AI Literacy for Citizen Participation Report

Analysis of 1,069 AI literacy sources drawn from a 5,033-article week reveals a discourse organized almost entirely around threat and defense — deepfakes, voice-cloning fraud, monitored teenagers, poisoned search results — while the citizen as an active participant in AI governance barely registers. Across the pieces surfacing this week, the citizen-as-participant framing appears in fewer than one in five; most treat literacy as a protective reflex, a way to keep people from being deceived, surveilled, or defrauded, rather than a capacity to act.

Watch that move, because it quietly changes what “literacy” is for.

The Landscape

The dominant definition of AI literacy right now is defensive. To be literate is to detect: to spot the synthetic video, catch the cloned voice, recognize when a chatbot is laundering a falsehood. The week’s most-cited work runs this way — employers briefed on “information disorder” Information disorder in the age of AI: what it means for employers, OSINT analysts explaining detection Deepfakes et désinformation 2026 : comment les détecter (analyste OSINT …, the mechanics of a $1.8 billion voice-cloning fraud wave that defeats identity verification Clonage Vocal IA : La Crise à 1,8 Md$ qui Brise le KYC. This is real and urgent. But it defines the citizen by what can be done to them. The rarer framing — literacy as the ground for participating in decisions about how AI is deployed — surfaces only where the discourse touches voting: chatbots giving electoral advice that “could help and harm voters” Chatbot Voting Advice Could Help and Harm Voters. Let’s Regulate Accordingly., or the reminder that in elections the humans, not the models, remain the thing to worry about Generative AI and elections: why you should worry more about humans ….

Whose Literacy

The teaching voice this quarter belongs overwhelmingly to institutions with something to sell or defend — vendors, employers, platforms, regulators. Anthropic issues guidance Introducing Claude for Teachers; a data-protection authority publishes findings on conversational AI and young people’s mental health IA conversationnelle et santé mentale des jeunes - CNIL; TikTok reports having labeled three billion AI videos — and researchers document what those labels miss TikTok Has Labeled 3 Billion AI Videos: Here Is What the Research Says They Miss. The citizen appears mostly as the object of concern, rarely as an interlocutor. When schools deploy monitoring software on students, the students are not consulted about the literacy being imposed on them Programas de IA para monitorear a estudiantes tienen riesgos de ….

What’s Being Taught

The curriculum, read across these sources, splits unevenly. Detection and protection dominate: verifying sources against AI that “undermines” verification IA et OSINT : comment l’IA sape la vérification des sources, understanding how generative tools become recurring vectors for disinformation Des contenus générés par IA sources de plus en plus récurrentes de …. A thinner strand teaches genuine capability — accessibility as first-class use for people whose brains work differently When your brain works differently, AI isn’t a luxury—it’s accessibility — and a still-thinner one warns of the cost of over-reliance, the “deskilling” that atrophies the judgment of even skilled workers El desafío del “deskilling”. Missing is the middle: the political literacy of knowing when a system is deciding about your life and how to contest it.

What’s Missing

The largest gap is rights. Almost nothing this week teaches citizens what they are owed — algorithmic transparency treated as a legal claim rather than a courtesy Transparencia algorítmica: límites del derecho a saber. A second gap is standing: the discourse tells people how to avoid harm but not how to organize against it, or how the free-expression stakes of AI moderation cut both ways Are LLMs Stifling Political Speech?. A literacy that only teaches vigilance produces careful subjects, not participating citizens — and that difference is the whole argument.

Core Tensions

The concept of “AI literacy” conceals genuine tensions about what citizens need to know and why. Sift through the 5,033 sources this week and the disputes don’t resolve into a knowledge gap waiting to be filled—they resolve into a contested terrain, where the word “literacy” is doing quiet political work. The most fundamental fracture: consumer literacy versus citizen literacy. Are we teaching people to use these systems competently, or to govern the conditions under which the systems are deployed on them? Those are not the same skill, and much of what gets sold as the former quietly forecloses the latter.

Watch the move. When a platform teaches you to spot a labeled AI video, it hands you a consumer skill and keeps the architecture off the table. TikTok has now labeled some three billion AI-generated clips, and yet researchers find the labels systematically miss the content most engineered to deceive—the cheapfakes, the recontextualized real footage, the material that never trips an automated detector TikTok Has Labeled 3 Billion AI Videos: Here Is What the Research Says They Miss. A citizen literate in the fuller sense would ask who decides what gets labeled and why the burden of detection lands on the viewer at all. That is the second tension underneath: individual competency versus collective governance. The Oxford Internet Institute’s work on elections keeps insisting the danger isn’t chiefly the synthetic artifact but the human incentives distributing it—that we should “worry more about humans” than about the models Generative AI and elections: why you should worry more about humans. Individual media-savvy cannot patch a problem that is structural.

The third tension cuts against the grain of nearly every literacy program on offer: use versus refusal. The dominant curriculum assumes competence means fluent operation—prompt well, verify outputs, integrate the tool. But there’s a defensible literacy that includes knowing when not to reach for the system, and knowing what atrophies when you do. The “deskilling” literature is now explicit that offloading judgment to AI can hollow out the very expertise of a firm’s strongest people El desafío del “deskilling”. For citizens, refusal is a competence, not a failure to adapt—and no vendor-authored syllabus will teach it, because the vendor’s interest runs the other way. This is where protection from versus empowerment with gets slippery. A report this year found Google’s AI search features posed what its authors called an “unacceptable risk” to children risks Google’s AI search features pose to kids; the reflex is to demand protective guardrails. Yet with mental-health chatbots, clinicians argue the opposite reflex—that teenagers “need guardrails, not bans” Teens need guardrails, not bans, for mental health chatbots. Protection and empowerment aren’t a dial you set once; they’re a fight over who counts as capable of judgment.

Now the metaphors, because they set the ceiling on what any literacy can imagine. Across this week’s corpus, AI is overwhelmingly a Tool (304 instances) and occasionally a Threat (52). Both framings position the citizen as external to the system—wielding it or fending it off, but never negotiating with it. The Tool metaphor is the more insidious, because a tool implies a neutral instrument answerable entirely to the user’s intent, which is precisely the claim the evidence keeps puncturing: voice-cloning fraud now runs into the billions and breaks identity verification that was itself sold as a tool Clonage Vocal IA : La Crise à 1,8 Md$ qui Brise le KYC. Tools don’t have incentives; platforms do. The Partner framing appears only 7 times, and it would demand something the other two never ask—reciprocity, contestation, a sense that the system acts back. When chatbots dispense voting advice, they are not passive instruments; they shape political speech in ways that regulators are only beginning to name Chatbot Voting Advice Could Help and Harm Voters, and the Oversight Board has already documented models suppressing legitimate political expression Are LLMs Stifling Political Speech?.

Here is the test a citizen can run without expert credentials: when someone offers you “AI literacy,” ask which metaphor it assumes. If it treats the system as a tool you must master, it is training a consumer. A citizen literacy starts by asking who built the tool, whose interests it encodes, and whether the honest answer is sometimes to put it down.

Power & Agency Analysis

Power in AI literacy operates through definition: whoever decides what citizens “need to know” also decides what gets to stay invisible. Our reading of this week’s sources finds a lopsided pattern in how AI agency is portrayed — the dominant frame is AI-as-tool, appearing in 304 instances across the corpus, against just 52 that cast it as threat. That ratio is not neutral. It quietly teaches citizens that these systems are inert instruments waiting for a human hand, right up until the moment something goes wrong and the same systems are described as forces no one could have controlled.

How AI is portrayed

Watch how agency migrates depending on who is speaking. When the benefit is being sold, the human is in charge: a journalist “uses” AI to verify sources faster, a neurodivergent worker “uses” AI to draft and organize When your brain works differently, AI isn’t a luxury—it’s accessibility. When harm arrives, agency drifts toward the machine: content “generated by AI” becomes a recurring source of disinformation Des contenus générés par IA sources de plus en plus récurrentes de …, ChatGPT becomes a “superspreader” The Next Great Misinformation Superspreader. But the Reuters Institute pushes back on exactly this grammar, arguing that in elections you should worry more about the humans deploying AI than the systems themselves Generative AI and elections: why you should worry more about humans. The lesson for a citizen is simple and rarely stated: when a system is described as having “decided” something, ask who built it, who deployed it, and who profits from the passive voice.

Who defines literacy

The people writing the definition of AI literacy are overwhelmingly the people selling the tools or managing the risk. Anthropic introduces “Claude for Teachers” Introducing Claude for Teachers; law firms brief employers on “information disorder” as a workforce liability Information disorder in the age of AI: what it means for employers. Each frames literacy as competent adoption — knowing how to prompt, how to comply, how to stay productive. What is missing is the citizen’s own voice defining what they need to know to defend themselves. That gap is the whole game. When a chatbot dispenses voting advice, the question of whether that advice helps or manipulates is being settled by vendors and regulators, not voters Chatbot Voting Advice Could Help and Harm Voters.

What metaphors teach

The “tool” metaphor’s dominance does specific work: a tool has no interests, so blaming it is a category error, and scrutinizing its owner feels like paranoia. That framing collapses when you look at what these systems actually do to public speech — Meta’s own Oversight Board found reason to ask whether large language models are quietly narrowing political expression Are LLMs Stifling Political Speech?. Tools do not shape what you are allowed to say; infrastructures do. The rarer “threat” metaphor enables the opposite distortion — it licenses panic and hands power to whoever promises protection, as when a lawmaker warns that a rival state has built “digital twins” of every legislator Cammack says China has deployed ‘digital twins’ of every lawmaker. Critical metaphor literacy means noticing that both framings — inert helper, looming menace — conveniently remove the human operators from view. TikTok’s labeling of three billion AI videos sounds like control until research shows how much the labels miss TikTok Has Labeled 3 Billion AI Videos.

Citizen agency

What power do citizens actually hold? Less than the transparency rhetoric implies — legal scholars note that the “right to know” how an algorithm decides runs into hard limits in practice Transparencia algorítmica: límites del derecho a saber. Individual vigilance against a voice-cloning fraud scheme worth $1.8 billion is not a strategy Clonage Vocal IA : La Crise à 1,8 Md$ qui Brise le KYC. The realistic form of agency is collective: public algorithm registries, verification craft that names what AI can and cannot do Lo que la IA puede y no puede hacer por el periodismo de verificación, and a refusal to accept definitions of literacy written entirely by the parties who benefit from your compliance. Knowledge here is not mastery of the tool. It is the ability to see who is holding it.

Drawn from analysis of 5033 sources.

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 5,033 sources documents five recurring patterns in how understanding breaks down—and, more usefully, who profits when it does.

Where Understanding Fails

The most expensive misconception is the simplest: that your senses still work. Voice cloning has turned a familiar heuristic—I recognize this person—into a liability, driving a fraud wave estimated at $1.8 billion that specifically defeats the “know your customer” checks banks once trusted Clonage Vocal IA : La Crise à 1,8 Md$ qui Brise le KYC. The same collapse of trust-by-recognition drives the 2026 wave of vishing scams, where a cloned voice on the phone does what no phishing email ever could Vishing 2026 : un spécialiste décrypte les arnaques vocales IA.

Detection is the second failure, and it does not scale the way platforms imply. TikTok has now labeled three billion AI-generated videos—an impressive number that conceals the gap researchers keep flagging: the labels miss vast swaths of synthetic content, and worse, teach users that unlabeled means authentic TikTok Has Labeled 3 Billion AI Videos: Here Is What the Research Says They Miss. Belgian public broadcasting documents the same labeling shortfall as AI content becomes a recurrent disinformation vector Des contenus générés par IA sources de plus en plus récurrentes de désinformation. Even professional verifiers—OSINT analysts trained to trace sources—report that generative tools are actively eroding their methods IA et OSINT : comment l’IA sape la vérification des sources. If the experts are struggling, “just look closer” is not a literacy strategy.

What Assumptions Mislead

The load-bearing assumption is that a fluent answer is a competent one. Chatbots offering voting advice can genuinely help citizens navigate ballots—and can just as easily launder a confident error into a decision, which is precisely why researchers argue the interface itself demands regulation rather than trust Chatbot Voting Advice Could Help and Harm Voters. Let’s Regulate Accordingly.. A second assumption—that the machine is neutral—dissolves under scrutiny: the Oversight Board’s assessment found AI models systematically shaping what political speech they will and won’t produce, a quiet editorial hand most users never see Are LLMs Stifling Political Speech?. The Reuters Institute adds the correction citizens most need: in elections, the danger is less the synthetic media than the humans who deploy, amplify, and believe it Generative AI and elections: why you should worry more about humans.

Consequences of Gaps

The costs land unevenly. Representative Cammack’s warning that a foreign state has built “digital twins” of every U.S. lawmaker names the ceiling of the threat Cammack says China has deployed ‘digital twins’ of every lawmaker, but the floor is where most citizens live: the employer navigating information disorder without tools Information disorder in the age of AI, and the person leaning on a chatbot for health information it presents with unearned authority—the “generative illusion” that fluency signals accuracy The generative illusion. A scoping review confirms the pattern is structural, not anecdotal: generative systems reshape misinformation’s supply and spread simultaneously Generative AI and misinformation: a scoping review.

What Would Help

The honest answer is that no individual skill closes these gaps, because most are engineered upstream of the user. What helps is a shift in the default question—from is this real? to who benefits if I believe it? Verification journalism offers the durable model: AI extends reach but cannot supply judgment, which stays human Lo que la IA puede y no puede hacer por el periodismo de verificación. Literacy that stops at detection will keep losing; literacy that interrogates incentives at least fails in the right direction.

Evidence Synthesis

Synthesizing this week’s 5,033 sources, the evidence on AI literacy points to an uncomfortable inversion: the skills we told citizens to acquire — spot the fake, check the label, verify the source — are precisely the skills the current generation of tools has engineered past. This goes beyond technical competence. The evidence suggests literacy for citizen participation now means understanding that detection is not something an individual can reliably perform, and organizing accordingly.

What the evidence shows. The convergent finding across this week’s reporting is that labeling and detection, the twin pillars of the “just teach people to spot it” approach, do not hold. TikTok has attached AI labels to roughly three billion videos, yet research finds the system misses the content that matters most — labels catch declared uploads while the manipulative material slips through TikTok Has Labeled 3 Billion AI Videos: Here Is What the Research Says They Miss. OSINT analysts who verify sources for a living now report that generative systems are actively degrading their craft IA et OSINT : comment l’IA sape la vérification des sources, and voice cloning has broken identity-verification systems that banks assumed only humans could pass Clonage Vocal IA : La Crise à 1,8 Md$ qui Brise le KYC. What works, per the fact-checking literature, is narrower than promised: AI accelerates monitoring and pattern-matching but cannot adjudicate truth Lo que la IA puede y no puede hacer por el periodismo de verificación.

Contested terrain. Here the evidence splits. The Reuters Institute argues the election panic is misdirected — that human behavior and existing political incentives, not synthetic media, drive most disinformation, and that literacy campaigns fixated on deepfakes miss the real machinery Generative AI and elections: why you should worry more about humans than AI systems. Against this, a scoping review of generative AI and misinformation documents a genuine scale shift in production capacity Generative AI and misinformation: a scoping review of the role of large language models, and NewsGuard’s early tests showed chatbots generating fluent falsehood on demand The Next Great Misinformation Superspreader. Both can be true; what “literacy” should prioritize under that condition is unresolved.

Across domains. Tool-specific literacy now includes knowing that the chatbot itself is an actor in civic life. When systems dispense voting advice, they help and mislead in the same breath, which is why researchers argue for regulation rather than user vigilance Chatbot Voting Advice Could Help and Harm Voters. Let’s Regulate Accordingly.. The Oversight Board’s finding that large language models quietly suppress legitimate political expression Are LLMs Stifling Political Speech? adds a second-order problem: the tool shapes the discourse before the citizen ever evaluates its output. The social dimension surfaces where the harm concentrates — Representative Cammack’s warning that China has built “digital twins” of lawmakers Cammack says China has deployed ‘digital twins’ of every lawmaker points to targeting no individual literacy can absorb.

Gaps and uncertainty. We do not know whether media-literacy interventions actually reduce susceptibility at population scale, nor how transparency mandates translate into anything citizens can use — the legal “right to know” about algorithms remains largely aspirational Transparencia algorítmica: límites del derecho a saber.

For citizens. The evidence favors institutional defense over heroic individual vigilance. Personally: slow down, treat urgency and unverifiable audio as red flags Vishing 2026 : un spécialiste décrypte les arnaques vocales IA, and stop assuming a label means safety. Collectively: the durable fixes — mandatory provenance, chatbot regulation, verification infrastructure — are policy, not personality. Literacy’s honest job is teaching citizens which burden is theirs and which they should refuse to carry alone.