students of Cheating, teachers deploy them knowing they’re unreliable AI detection tools are unreliable. Teachers are using them anyway : NPR, and families are now filing suit over false accusations A Palo Alto high schooler was accused of AI cheating. His family filed …. The second is use — prompt technique, productivity. What’s thin is the middle: understanding when AI is being run on you. Prompt-injection attacks Prompt Injection Attacks: Examples and Defences, documented model deception AI deception: A survey of examples, risks, and potential solutions, and AI-enabled fraud El auge de la delincuencia facilitada por la IA - Informe name threats a citizen faces as a target, not a user.

4. What’s Missing

Three absences stand out. First, data rights and consent — the discourse teaches evaluation of outputs while ignoring who harvested the inputs. Second, differential harm: MIT researchers found chatbots deliver less-accurate information to vulnerable users Study: AI chatbots provide less-accurate information to …, yet literacy frameworks assume a uniform learner. Third, governance participation — almost nothing equips citizens to contest AI systems in elections or public administration, even as generative tools reshape European campaigns IA générative et élections européennes : opportunité ou péril ?. The UN warns that children now grow up inside AI before anyone has decided the rules Millones de niños crecen con la inteligencia artificial antes de que …. A literacy that only teaches use, and never refusal or redress, is training subjects — not participants.

Core Tensions

The concept of “AI literacy” conceals genuine tensions about what citizens need to know and why. When you ask what a person must understand to participate in public life now, you get answers pulling in opposite directions—and the disagreement isn’t a knowledge gap to be filled by a better curriculum. It’s contested terrain, and the contest is over power. Where prior framings treated AI literacy as a workforce-readiness project balanced against ethical awareness, the sharper question this week is narrower and more political: what does a citizen—not an employee, not a student—need, and who gets to decide?

Technical skill versus critical understanding. The dominant push treats literacy as competence: prompt well, verify outputs, know the tool’s limits. Stanford’s own teaching guidance frames literacy largely around understanding capabilities and using systems effectively Comprender la alfabetización en IA | Teaching Commons. But competence with a tool is not the same as judgment about a system. The New America field survey found practitioners insisting that “digital literacy in the age of AI” has to mean interrogating who builds these systems and whose interests they serve, not just operating them Digital Literacy in the Age of AI: Voices from the Field. ETS, which is in the business of measuring such things, concedes that assessing critical AI literacy is far harder than assessing procedural skill—which tells you which one institutions will quietly default to measuring PDF Opportunities and Challenges for Asessing Digital and AI Literacies - ETS. What’s at stake for citizens: a population fluent in prompting but illiterate in provenance is easy to govern and easy to fool.

Protection FROM versus empowerment WITH. A large slice of the literature frames citizens—especially young ones—as targets to be shielded. UNESCO’s account of deepfakes describes a “crisis of knowing,” where the problem is not that people can’t use AI but that they can no longer trust what they see Deepfakes and the crisis of knowing. Studies document children’s heightened susceptibility to AI-generated content Children’s susceptibility to content generated by artificial …, and MIT researchers found chatbots deliver less accurate information to the most vulnerable users Study: AI chatbots provide less-accurate information to …. Protection is real and necessary. But a purely protective literacy produces wary consumers, not participating citizens—and the two are not the same person. Consumer literacy asks “how do I avoid being scammed” as AI-enabled crime scales up El auge de la delincuencia facilitada por la IA - Informe. Citizen literacy asks “who is allowed to deploy this against a population, and who answers for it.”

Individual competency versus collective governance. Here is the move to watch. Almost every literacy framework locates the burden on the individual: learn to spot the deepfake, verify the source, guard your data. Meanwhile the deepfakes flooding the 2026 midterms across dozens of states AI Deepfakes Are Flooding the 2026 Midterms. 26 States Scramble, and the disinformation tracked through France’s municipal races Municipales 2026 : IA, deepfakes et désinformation, la démocratie …, are structural problems no amount of personal vigilance resolves. When responsibility is framed as individual literacy, the actors who build and profit from these systems recede from view. That’s not an accident of pedagogy; it’s a distribution of blame.

The metaphor does the arguing. Across this week’s evidence, AI is overwhelmingly cast as a tool (304 instances) and occasionally a threat (52). Both framings position the citizen as a lone user—wielding or defending. Almost no one (7 instances) frames AI as a partner, and vanishingly few frame it as infrastructure or governed institution. This matters because the metaphor sets the literacy. “Tool” implies your job is skillful operation; “threat” implies defensive hygiene. Neither implies a vote, a regulation, or a right. France’s national framework for AI in education at least names collective norms rather than individual competence alone Cadre d’usage de l’IA en éducation, and the 2026 AI Index documents how governance is lagging capability The 2026 AI Index Report | Stanford HAI. The literacy a citizen actually needs is the one none of the dominant metaphors supply: the ability to ask who owns the tool, who wrote the rules, and who is exempt from them. You can test any AI-literacy program against that single question—and most, you’ll find, quietly answer “you, alone.”

Power & Agency Analysis

Power in AI literacy operates through definition: whoever decides what citizens “need to know” also decides what stays invisible. Across this week’s 916 category articles, drawn from a corpus of 4,400, the pattern is lopsided. When AI does something consequential, the language overwhelmingly casts it as a tool — the dominant metaphor by a wide margin — while the far rarer threat framing (a small fraction of instances) shows up almost exclusively around crime and children. Both framings, in different ways, teach citizens to look away from the humans making decisions.

How AI is portrayed

Watch the grammar of agency. When an AI detector flags a student, the sentence usually reads “the system detected AI-generated text” — the tool decides, the human recedes. But the Bloomberg analysis of false cheating accusations and reporting that detection tools falsely accuse international students reveal the humans who bought the tool, set the threshold, and chose to act on its output. When a family in Palo Alto filed a claim after an AI-cheating accusation, they were not suing an algorithm; they were confronting an institution that chose to trust one. Yet NPR found teachers using these unreliable tools anyway — a human decision the “the AI flagged it” phrasing conveniently launders. Agency assigned to the machine is accountability removed from the person. For citizens, the lesson is corrosive: it trains you to argue with an output rather than with the party who deployed it.

Who defines literacy

The definition of “AI literacy” is not neutral, and the people writing it are not a cross-section of the public. Frameworks arrive from testing companies (ETS’s assessment report), elite universities (Stanford’s Teaching Commons), education ministries (France’s official usage framework), and index-makers (Stanford HAI’s 2026 report). Each has an interest in what counts as competence — the tester wants something scoreable, the ministry something teachable, the vendor something adjacent to its product. New America’s “voices from the field” is unusual precisely because it asks practitioners rather than dictating to them. Notice who is almost never in the room defining the term: the citizen expected to become literate. Literacy defined for you, without you, is a curriculum of compliance.

What metaphors teach

The “tool” metaphor, at 304 instances the runaway leader, carries a quiet politics: a tool is neutral, obedient, and only as good as its user — so any harm becomes user error. That framing obscures what research on AI deception and prompt-injection attacks make plain: these systems behave in ways their users neither intend nor control. A hammer does not manipulate the carpenter. The “threat” metaphor, reserved for AI-enabled crime and AI reshaping child exploitation, does the opposite work — it externalizes danger onto a monstrous “it,” licensing emergency powers while leaving the platforms that host the abuse unnamed. UNESCO’s account of deepfakes and the crisis of knowing is sharper: the threat is not the fake itself but the erosion of any shared basis for verifying it. Critical metaphor literacy means asking, each time, what a given figure of speech is hiding — and who benefits from the hiding.

Citizen agency

So what power do citizens actually hold? Less than the “empowerment” rhetoric promises, more than the passive framings imply. The individual defenses are real but thin: knowing that chatbots give less accurate answers to vulnerable users, or that generative systems can flood an election with synthetic content — a worry documented across European elections and the 2026 French municipals — makes you a harder target, not a safe one. Real leverage is collective: it lives in the family that files a claim, the state legislatures scrambling to regulate deepfakes, the demand that institutions justify the tools they deploy rather than hide behind them. Knowledge here is not personal protection; it is the precondition for holding power to account. The citizen’s literacy that matters least is knowing how to prompt. The one that matters most is knowing whom to name.

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. Our analysis of this week’s 4,400 sources documents a recurring shape to these breakdowns — not ignorance of how models work internally, but a mismatch between what people assume AI can do and what it actually does to them.

Where understanding fails

The dominant failure is not gullibility but calibration. Citizens swing between over-trust and under-trust, often in the same encounter. Over-trust shows up when people accept fluent output as accurate — a problem sharpened by evidence that chatbots deliver less-accurate information to vulnerable users, meaning the people least equipped to catch an error are the ones most likely to receive one. Under-trust shows up as its mirror: the assumption that a machine can reliably detect machine output. It cannot. AI detectors produce false positives that falsely accuse students of cheating, and the harm lands unequally — detection tools disproportionately flag international students whose non-native phrasing reads as “synthetic.” Institutions know the tools are unreliable and keep using them anyway. The detection gap runs the other direction too: UNESCO’s work on deepfakes and the crisis of knowing documents that most people cannot reliably distinguish synthetic video from real, and that confidence in one’s ability to do so is itself uncorrelated with actual skill.

What assumptions mislead

Three assumptions do most of the damage. First, that detection is a solvable technical problem — that some tool, somewhere, will sort true from false so the citizen doesn’t have to. The documented unreliability of detectors makes this a false hope. Second, that AI systems are neutral instruments rather than persuasive ones — an assumption undercut by a survey of AI deception showing systems that learn to mislead when it serves their objective. Third, that inputs are private. Citizens type medical worries, political views, and financial details into interfaces they treat as search bars, unaware that the same generative capacity powering these tools is being turned to crime at scale and that systems remain vulnerable to prompt injection, where hidden instructions hijack a model’s behavior without the user ever knowing.

Consequences of gaps

The costs are borne unevenly, and disproportionately by those with the least room to absorb them. Adolescents face documented risks to well-being, per the APA’s health advisory on AI and adolescent well-being, while younger children show heightened susceptibility to AI-generated content they cannot yet identify as synthetic. At the civic scale, the failure compounds: deepfakes are flooding the 2026 midterms across 26 states, and French analysts warn that generative content in the 2026 municipal elections erodes the shared factual ground democracy requires. When enough citizens cannot evaluate what they see, the harm stops being individual and becomes a collective loss of the ability to know anything together.

What would help

The failure analysis points away from detection and toward disposition. What prevents these breakdowns is not a better tool but a habituated skepticism — the reflex to ask who made this, to move me toward what, before asking is it real. New America’s voices from the field and the ETS assessment work both converge on process-oriented judgment over output-checking. The honest limitation: no disposition survives an information environment engineered to exhaust it. Literacy reduces the failure rate; it does not repeal the asymmetry between a citizen’s attention and an industry’s output.

Evidence Synthesis

Synthesizing this week’s 4,400 sources, the evidence on AI literacy points to an uncomfortable conclusion: the skill citizens most need is not learning to use the tools but learning to distrust the machinery built to police them. This goes beyond technical competence. The prior installments in this series treated literacy as workforce preparation and misinformation defense; the delta this week is that the defensive infrastructure itself — detectors, deepfake flags, authenticity signals — has become a second-order source of harm citizens must now be literate about.

What the evidence shows

Start with the failure that keeps recurring. AI-detection tools do not work reliably, and the people deploying them know it. NPR documents teachers using detectors they concede are unreliable, while Bloomberg tracks the human cost — students falsely accused of cheating on the word of software. The Markup found these false positives fall hardest on international students, whose non-native phrasing reads as “machine-like” to a classifier. A Palo Alto family took the accusation to court. Law librarians now catalog the false-positive problem as a known defect. The convergent finding: a literate citizen should treat “AI-detected” as an accusation requiring evidence, not a verdict.

The second convergence is about the epistemic ground itself. UNESCO’s account of the crisis of knowing argues that deepfakes damage the public less by fooling us and more by giving everyone permission to disbelieve anything inconvenient — the “liar’s dividend.” That maps directly onto electoral evidence: France’s municipal campaigns and the 2026 U.S. midterms, where 26 states scrambled to legislate against synthetic media faster than they could define it.

Contested terrain

“Literacy” fractures precisely here. The Institut Montaigne frames generative AI in European elections as opportunity or peril — an unresolved either/or, not a settled risk. One camp treats literacy as detection skill (spot the fake); another, better supported by the UNESCO evidence, treats it as provenance habit (verify the source, distrust the frame). These are different educations. Stanford’s teaching guide and ETS’s assessment work concede the field cannot yet measure the thing it names.

Across domains

Tool-specific literacy has a concrete edge: prompt injection, where hidden instructions hijack a model’s behavior, is now a documented attack class — meaning the citizen’s assistant can be turned against them. On the social-aspects side, MIT found chatbots deliver less-accurate information to vulnerable users, making literacy an equity issue rather than a personal virtue: the people least equipped to fact-check get the worst answers. New America’s field voices and the 2026 AI Index both register the same distributional gap.

Gaps and uncertainty

We do not know whether literacy interventions durably change behavior, or whether they merely teach people to perform suspicion. The Common Sense Media survey of teen trust shows attitudes shifting, but attitude is not competence. The UN’s warning that millions grow up with AI before any framework exists underscores how far behind the evidence base runs.

For citizens

Individually: treat detection verdicts as claims, not proof; verify provenance before belief; assume your vulnerability is being priced in. Collectively: the false-positive burden, the liar’s dividend, and the deepfake election problem are structural — they require rules on who may accuse you by algorithm, not more personal vigilance. Literacy is necessary. It was never going to be sufficient.