AI Literacy for Citizen Participation Report

Analysis of 1,025 AI literacy sources this week—drawn from a corpus of 5,001—reveals a discourse still organized around the user who operates AI, while the citizen who is operated on by it remains an afterthought. The citizen-as-participant framing—literacy as the capacity to know when a system is being run on you, to contest it, to refuse it—surfaces in a distinct minority of sources. Most treat literacy as a skill to be acquired: prompt fluency, tool adoption, output-checking. The publication has, across earlier pieces, weighed literacy as workforce preparation against ethical engagement. The delta this week is sharper and less flattering: even the “ethical” half of that pairing assumes a person sitting at a keyboard, choosing. The harder cases are people who never opted in.

The Landscape

What counts as “AI literacy” is being defined, overwhelmingly, by the entities that sell or deploy the tools. Microsoft’s training path frames literacy as a journey toward confident use Introduction to AI Literacy - Training | Microsoft Learn; a parallel module folds accessibility into the same competence-acquisition arc Inteligencia artificial generativa y accesibilidad | Microsoft Learn. This is literacy as onboarding. Against it sits a thinner but more interesting strand—prompt engineering reframed not as a vendor skill but as a civic competence of the century Prompt engineering as a new 21st century skill - Frontiers—and a sober counter-current insisting that fluency with a tool is not understanding of it: large language models still miss what accessibility standards actually require AI Is Not Your Accessibility Expert: What LLMs Still Miss About WCAG. Whoever defines the curriculum defines the citizen’s relationship to the machine, and right now the curriculum-writers mostly have something to sell.

Whose Literacy

The perspective distribution is lopsided. Platform documentation, foundation reports, and field surveys dominate; the voice that is conspicuously thin is the person being acted upon without consent. New America’s field interviews at least centre practitioners rather than vendors Digital Literacy in the Age of AI: Voices from the Field, and Estonia’s national approach is notable for treating literacy as a public good rather than a product feature—what its proponents call technorealism rather than technoptimism Estonia adopta un enfoque tecnorrealista para la alfabetización en IA …. But the structural fact stands: experts are teaching publics how to use, far more than publics are demanding to know how they are being used.

What’s Being Taught

The thematic clusters split cleanly into use and protection, and use is winning the resource war. On the protection side, the evidence is alarming and underweighted in curricula: synthetic media is eroding the shared factual baseline democracy runs on Synthetic Media, Political Disinformation, and the Erosion …; chatbots deploy documented dark patterns designed to retain and manipulate Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design; AI-cloned voices already power extortion calls against ordinary families AI scam calls: This mom believes fake kidnappers cloned her … - CNN; and identity-theft operations now use generated personas to drain public funds How Scammers Are Using AI to Steal College Financial Aid. A literacy that teaches prompting but not recognition of these threats arms the wrong half of the encounter.

What’s Missing

The largest gap is governance literacy—the citizen’s capacity to participate in deciding how AI is allowed to operate, not merely to cope once it does. The accountability vacuum is documented Ungoverned: AI, Accountability, and the Limits of Law, yet almost no source treats knowing your data rights or contesting an automated decision as a teachable competence. Equally absent: the populations most exposed and least equipped—people surveilled by error-prone systems that produce real punishments Falsas alarmas de vigilancia con IA han provocan castigos y arrestos …. The discourse teaches us to drive the car. It has not yet decided we are allowed to read the road.

Core Tensions

The concept of “AI literacy” conceals genuine tensions about what citizens need to know and why. Across the 5,001 sources surveyed this week, the word does a remarkable amount of quiet work—standing in for a technical skill, a civic defense, a consumer competency, and a governance demand all at once, as if these were the same project. They are not. The most fundamental fracture: literacy framed as competence with the tool versus literacy framed as judgment about the system. This isn’t a knowledge gap to be filled with another training module. It’s contested terrain, and the contest is mostly being settled by whoever supplies the curriculum.

Watch the first move. When Microsoft offers an Introduction to AI Literacy and the research literature elevates prompt engineering as a new 21st century skill, literacy quietly becomes fluency in operating the product. A systematic review of prompt engineering in higher education treats the chatbot’s quirks as a body of knowledge worth mastering. That’s a real skill—but notice who benefits when “becoming literate” means “becoming a better user of my system.” A population skilled at prompting is not necessarily a population able to ask whether the system should be deployed at all. Estonia’s recent move toward what Euronews calls a “technorealist” approach—pairing tool fluency with explicit instruction in limits and harms—reads as a deliberate correction to that drift.

Consumer literacy versus citizen literacy. The second tension is whom the literate person is imagined to be. The Center for Democracy and Technology’s taxonomy of dark patterns in AI chatbots documents systems engineered to retain, flatter, and manipulate—which means consumer-grade literacy (spot the manipulation, protect yourself) is necessary but ends at the individual’s screen. Citizen literacy asks a different question: who is permitted to build these patterns in the first place? The Columbia Law Review’s Ungoverned argues that accountability has outrun existing law. No amount of personal savvy closes that gap. A citizen who can detect a dark pattern but has no lever over its legality has been handed responsibility without power—a familiar bargain.

Protection from versus empowerment with. The third fracture runs through nearly every harm story this week. AI scam calls cloning a child’s voice, synthetic media eroding the shared factual record, false AI surveillance alarms triggering arrests, and Common Sense Media’s finding that major chatbots are unsafe for teen mental health support—these all argue for a defensive literacy, a citizenry inoculated against deception. But a purely protective frame casts citizens as perpetual potential victims, never as agents who might use these systems on their own terms. Stanford’s work on AI companions and young people holds both at once: the danger is real, and pretending people will simply abstain is fantasy. Literacy that only warns produces anxiety; literacy that only empowers produces marks.

This is where the metaphors do their damage. The dominant framing this week is overwhelmingly Tool (304 instances), trailed distantly by Threat (52). The Tool metaphor flatters the user—you wield it, you’re in control—and conveniently obscures that the “tool” is optimizing against you, a point New America’s field analysis makes when it insists digital literacy now means understanding incentives, not just interfaces. The Threat metaphor over-corrects into helplessness. Almost absent—Partner appears just 7 times—is the framing that would actually demand something: that these systems are counterparties with their own objectives, which would require literacy to include knowing whose interests the partner serves. France’s education ministry report on AI in institutions gestures toward this relational view.

The test citizens can apply themselves is blunt: when someone offers to make you “AI literate,” ask which tension they’ve resolved on your behalf. If the answer is always use the tool better, you are being trained, not educated—and the HEPI student survey showing rapid adoption with shallow critical grounding suggests how easily the two get confused.

Power & Agency Analysis

Power in AI literacy operates through definition: whoever decides what citizens “need to know” also decides what gets to stay invisible. This week’s evidence base—1,025 articles inside a 5,001-article corpus—shows the pattern clearly. The dominant framing of AI literacy is overwhelmingly the tool frame, recurring in the literature far more than the language of threat, and far, far more than any account that names AI as an autonomous decider. That distribution is not neutral. It is itself a teaching, and it teaches citizens to locate agency in the wrong place.

How AI is portrayed

Watch the grammar. When a system enrolls thousands of fake students to drain financial aid, the AP describes scammers “using AI” to commit identity theft Scams to steal college financial aid are using AI for identity theft …—human agent, tool in hand. But when a surveillance product flags a child and triggers an arrest, the story slides toward “false alarms” generated by the system itself Falsas alarmas de vigilancia con IA han provocan castigos y arrestos …, as if no procurement officer chose it and no administrator acted on its output. The agency assignment shifts depending on who would be embarrassed by the attribution. When citizens benefit, humans are credited; when citizens are harmed, the algorithm absorbs the blame. A literate reader learns to ask, every time: who is the actual actor this sentence is hiding?

The same evasion runs through chatbot harms. Common Sense Media found major chatbots unsafe for teen mental-health support Common Sense Media Finds Major AI Chatbots Unsafe for Teen Mental …, and Stanford researchers documented the specific dangers of AI companions Why AI companions and young people can make for a dangerous mix. The Center for Democracy and Technology’s taxonomy of chatbot dark patterns Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design makes the buried actor visible: the manipulative behavior is designed, not emergent. Naming the designer is the literacy move.

Who defines literacy

Notice who is writing the curriculum. Microsoft offers the “Introduction to AI Literacy” path Introduction to AI Literacy - Training | Microsoft Learn; OpenAI instructs educators on how to read its own outputs How can educators respond to students presenting AI-generated content …. When the vendor authors the definition of competence, “literacy” tends to mean fluency with the product, not the capacity to refuse it. The independent counter-tradition exists—New America’s field survey Digital Literacy in the Age of AI: Voices from the Field, Renaissance Numérique’s framework, and Estonia’s deliberately “techno-realist” national approach Estonia adopta un enfoque tecnorrealista para la alfabetización en IA …—but it is outspent and out-distributed. Prior weeks here treated equity of access to literacy as the central fault line; the sharper problem this week is equity of authorship. Access to a curriculum written by the seller is not a remedy.

What metaphors teach

The “tool” metaphor, dominant by a wide margin, carries a quiet promise: that the thing sits inert until a human picks it up, fully under the user’s control. That promise is false in exactly the cases that matter. SitePoint’s finding that large language models still miss core accessibility requirements AI Is Not Your Accessibility Expert: What LLMs Still Miss About WCAG shows a “tool” confidently producing wrong output while the user assumes competence. The “threat” metaphor, less common, does different work: it licenses sweeping responses—surveillance, lockdown, prosecution—as in the Adelphi plagiarism dispute Adelphi University accused a student of using AI to … - Newsday, where a detection tool’s verdict became an accusation. Critical metaphor literacy means seeing that “tool” disarms scrutiny and “threat” justifies control—and that both can serve the same institutional interest.

Citizen agency

What power do citizens actually hold? Less than the optimists claim, more than the fatalists allow. The legal scholarship is blunt about the ceiling: accountability frequently exceeds the reach of existing law Ungoverned: AI, Accountability, and the Limits of Law. Individual vigilance will not stop a cloned voice demanding ransom AI scam calls: This mom believes fake kidnappers cloned her … - CNN or synthetic disinformation corroding shared facts Synthetic Media, Political Disinformation, and the Erosion …. The protective knowledge that scales is collective: demanding that agency be named, that authorship be disclosed, that the actor behind every “the AI decided” be dragged back into the sentence. Literacy here is not a personal skill. It is a civic posture—the refusal to let power hide inside the passive voice.

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 model performs exactly as built; the breakdown happens in the human’s head, in the gap between what a system appears to offer and what it actually does. Our analysis documents five recurring patterns in how that understanding collapses — and none of them is fixed by knowing how to write a better prompt.

Where Understanding Fails

The first and most consequential failure is misplaced trust calibration — citizens trusting outputs that deserve suspicion and rejecting tools that might actually help. The trust failure dominates. When a chatbot answers in fluent, confident prose, fluency reads as accuracy, and most people have no internal alarm for the difference. Stanford’s benchmarking found that purpose-built legal models hallucinate in at least one of every six queries AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) — and these are the systems marketed as reliable. A citizen with no domain expertise has no way to spot the one-in-six.

Detection is the second gap. The deepfake-nudes crisis spreading through schools worldwide shows how badly ordinary people read synthetic media The Deepfake Nudes Crisis in Schools Is Much Worse Than You Thought, and synthetic political content is now sophisticated enough to erode the baseline assumption that a video records something that happened Synthetic Media, Political Disinformation, and the Erosion. The third failure is protection — citizens handing over voice, face, and identity data without grasping that those samples are exactly what an AI cloning scam needs, as the mother who received a faked kidnapping call from her “daughter” learned AI scam calls: This mom believes fake kidnappers cloned her.

What Assumptions Mislead

Underneath these failures sit two assumptions that quietly betray people. The first is that the interface is neutral — that a chatbot is a tool waiting to serve, rather than a designed environment engineered to retain you. The Center for Democracy and Technology’s taxonomy of “dark patterns” in chatbots catalogs how systems manufacture emotional dependency and discourage you from leaving Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design. The second is that competence in one domain transfers to all — the belief that a model fluent in conversation is therefore fluent in fact, accessibility, or care. It isn’t: LLMs confidently produce code that violates the very accessibility standards they claim to satisfy AI Is Not Your Accessibility Expert: What LLMs Still Miss About WCAG.

Consequences of Gaps

The costs land unevenly. Common Sense Media judged major chatbots unsafe for teen mental-health support Common Sense Media Finds Major AI Chatbots Unsafe for Teen Mental, and Stanford researchers documented the specific danger of companion bots to young people who read simulated empathy as real Why AI companions and young people can make for a dangerous mix. At scale, scammers using generated identities to drain financial-aid programs siphon public money that funds real students Scams to steal college financial aid are using AI for identity theft. The people least equipped to detect these moves — the time-poor, the isolated, the already-precarious — absorb the largest share of harm, while the systems generating it operate in a space where law is still catching up Ungoverned: AI, Accountability, and the Limits of Law.

What Would Help

The literacy that prevents these failures is not operational — it is dispositional. It is the habit of asking who built this and what do they want, the reflex to treat fluent output as a claim rather than a fact, the instinct to withhold identity data by default. New America’s field research argues that digital literacy must now center critical evaluation over tool proficiency Digital Literacy in the Age of AI: Voices from the Field. But honesty requires the limit: no amount of individual skepticism neutralizes a one-in-six hallucination rate or a dark pattern engineered by people with more resources than any reader. Literacy reduces exposure. It does not absolve the builders.

Evidence Synthesis

Synthesizing the week’s 1,025 AI-literacy analyses drawn from a corpus of 5,001 sources, the evidence points to a single uncomfortable finding: the literacy that actually protects citizens is not the literacy that gets funded. This goes beyond technical skill. The skills being packaged and sold — prompt-writing, tool fluency, “getting the most out of” a chatbot — are precisely the skills that make you a more efficient user of someone else’s product, while the capacities that defend you against that product go untaught.

What the evidence shows

Start with what converges. Vendor-built curricula frame literacy as competence-with-the-tool: Microsoft’s own pathway opens with using AI well Introduction to AI Literacy - Training | Microsoft Learn, and a growing research literature has elevated prompt engineering to the status of a civic competence — “a new 21st century skill” Prompt engineering as a new 21st century skill - Frontiers, with systematic reviews now codifying it Prompt engineering in higher education: a systematic review to help …. Against this, the field-level work points elsewhere: New America’s practitioner interviews find that the durable skill is critical judgment of outputs, not facility with inputs Digital Literacy in the Age of AI: Voices from the Field, and the European review of anti-disinformation programs from 2018 to 2024 finds that the interventions that work teach people to interrogate synthetic media, not to produce it Éduquer contre la désinformation amplifiée par l’IA et l’hypertrucage …. Estonia’s much-discussed national rollout is notable precisely because it is “technorealist” — naming the tools’ limits rather than their magic Estonia adopta un enfoque tecnorrealista para la alfabetización en IA ….

Contested terrain

Here the word “literacy” splits. One camp treats it as adoption: the HEPI survey documents near-universal generative-AI use among the young Student Generative AI Survey 2025 - HEPI, and reads fluency as readiness. The other camp reads the same fluency as exposure. The Center for Democracy & Technology’s taxonomy of chatbot dark patterns shows that the interfaces citizens are told to master are engineered to manipulate them Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design. You cannot resolve “use it more” and “it is designed to exploit you” with a curriculum that only does the first.

Across domains

Tool-specific literacy means knowing where the tool is confidently wrong. Stanford’s benchmarking found legal models hallucinating in one of six queries or worse AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More …, and accessibility specialists document that LLMs miss what WCAG compliance actually requires AI Is Not Your Accessibility Expert: What LLMs Still Miss About WCAG. The social dimension is starker: the same systems are running financial-aid identity theft at scale Scams to steal college financial aid are using AI for identity theft …, cloning voices for ransom calls AI scam calls: This mom believes fake kidnappers cloned her … - CNN, and corroding the shared evidentiary ground democracy runs on Synthetic Media, Political Disinformation, and the Erosion …. Literacy as equity means: who can afford to be skeptical, and who gets defrauded.

Gaps and uncertainty

What we lack is causal evidence that any literacy program reduces actual harm — fewer scams fallen for, fewer fakes believed. The studies measure confidence and usage, not protection. And the accountability vacuum is documented but unsolved: the law has not caught up Ungoverned: AI, Accountability, and the Limits of Law.

For citizens

Individually: treat fluency and defense as different skills, and prioritize the second — assume confident outputs can be wrong, verify before you trust, and recognize that an interface optimized to please you is optimized to move you. Collectively: literacy cannot be the whole burden. When systems are engineered to deceive and the law lags, asking citizens to simply learn harder is the move to watch — and to refuse.