AI Fluency, how to teach people what AI actually is

"AI fluency" is used loosely. In the research it splits into two questions that are easy to run together: what should a learner end up able to do, and what wrong ideas have to be dismantled on the way there. The first has good answers. The second is where most of the unfinished work is.

The competency side is well developed. Long & Magerko (2020) remains the reference synthesis of AI literacy competencies, and Bewersdorff et al. (2023) reviews the myths and preconceptions found in schooling. What those frameworks do not say is which wrong intuition to take apart first. That ordering question is the part a curriculum has to decide for itself.

This note works from AI, plainly, a beginner AI-fluency site built outward from a screened literature base of 319 papers: fifteen lessons for adults starting from zero, each aimed at one documented misconception. The site is an evidence-informed design, not a validated intervention, and this note keeps that distinction in view throughout. Where a claim rests on a user study it says so; where it rests on conceptual analysis or on the author's judgement it says that instead.

Two dials, opposite directions

The organising finding is a survey of 319 knowledge workers (Lee et al. 2025): higher confidence in the AI predicted less critical thinking, while higher confidence in oneself predicted more.

Those are separable targets, and they pull opposite ways. Beginner AI content tends to raise the first dial (it is impressive, it is helpful, you can trust it) and ignore the second. The result reads as reassurance and produces overtrust. The teaching goal that follows is to raise self-efficacy while pulling trust in the tool down to something calibrated: a learner who will try things without asking permission and check things without being told.

Ten misconceptions, five distinctions

Ten misconceptions organise most of what goes wrong when a newcomer meets a chatbot. They are not ten independent errors. Each one collapses a distinction the system's behaviour depends on, and the same five distinctions recur underneath all of them.

Four of the five come from Lin (2026), written for a scientific readership: the pretraining process versus the deployed system, the learned distribution versus any one sample, parametric versus contextual versus external memory, and task competence versus agency. The curriculum adapts them for general adults and adds a fifth the expert version does not need, interface cues versus mechanism, because for a beginner the chat window is the main evidence about what the thing is. Fluency, first-person pronouns and politeness are design choices, not facts about the internals.

This is why a short curriculum is possible at all. A learner who holds a distinction defuses misconceptions the lessons never mention. Pick a distinction below to see which misconceptions sit on it, and the concrete move each lesson ends on.

"It sounds sure, so it's right"

Of the ten, M2 has the most evidence behind it and the highest traffic. Overreliance, accepting an AI suggestion that is wrong, is resilient, and adding explanations does not reliably fix it (Buçinca et al. 2021). Fluent prose hides misinformation. In logged interactions, users who over-relied were identifiable by behaviour: frequent copy-paste, skipping task comprehension, coarse navigation, and accepting a claim despite visible hesitation (Zhang et al. 2026). Students describe the same texture from the inside: fabricated citations, overconfident but misleading answers, persistence in a wrong answer, sycophancy (arXiv:2602.17671).

The philosophical framing is that such a system is indifferent to truth rather than lying about it (Hicks et al. 2024), which is what makes a good corrective intuition possible: confidence is a writing style, generated by the same process as the content.

The anthropomorphism evidence points somewhere less obvious. In a preregistered experiment (N = 410), attributing intelligence to the model (reasoning, planning) correlated strongly with accepting its advice, while attributing experience (feelings, consciousness) did not, and trended the other way (Colombatto et al. 2025). So the "is it conscious?" argument is not the risk channel and does not need to be won. "Does it know?" is. And simply telling people to stop anthropomorphising is a weak intervention when the systems now mimic human communication convincingly (Peter et al. 2025).

Friction works, and people dislike it

Three findings constrain a teaching design more than any others, and they interact.

Forcing a pause reduces overreliance but is the least liked option. Cognitive forcing interventions significantly cut overreliance relative to simple explainable-AI approaches, yet drew the least favourable subjective ratings of any condition, and benefited participants high in Need for Cognition more (Buçinca et al. 2021). The authors audited this explicitly as an intervention-generated inequality. Anything that makes verification effortful will be skipped by the people who most need it.

Explanations are strategic, not inert. Across five studies (N = 731), whether an explanation reduced overreliance depended on the costs and benefits of engaging with it (task difficulty, explanation difficulty), not on explanation quality alone (Vasconcelos et al. 2022). The lever is the cost of checking, not the volume of explanation.

Help widens gaps before it closes them. Twenty-one novice programmers observed with generative tools split sharply: the students who accelerated used the tools differently from those who struggled, and the struggling students' metacognitive difficulties were amplified rather than relieved (arXiv:2405.17739). A beginner course has to teach the use pattern, not only the concept.

What this adds up to for teaching

Eight practices follow from the evidence. They are not equally well supported, and the tag on each says how. Direct means a user study speaks to it; Inferred means it follows from conceptual analysis or from a study of a neighbouring question; Judgement means it is a design call the literature does not settle.

1. 01 Name the wrong intuition before correcting it Inferred Laypeople build folk theories by decoding the phrase "machine learning", probing the bot, and reaching for analogies to familiar objects (Li et al. 2026), so the prior belief is already in the room. Each lesson opens with the plausible wrong version, stated sympathetically enough that a holder of it feels recognised rather than caught. The conceptual-change reasoning is the design's; the folk-theory finding is the study's.
2. 02 Teach distinctions, not a list of facts Inferred Five distinctions underlie ten misconceptions (Lin 2026, adapted). Organise lessons by distinction and treat misconceptions as symptoms. The framework is conceptual rather than tested on learners.
3. 03 Put something in the learner's hands first Direct Surveyed adults aged 50+ (n = 103, mean age 64) were motivated and thought AI literacy important, but did not understand it or know how to start, and strongly preferred hands-on learning to explanation (arXiv:2504.14649). Note the limit: this establishes a stated preference and a missing starting point, not that hands-on teaching produces better outcomes.
4. 04 Move both dials: self-efficacy up, trust calibrated Direct The two-dials result (Lee et al. 2025). Reassurance-only content raises the wrong dial. Correlational, so it motivates the separation rather than quantifying an effect.
5. 05 Make verification cheap instead of mandatory Direct No quizzes, gates or interstitials, since forcing is disliked and favours people who need it least (Buçinca et al. 2021). Pull the cost lever instead (Vasconcelos et al. 2022): a few checking habits short enough to survive a real week. That specific design response is a hypothesis, not a tested result.
6. 06 End every lesson on a behaviour, not a fact Direct Generative tools widened the gap between strong and struggling novices, and the difference was in how they used them (arXiv:2405.17739). Each lesson closes on something to do: ask the same question twice in fresh chats; ask for three sources on a niche topic and try to find them. The sample is 21 programming students, so the transfer to general adults is assumed.
7. 07 Spend the anthropomorphism budget on competence Direct Attack "it knows" and leave "it feels" alone (Colombatto et al. 2025), and expect plain warnings against anthropomorphising to be weak (Peter et al. 2025). Public metaphors for AI have also drifted toward warmth and human-likeness (Cheng et al. 2026), so the default pull runs the other way.
8. 08 One analogy per concept, with its limit stated Judgement Folk theories form by analogy, so an unbounded analogy becomes the next misconception. Follow every analogy immediately with where it breaks, and ban the ones that seed M1 or M3: "digital brain", "it reads the internet", "like a very fast librarian". No study tested this; it is inferred from how folk theories form.

Where the evidence runs thin

The honest reading is that the confident part of this literature is narrow. Two of the ten misconceptions (M6 on memory, M7 on reach) rest on conceptual analysis, not studies of what users believe. Self-efficacy as a theme returned only 6 papers in the 319-paper screened corpus, against 63 on trust and reliance, so the half of the thesis about confidence in yourself leans on one survey more heavily than is comfortable.

The participants are also a patchwork: knowledge workers, students, novice programmers, adults over fifty, focus-group members. None is a general adult population meeting AI for the first time, and the studies were run on particular products at particular moments. And the design built from all this has not been user-tested. The success conditions are written to be testable (say where the words come from without the word "search"; predict that the same prompt can give a different answer, and check it; name one thing the model cannot reach; describe one verification habit you would actually use; report that trying things carries no risk of breaking anything), and running that test is the obvious next step.

Open questions

• QDoes hands-on-first actually produce better calibration than explanation-first? The preference is documented; the learning gain is not. A comparison with trust calibration as the outcome, not satisfaction, would settle it.
• QDoes cheap verification get used? "Make checking cheap" follows from the engagement-cost result but has not been measured against forcing, either for uptake or for who benefits. The risk to test is the inequality Buçinca et al. flagged: that even cheap friction helps the already-careful most.
• QWhich misconception should come first? Competency frameworks say what a learner should end up able to do, not the order in which to dismantle wrong intuitions. The dependency graph used here is a judgement call, and a different order might teach faster.
• QDo the distinctions survive contact with changing products? M6 and M7 describe memory and reach, and those are exactly the features products keep changing. What does a durable lesson on them look like when the underlying behaviour moves every few months?