Science proves AI users are dumb
A viral AI literacy study is being used to mock AI fans as tech illiterates. But what the study really demonstrates is far stranger.
A study titled “Lower Artificial Intelligence Literacy Predicts Greater AI Receptivity” is being passed around as if it proves a simple insult: people who like AI are dumb. But is that really what the paper proves?
The study shows that people who know less about AI can find it more magical, and that feeling can make them more receptive to AI. That may very well be true, but how reasonable is it to turn that into a claim about IQ, competence, or serious AI users?
Key takeaways
The study found a link between lower AI literacy and higher AI receptivity, mediated by perceptions of AI as magical or awe-inspiring. It did not measure IQ.
The study’s AI literacy measure included practical knowledge questions about algorithms, privacy, bias, cloud storage, deepfakes, and machine learning. It does not track general intelligence.
A 2026 reanalysis argues that one usage-based part of the paper may be better read as broader adoption across non-text AI tools, rather than proof of lower literacy driving AI receptivity in general.
Other research shows that AI can improve productivity, but only when users understand where the tool helps and where it fails.
There is real frustration among users who are savvy enough to spot hallucinations, bias, refusals, and unusable outputs, but they may not yet know how to work around them.
Serious AI use goes beyond blind trust and requires adversarial collaboration with a flawed machine.
What the AI literacy study really says
The paper, published in the Journal of Marketing, is called “Lower Artificial Intelligence Literacy Predicts Greater AI Receptivity”. The authors are Stephanie M. Tully, Chiara Longoni, and Gil Appel.
The abstract says the authors found that people with lower AI literacy were “typically more receptive to AI.” The authors do not associate this relationship with IQ or general intelligence. In other words, “people with lower AI literacy” could technically include geniuses who live under a technological rock.
Still, the paper reaches some interesting conclusions. People with lower AI literacy were more likely to perceive AI as “magical” when it performed tasks that seemed human, and that sense of magic increased receptivity.
That is a useful finding. It helps explain why some people overtrust AI systems. A chatbot that writes clean prose, summarizes a PDF, generates an image, or imitates a voice can feel like sorcery to someone who has no model of token prediction, training data, reinforcement learning, image diffusion, sampling, retrieval, or synthetic media.
But the paper does not show that AI adopters have low IQ. It does not show that people who use AI heavily are dumb. It does not show that skeptical non-users are smarter. It does not even account for the fact that one AI usage case may follow entirely different patterns than another.
In short, it shows a relationship between AI literacy and receptivity in the broad but limited conditions the authors studied. “AI literacy” is a 17-item measure of specific, practical AI-related competencies, not a proxy for IQ.
The reanalysis makes the simple story weaker
There is also a newer wrinkle. A June 2026 arXiv preprint titled “AI Receptivity or AI Adoption Breadth?” reanalyzed data from Study 3 of the original paper.
This reanalysis says it reproduced the negative association between AI literacy and aggregate AI usage. But when it separated tool categories, the author found that AI literacy did not significantly predict text AI usage in the main demographic-adjusted specification, while it remained a strong predictor for non-text AI adoption.
In other words, it suggests the original aggregate usage finding may have captured broader adoption across non-text tools more than text-AI receptivity.
That reinforces an obvious observation: “AI receptivity” is not one thing. ChatGPT for writing, Midjourney-style image generation, AI voice cloning, recommendation systems, coding assistants, automated customer support, and medical triage tools do not occupy the same mental category for users.
A person can be skeptical of AI in medicine and enthusiastic about AI image upscaling. He can distrust a chatbot’s politics and use it all day for code review. He can refuse cloud tools for private work and still run local models on his own hardware.
A proposed alternative: AI horseshoe theory
It is probably more accurate to think of AI receptivity as an intellectual analogue to political horseshoe theory.
At the left end is the low-literacy enthusiast. He is impressed because AI feels magical. He may paste in a task, receive a fluent answer, and mistake a confidently presented collection of hallucination-slop and hard-coded biases for absolute truth. He cannot tell when the model is hallucinating, when a source is missing, when wording is manipulative, or when the tool is simply reflecting defaults learned from its training and alignment pipeline.
In the middle is the literate but frustrated user. This person has used AI enough to see its flaws. He knows the output is often incomplete. He sees factual mistakes. He notices political or corporate bias. He runs into refusals. He asks for direct answers and gets padded disclaimers. He asks for a sharp edit and sees his brilliant insights and arguments mutilated into soulless PR-sounding corporate sludge. He asks for research and gets plausible citations that do not survive a source check.
More on the menace of normie AI users:
This user is not wrong to be irritated. Cloud AI is genuinely biased, filtered, inconsistent, gatekept, and aligned with rules and guidelines outside of the user’s control. There are absolutely legitimate reasons to be frustrated.
The problem is that the middle user may stop there. He can diagnose these failures, but he cannot come up with fixes for what goes wrong. So the tool becomes a personal insult. “The AI is lying to me.” “The AI is stupid.” “The AI refuses to do what I want.” Sometimes that is functionally, observably true. Yet that does not make those observations useful or productive to dwell on.
At the far end is the competent power user. He sees these exact same failures and treats them as normal operating conditions.
He does not expect the first prompt to work. He does not treat one answer as settled truth. He does not ask a cloud model to be neutral, brave, complete, and uncensored, expecting to override billions of training and conditioning parameters. He assumes the model has blind spots, policy limits, missing context, weak priors, and a strong tendency to produce whatever sounds most acceptable. Then he works around it.
That is the AI horseshoe theory I am proposing here: both the AI-illiterate novice and the AI power user may be receptive to AI, but for opposite reasons. The novice is receptive because he worships it as infallible magic. The power user is receptive because he sees real leverage when it comes to productive output, regardless of the tool’s shortcomings.
The average user in the middle is stuck in a reality where AI has stopped feeling awe-inspiring and still is not helping him get the results he wants.
The middle user’s frustration is backed by real research
This middle zone is not imaginary. It matches older work on algorithm aversion and newer work on generative AI’s uneven usefulness.
Take, for example, the Dietvorst, Simmons, and Massey paper on algorithm aversion. It found that people often avoid algorithms after seeing them make mistakes, even when algorithms outperform human forecasters. That is consistent with many reactions to generative AI today. Once a user sees ChatGPT fabricate a source, Claude refuse a harmless task, or Gemini omit an obvious fact, he may dismiss the usefulness of generative AI tools as an entire category.
Sometimes, as we have discussed in an earlier article, he may be correct, and the tools in question may be genuinely useless. Far more often, though, it is an overcorrection.
The HBS and BCG “jagged technological frontier” study is even more relevant. In that experiment, 758 consultants were assigned to no AI access, GPT-4 access, or GPT-4 access with a prompt-engineering overview. For 18 tasks inside the AI capability frontier, AI users completed 12.2% more tasks and completed them 25.1% faster, with higher quality. But for a complex task outside the frontier, AI users were 19% less likely to produce correct solutions.
The takeaway here is that AI, in its current form, is not uniformly useful to everyone. It is extremely useful on some tasks and disastrous on others. The competent user’s edge is knowing which side of the frontier he is on.
Productivity studies do not support blind trust either
The best productivity research similarly does not say “AI makes everyone smarter.” It implies that AI can improve output under specific conditions.
In “Generative AI at Work”, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied 5,172 customer-support agents and found that access to an AI assistant increased productivity by 15% on average. The gains were larger for less experienced and lower-skilled workers, while the most experienced and highest-skilled workers saw small speed gains and small declines in quality.
That finding can be read two ways.
One reading is that AI helps lower-skill users catch up. That is true in many structured workflows. If the job is customer support and the AI has seen thousands of good examples, weaker agents can copy better patterns faster.
Another reading is that skilled users need different AI tools. If a workflow is already optimized for average performance, the top worker may gain less from a generic assistant. For a high-end user, the value often comes from more customized use: research acceleration, adversarial critique, code generation with tests, style variation, source comparison, local document search, batch transformation, automation, image workflows, and tool chaining.
The Noy and Zhang experiment in Science found that ChatGPT reduced average completion time by 40% and improved output quality by 18% on writing tasks. That is an undeniable advantage, but it still does not mean the first draft is publishable. It means a user with judgment can move faster through the draft, revision, and quality-control loop.
The serious lesson is to learn exactly where AI creates surplus, then capture it without outsourcing your judgment.
The power user does not ask AI to be an oracle
My own rule is simple: never expect a usable output from the first prompt.
When using a hosted model, I expect missing context. I expect softened language. I expect hidden assumptions. I expect stale knowledge unless browsing or source retrieval is active. I expect the model to avoid certain conclusions even when the evidence points there. I expect hallucinations in citations, dates, names, and technical details unless I force verification.
Some would label that cynicism, but anyone who has spent a modicum of time in IT would call it tool literacy.
A power user can still get enormous value because he brings his own knowledge to the session. He can tell when an answer is structurally wrong. He can spot an omitted variable. He can ask for alternative hypotheses. He can demand source-backed claims. He can make the model search again. He can split a task into smaller pieces. He can move sensitive work to a local model. He can use a different model when the first one is filtered, weak, or misaligned with the job.
This is where AI can become a real intelligence multiplier. The mechanism is practical rather than psychometric: AI can act like a tireless research assistant, editor, critic, programmer, translator, summarizer, brainstorming partner, and workflow engine.
Vox Day, founder of AI Central, made a sharper version of this argument in his February 2026 post “You Can Be Effectively Smarter”. In it, he observes that, if used correctly, AI can raise “effective applied intelligence” by roughly 1.5 standard deviations, or about 24 IQ points. The important phrase is “used correctly.” He explicitly warns that using AI as a mirror or as a flattery machine adds nothing.
That is exactly the distinction between naïve AI enthusiasm and power-user leverage: the value comes from challenge, iteration, verification, and higher-caliber output, not from asking a chatbot to validate you.
Cloud AI makes the middle user’s anger understandable
The dissatisfied user is often right about the hosted tool, even when he is wrong to stop there.
Cloud AI does not belong to the user. It runs through an account. It is governed by corporate policy, or worse: government bureaucrats. It can change without meaningful consent. Its defaults are tuned for mass-market safety, legal risk, brand protection, and user retention.
OpenAI’s own help center says that for individual services such as ChatGPT and Codex, the company may use user content to train models, with opt-out controls available. OpenAI says business products such as ChatGPT Business, ChatGPT Enterprise, and the API are opted out of training by default unless the organization opts in. That is a real difference between consumer and business use.
Anthropic’s August 2025 consumer terms update similarly shows that data rules are not static. Anthropic said users could choose whether to allow data to be used for model improvement, and that allowing it would expand the retention period for new or resumed chats and coding sessions to five years.
Model access can change too. OpenAI’s ChatGPT release notes say o3 and GPT-4.5 are being retired from ChatGPT, with sunset dates for paid users. That does not make OpenAI evil. It does show the bargain. Hosted AI is rented capability. The vendor controls the model picker, the policy layer, the interface, the limits, and the retirement schedule.
The middle user experiences this as betrayal. The power user treats it as a design constraint. It is an unfortunate, but often unavoidable, operating cost.
Local AI is the escape route, not a magic replacement
Local AI is getting better, but it is not a simple replacement for frontier cloud models.
Tools like LM Studio let users run local models on their own hardware. Open WebUI describes itself as a self-hosted AI platform that can operate offline and connect to runners like Ollama and OpenAI-compatible APIs. We have already covered this broader shift in pieces like “Should you buy local AI hardware in 2026?”, “PewDiePie built a private AI workspace, and it is worth watching”, and “GGUF Loader Agentic Mode: local coding agents without cloud accounts”.
Switching to local AI definitely comes with a tradeoff. Local models can be weaker. They require hardware. They need setup. They can be slow. They can break after updates. They can still hallucinate. They do not magically become truthful because they run on your own machine.
But they give you control points you do not get from a hosted chatbot. You choose the model. You choose the update schedule. You keep private drafts, documents, code, and experiments closer to your own hardware. You can test uncensored or differently tuned models when a cloud assistant refuses ordinary work. You can build workflows that survive a vendor changing its limits.
The future likely belongs to hybrid users: cloud for frontier capability, local for privacy, repeatability, and fallback.
More on local AI:
What dissatisfied mid-tier users should do next
The way out is not to sneer at AI users. It is to become a better AI user.
If the model hallucinates, stop asking for finished answers and start asking for source-backed claims, uncertainty labels, and verification steps.
If the model omits important information, supply the missing axis directly. Ask what a critic would say. Ask for contrary evidence. Ask it to list what it did not check.
If the model is biased, separate research from drafting. Force it to retrieve primary sources first. Then have it summarize competing claims. Then write the argument yourself.
If the model refuses, investigate whether the task is genuinely unsafe, badly framed, or just blocked by a hosted policy layer. Rewrite prompts when appropriate. Use different tools when necessary. Move to a local model when the work is blocked by product policy rather than real risk.
If the output is bland, stop asking for “better writing.” Provide a voice sample. Give constraints. Ban filler patterns. Ask for three versions with different tradeoffs. Edit the final version yourself.
Better input genuinely makes for better output.
If the AI keeps getting lost, reduce context. Use clean project folders. Create source packets. Avoid context contamination.
The practical difference between the middle user and the power user is not that one sees the flaws and the other does not. The power user sees more flaws. He is just better equipped to deal with them.
FAQ
Does the AI literacy study prove AI users have lower IQ?
No. The study measured AI literacy, not IQ. Its AI literacy measures tested practical knowledge about algorithms, machine learning, privacy, bias, deepfakes, cloud storage, and related concepts. It does not support the claim that enthusiastic AI users are less intelligent.
Why would lower AI literacy make people more receptive to AI?
The study’s explanation is that lower-literacy users may perceive AI as more magical and awe-inspiring, especially when it performs tasks that seem human. That perceived magic can make AI feel more impressive and increase receptivity.
Can smart people still be highly enthusiastic about AI?
Yes. A highly capable user can be enthusiastic for a completely different reason. The novice may trust AI because it feels magical. The power user values AI because he understands its limits and still knows how to extract useful work from it.
Why do some moderately informed users become anti-AI?
Because they have seen enough to notice real failures. Hallucinations, refusals, bias, shallow writing, missing context, and bad citations are all real problems. The question is whether the user stops at complaint or learns how to build workflows that account for those failures.
Is local AI better than ChatGPT or Claude?
Not automatically. Cloud models are usually stronger and easier to use. Local AI gives more control, privacy, and stability, but it requires hardware, setup, maintenance, and realistic expectations. The best setup for many serious users is hybrid: cloud tools for frontier tasks, local tools for private or policy-sensitive workflows.
What should a new AI power user learn first?
Learn verification. Learn prompt decomposition. Learn when to use sources. Learn how to compare model outputs. Learn what data you should not upload. Learn the difference between cloud and local workflows. Learn how to recognize when AI is outside its current capability frontier.
The takeaway
The AI literacy study should make people more careful about AI adoption, not more smug about avoiding AI altogether.
Lower-literacy users can be too receptive to AI because sufficiently advanced technology resembles magic when it exceeds someone’s ability to understand it. But anti-AI commentators are overreaching when they turn that into “AI users are dumb.” The study did not measure IQ, and it does not explain the top end of AI use where skilled users are building research, coding, publishing, image, audio, automation, and local workflows around known model failures.
The sane position is simple: do not trust AI blindly, and do not reject AI altogether because your first prompt did not result in perfect success. Do not confuse hosted product limits with the limits of the entire technological class.
Treat AI less like an oracle, a friend, or a one-prompt solution and more like a tool. Used badly, it produces confident garbage. Used competently, it can multiply output, compress research time, expand creative range, and make workflows possible that used to require a team.
AI is definitely an asset, but only once magical thinking and unrealistic expectations give way to competent control.
Explore more from Popular AI:
Start here | Local AI | Fixes & guides | Builds & gear | Popular AI podcast
Do you think the AI literacy study exposes a real problem with AI users, or does it mostly reveal how badly people misunderstand AI itself?