AI Literacy for Citizen Participation Report

Analysis of 1,017 AI literacy sources within this week’s corpus of 4,785 reveals a discourse focused on detection and defense while neglecting the citizen as an agent with standing to shape how AI is used at all. The citizen-as-participant framing — the reader as someone who governs, consents, and contests, not merely someone who spots fakes — appears in roughly one in seven sources. Most treat literacy as a threat-response reflex: learn to recognize the deepfake, the cloned voice, the fabricated video, before it gets you.

That reflex is not wrong. It is just narrow. And the narrowing is worth watching, because it quietly redefines what a competent citizen is supposed to be.

The Landscape

“AI literacy” this week means, overwhelmingly, the ability to survive AI-generated deception. The dominant vocabulary is forensic: tools that trace the sources of fake video, provenance standards like Content Credentials, methods to keep kids safe from illegal AI-generated content. The framing is defensive because the threats are real — AI-generated content is an increasingly recurrent source of disinformation, and UNESCO frames deepfakes as a crisis of knowing itself. But notice the move: literacy becomes a personal firewall. The question shifts from “how should AI be used in public life” to “how do I avoid being fooled.” The first is political. The second is hygiene.

Whose Literacy

The teaching runs almost entirely one direction — from institutions that have already decided what you need to know down to a public presumed vulnerable. Fact-checking outfits explain how deepfakes are used to scam you; security firms catalogue voice-cloning fraud against families; advocacy groups warn that 96% of deepfakes are non-consensual pornography. What’s scarce is the citizen speaking back. A genuine exception is Microsoft’s account of a data approach that puts communities in charge — but even there the vendor narrates the empowerment. The people whose literacy is at stake rarely set the terms of what “literate” means.

What’s Being Taught

Two clusters dominate. The first is technical recognition — spot the artifact, check the provenance. The second is institutional trust — outsource verification to a tool or a label. Both crowd out the harder competency: understanding when AI is being run on you and by whom. The week’s most instructive cases are political. AI chatbots gave inaccurate and unreliable voting advice in Hungary; the UN flagged new risks as AI enters Latin American elections; the Journal documented a $50 million campaign to shift U.S. public opinion. None of these are solved by knowing what a deepfake looks like. They require what a few sources actually name — critical and technopolitical literacy, reading the interests behind the output, not just its pixels.

What’s Missing

The gap is agency. Detection literacy tells you what to fear; almost nothing tells you what you can do — how to exercise data rights, withhold consent, or demand explicability as a public resource. The discourse also underserves the people least equipped to self-defend: elderly targets of voice fraud, children exploited at scale, non-English speakers reading past machine-translated warnings. And the sharpest absence: a public that can lose in this system while doing everything right. When even a cabinet minister must file a police complaint over a deepfake of himself, individual vigilance has plainly hit its ceiling. Literacy that stops at self-defense leaves the governing to someone else.

Core Tensions

The phrase “AI literacy” behaves like a settled goal, as if we all knew what a citizen needs to know and were merely arguing about delivery. We don’t. Beneath the consensus sits a contested terrain, and the most fundamental fracture runs between two incompatible pictures of the literate citizen: one who can operate the technology, and one who can interrogate the conditions under which it operates. Call it technical skill versus critical understanding. The distinction sounds academic until you notice that almost every institution offering “literacy” quietly picks the first and calls it the whole.

Watch the move. When a course teaches you to prompt a chatbot efficiently, it has taught you to be a better user—which is exactly what the vendor wants, and exactly what leaves you defenseless when the same chatbot dispenses confidently wrong election guidance. That is not hypothetical: when researchers tested AI chatbots on voting logistics ahead of Hungary’s elections, the advice came back “inaccurate and unreliable,” misstating registration rules and polling procedures Election voting advice from AI chatbots ‘inaccurate and unreliable’. Operational fluency does nothing here. What protects you is the critical instinct to distrust a fluent answer about civic mechanics—a capacity no prompt-engineering tutorial cultivates.

The second tension is sharper because it is about power, not skill: individual competency versus collective governance. The dominant literacy story is relentlessly individual—learn to spot the deepfake, verify the source, protect yourself. But some threats are structural and immune to personal vigilance. Consider the $50 million influence operation the Wall Street Journal documented, an organized, well-funded effort to reshape public opinion at scale Israel’s $50 Million Experiment to Change U.S. Public Opinion. No amount of individual media literacy neutralizes a coordinated, capitalized campaign; that requires disclosure regimes, platform accountability, provenance standards. Framing literacy as personal responsibility conveniently offloads a governance failure onto the citizen. It tells you to read the water more carefully while declining to name who poisoned the well. Europe’s debate over labelling AI-generated content Des contenus générés par IA sources de plus en plus récurrentes de désinformation and the push for cryptographic provenance in Content Credentials PDF Content Credentials: Strengthening Multimedia Integrity are collective moves—and their existence is a tacit admission that individual literacy was never going to be enough.

The third fracture is the one the marketing least wants you to see: protection FROM versus empowerment WITH. Whole domains of AI literacy are, honestly, defensive crouches. The material on voice-cloning fraud that fakes a relative’s voice AI Voice Cloning Scam: How Fraudsters Fake Your Family’s Voice, the documentation that 96% of deepfakes are non-consensual pornography Deepfake Crisis 2025: 96% Is Non-Consensual Porn, the reporting on AI reshaping child sexual exploitation at scale Abuse at scale: Artificial intelligence is reshaping child sexual exploitation—none of this is about empowerment. It is about survival in a hostile information environment. And when Piyush Goyal filed a police complaint over a deepfake of himself Piyush Goyal files police complaint over AI-generated deepfake video, he demonstrated that even the powerful reach for law, not literacy, when targeted. If the powerful can’t self-educate their way out, the “just learn to spot it” prescription for ordinary citizens is close to a bad-faith deflection.

These tensions are held in place by metaphor. Across the discourse, AI is overwhelmingly framed as a tool (304 instances) and only occasionally as a threat (52). The tool metaphor is not innocent: a tool is neutral, obedient, and its misuse is your fault. It naturalizes the individual-competency framing and makes the vendor invisible. The partner framing—vanishingly rare at 7 instances—would demand something the tool story never asks: reciprocal obligations, transparency about what the system knows and wants, the kind of explainability one French analysis calls the critical resource of a functioning social contract Effondrement ou reconquête du contrat social algorithmique.

Here is the literacy that matters, then: not learning to use the tool, but learning to notice which metaphor you’ve been handed—and asking who benefits from your accepting it. UNESCO’s framing of deepfakes as a “crisis of knowing” Deepfakes and the crisis of knowing gets the stakes right. The question was never whether you can operate the machine. It is whether you can still tell who is operating you.

Power & Agency Analysis

Power in AI literacy operates through definition: who decides what citizens “need to know” shapes what remains invisible. Our analysis of this week’s 4,785 sources finds a persistent grammatical sleight-of-hand in how AI agency is portrayed—the “tool” framing dominates by an order of magnitude over the “threat” framing, and both quietly displace the human hand on the switch. This framing matters for citizens, because the words we use to describe who acted determine whom we think to hold responsible.

How AI is portrayed

Watch the verbs. When a deepfake circulates, the coverage says the AI “generated” it, the algorithm “decided,” the system “produced” a fabricated video of a public official—as when India’s commerce minister filed a police complaint over a manufactured clip Piyush Goyal files police complaint over AI-generated deepfake video, warns of strict legal action. The machine becomes the actor; the person who commissioned, funded, and released the fabrication recedes into grammar. Contrast the WSJ account of a $50 million influence operation aimed at American opinion Israel’s $50 Million Experiment to Change U.S. Public Opinion—here the agents are named, funded, and strategic. The lesson for citizens is that autonomous-AI language is not neutral description; it is a liability shield. When chatbots dispense “inaccurate and unreliable” voting advice Election voting advice from AI chatbots ‘inaccurate and unreliable’, the error was engineered by a vendor who chose to ship a system that speaks with unearned confidence. Assigning agency to “the AI” lets that vendor off the hook. The first act of literacy is refusing the passive voice.

Who defines literacy

The perspective distribution in this week’s evidence is lopsided. Definitions of what citizens should understand flow overwhelmingly from institutions with something to sell or govern: platform vendors, national security agencies, and standards bodies. The NSA-adjacent push for cryptographic provenance—Content Credentials PDF Content Credentials: Strengthening Multimedia Integrity in the …—defines the problem as one of technical verification, which conveniently makes the solution proprietary and the citizen a passive checker of badges. Microsoft frames data governance as something to be handed back to communities A new approach to AI data puts communities in charge—a generous gesture that still leaves the platform as the party doing the handing. Notably absent from nearly all of it: the citizens themselves, voting on what they need. UNESCO’s framing of a “crisis of knowing” Deepfakes and the crisis of knowing - UNESCO at least names the epistemic stakes, but even there the citizen is a subject to be protected, not a party consulted.

What metaphors teach

The “tool” metaphor is the workhorse, and it teaches a comforting falsehood: that AI is inert until picked up, morally neutral, an extension of the user’s will. That framing obscures the fact that these systems arrive with defaults, incentives, and biases already baked in—that a voice-cloning system is a “tool” the way a lockpick is a tool AI Voice Cloning Scam: How Fraudsters Fake Your Family’s Voice. The “threat” metaphor, rarer, does different work: it enables emergency powers, calls for surveillance, and content-scanning regimes justified by the genuinely horrifying—the industrialization of child exploitation Abuse at scale: Artificial intelligence is reshaping child sexual exploitation. Both metaphors are true and both are used. Critical metaphor literacy means noticing which one is deployed and what it authorizes—the “tool” story disclaims responsibility, the “threat” story expands control. What French analysts call the “algorithmic social contract” Effondrement ou reconquête du contrat social algorithmique hinges precisely on explicability—the right to be told, in plain terms, which story is being told to you.

Citizen agency

So what power do citizens actually hold? Less than the “empowerment” rhetoric claims, more than the “crisis” rhetoric admits. New forensic tools that trace fake video to its source New tool identifies the sources of fake video | UCR News shift the burden back toward the human actor—useful, but only in expert hands. The durable form of citizen power is not individual detection skill, which will always lose the arms race, but collective demand: for named accountability, for explicability as a right, and for a seat at the table where “literacy” gets defined. Knowledge here is protection precisely because it converts “the AI did it” back into “someone did it, and here is who.”

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. The tool can work exactly as designed and still leave you worse off, because the breakdown happens in your head, not in the machine. Our analysis of this week’s sources documents five recurring ways understanding collapses — and they cluster, tellingly, not around ignorance but around misplaced confidence.

Where understanding fails

The most expensive failure is not under-trust but over-trust in the wrong register. When Hungarian voters asked chatbots for practical guidance on how to cast a ballot, the answers were, in the researchers’ words, “inaccurate and unreliable” Election voting advice from AI chatbots ‘inaccurate and unreliable’. The failure here is not that the model lied — it does not know it is lying — but that citizens treated a fluent text generator as a civic reference desk. Detection runs the opposite direction: people cannot see what they are looking at. AI-generated material has become a recurring source of disinformation precisely because the surface signals we learned to trust — a confident face, a clean voice, a plausible dateline — no longer correlate with truth Des contenus générés par IA sources de plus en plus récurrentes de désinformation. When a sitting minister has to file a police complaint to establish that a video of him was fabricated Piyush Goyal files police complaint over AI-generated deepfake video, the ordinary viewer has no comparable recourse.

What assumptions mislead

Three assumptions do most of the damage. The first is that seeing is verifying — that a realistic image carries its own proof. The second is that a machine, lacking motives, is therefore neutral; the reality is that models reproduce and launder the biases and errors baked into their training and prompting Pourquoi l’IA fait des erreurs et comment les reconnaître. The third, and most corrosive, is that the familiar voice on the phone is the person it sounds like. Voice-cloning fraud weaponizes exactly this reflex, faking a relative in distress to extract money before the target thinks to doubt AI Voice Cloning Scam: How Fraudsters Fake Your Family’s Voice. Each assumption is a shortcut that served us well for a century and now betrays us.

Consequences of gaps

The costs are unevenly distributed, and that is the point. Coordinated influence operations exploit the detection gap at scale — a reported $50 million effort to shift U.S. opinion shows what industrialized persuasion looks like when audiences cannot trace provenance Israel’s $50 Million Experiment to Change U.S. Public Opinion. The most vulnerable bear the worst of it: the elderly targeted by cloned-voice scams Cómo se usa la inteligencia artificial en ‘deepfakes’ para estafar, and children, where synthetic exploitation material is being produced at industrial volume Abuse at scale: Artificial intelligence is reshaping child sexual exploitation. At the collective level, UNESCO names the deepest harm precisely: a “crisis of knowing,” where the erosion of shared evidentiary ground makes democratic deliberation itself harder Deepfakes and the crisis of knowing.

What would help — honestly

Literacy that would actually prevent these failures is not a checklist of “spot the fake” tells; those decay as generation improves. It is a shift in default posture — from authenticating content by appearance to demanding provenance, the verifiable chain of where a file came from, as the Content Credentials standard attempts to encode PDF Content Credentials: Strengthening Multimedia Integrity. But be honest about the ceiling: no individual skill offsets an asymmetry this large. New forensic tools that trace a fake video to its source New tool identifies the sources of fake video belong to investigators, not to you at 11 p.m. with a suspicious message. Personal literacy narrows the gap; it does not close it. The remainder is a job for infrastructure and law — and pretending otherwise is its own kind of failure.

Evidence Synthesis

Synthesizing this week’s 4,785 sources, the evidence on AI literacy points to a hard finding: the skills most often taught — spot the glitch, check the source, pause before sharing — are being outrun by the systems they were meant to counter. This goes beyond technical skill. What citizens actually need is the capacity to operate when individual verification no longer scales, and to demand the infrastructure that makes verification possible at all.

What the evidence shows

The convergent finding across this week’s reporting is that detection has migrated from the human eye to the toolchain — and that migration is itself the literacy story. Researchers at UC Riverside unveiled a system that identifies the source model behind a fake video rather than asking viewers to judge authenticity themselves New tool identifies the sources of fake video | UCR News. The U.S. defense establishment has been pushing Content Credentials — cryptographic provenance baked into files at capture — precisely because downstream judgment fails PDF Content Credentials: Strengthening Multimedia Integrity in the …. UNESCO frames the underlying condition bluntly as a “crisis of knowing,” where the reflex to authenticate everything itself becomes a vector for exhaustion and disengagement Deepfakes and the crisis of knowing - UNESCO. What works, on this evidence, is less “train the individual to be suspicious” and more “build systems where provenance travels with content” — with critical, technopolitical media education as the human layer on top Hacia una Alfabetización Crítica y Tecnopolítica en el Aula ….

Contested terrain

Where the evidence conflicts is on whether “literacy” is even the right unit of intervention. One camp treats it as a competence citizens acquire; another treats over-reliance on individual vigilance as the problem, since the labelling of AI content remains inconsistent and gameable Des contenus générés par IA sources de plus en plus récurrentes de …. Even OSINT verification methods that professional investigators rely on are being degraded by generative fakes IA y OSINT: cómo la IA mina la verificación de fuentes. The delta from prior framings: the question is no longer skills-versus-ethics, but whether any distributed individual skill can hold against industrial-scale generation Desinformación e inteligencia artificial: del clickbait a los ….

Across domains

Tool-specific literacy now means knowing what a tool does to you. Chatbots gave “inaccurate and unreliable” voting advice in a live European election test Election voting advice from AI chatbots ‘inaccurate and unreliable’ — a fluent-answer machine is not a civic-information machine, and the difference is invisible without knowing why models err Pourquoi l’IA fait des erreurs et comment les reconnaître. The social-aspects dimension is equity: influence operations are now capitalized ventures, as the reported $50 million campaign to shift U.S. opinion shows Israel’s $50 Million Experiment to Change U.S. Public Opinion, and voice-cloning scams prey hardest on those with least exposure to the technique AI Voice Cloning Scam: How Fraudsters Fake Your Family’s Voice. One promising thread hands data governance back to communities rather than platforms A new approach to AI data puts communities in charge.

Gaps and uncertainty

We do not know whether provenance standards will be adopted widely enough to matter, or whether they will fragment by jurisdiction. We lack evidence that any media-literacy curriculum measurably reduces susceptibility at population scale. And the sharpest harms — non-consensual deepfakes and child exploitation — resist literacy framing entirely; you cannot educate your way out of being a target Deepfake Crisis 2025: 96% Is Non-Consensual Porn.

For citizens

Two evidence-based moves. Individually: shift from detecting fakes to demanding provenance — treat unlabelled, unsourced content as unverified by default, and use the fallback that scams exploit trust, not gullibility Cómo se usa la inteligencia artificial (IA) en ‘deepfakes’ para estafar …. Collectively: the wins are structural — provenance mandates, community data control, accountable labelling. Fact-checking initiatives help Can This New Way to Fight Misinformation Work?, but the citizen’s real leverage is political: insisting the burden of proof sit with those generating content, not those receiving it Inteligencia artificial y desinformación - UNESCO.