Most People Have the Wrong Mental Model of AI. The Gap Is Getting Dangerous.

The most consequential AI knowledge gap right now is not between technical people and everyone else. It is between people who tried ChatGPT once eighteen months ago and formed a view, and people who have been using frontier tools daily since then. These two groups are sitting in the same meetings, making the same strategic decisions, and working from completely different pictures of what AI can actually do. One of them doesn't know it.

The gap is widening, not closing

Andrej Karpathy posted about this dynamic on X in early April — and it kept circulating through the week of April 21 because it named something real. The people whose view of AI is based on early ChatGPT encounters are not simply less informed. They are working with an actively inaccurate picture of what these tools can do today. The free-tier, mid-2024 version of ChatGPT and the frontier paid tools running now are in the same product family the way a sedan and a Formula 1 car are both "cars." People whose experience is the sedan are making judgments about the Formula 1 car they have never driven.

Ethan Mollick confirmed the enterprise version of this gap on April 23: senior leaders at major organisations now understand AI capability well enough to set strategy around it, but translating that understanding firm-wide is, in his words, the unsolved challenge. In most teams, it isn't happening. Leaders have a current model; the people executing work have the old one. That combination produces exactly the friction that makes AI rollouts fail — not because the tool doesn't work, but because the mental models are out of sync.

Simon Willison highlighted public sentiment research on April 24 that completes the picture: only 18% of Gen Z feels hopeful about AI, down from 27% the prior year, with 31% reporting feeling angry about it. Willison's read: "People do not yearn for automation." The framing of AI as a thing that replaces human work is producing resistance even among groups expected to be enthusiastic. The mental model problem is not just about capability — it's about what people believe the whole project is for.

What to do about it this week

If your mental model of AI is based on what it couldn't do in 2024, you are making career and workflow decisions based on a tool that no longer exists. Three specific things to close that gap:

Try a task you've written off. Most people who think AI is overhyped have a list of specific things it couldn't do when they tried it. Run one of those tasks again, on a current frontier model, with a clear and specific prompt. Some failures will still fail. Some won't. You need to know which is which.

Test the paid tier for two weeks. The gap between the free ChatGPT experience and Claude Sonnet or GPT-4o with extended context is not a small refinement. It is a different capability class. If you have never paid $20-30 for a month of a frontier tool, you are forming opinions about a product you have not actually used.

Reframe from automation to delegation. Willison's sentiment data points to a word problem. "Automation" means your work gets replaced. "Delegation" means you hand off the parts of your job you find least valuable so you can do more of the parts that matter. These are not the same conversation, and the difference in how colleagues react when you use one word versus the other is real.

What this looks like in practice

A senior account manager at a 90-person professional services firm heard her leadership team talk about AI "transformation" at an all-hands in February. Her mental model: AI is taking entry-level work, so she's safe for now. She hasn't tried it seriously since a 2024 ChatGPT session that produced generic output she couldn't use for anything.

What she doesn't know: her counterparts at competing firms are using AI to prepare client meeting briefs in 15 minutes instead of 90, run three proposal drafts in parallel before choosing which to develop, and flag renewal risk from usage patterns before calls. The gap between her and them is not talent or effort. It is six months of daily use — on better tools, with a more developed approach — compounding.

A harder version of the problem

A strategy director at a mid-size consultancy tried to roll AI into his team's proposal workflow in late 2024. It didn't work — the associates didn't trust the output, the partners weren't interested in reviewing AI drafts, and the tool (free-tier Gemini at the time) produced exactly the generic summaries everyone was worried about. He concluded AI wasn't ready for professional services.

Eighteen months later, a competitor's strategy director is using Claude and GPT-4o across the same workflow. Not without friction — there is a verification step, a clear protocol for when AI drafts go forward versus when they don't, and a specific briefing approach for complex analysis. The tool category is the same. The method is different. The first director's mental model was formed by one failed experiment and has not been updated since.

The experiment failed. The conclusion — "AI isn't ready for this" — was wrong, or at least is now out of date. Those two things are not the same.

The one thing that stays true

The professionals Karpathy is describing — the ones working with frontier tools daily — are not mindlessly outsourcing their judgment. The ones doing it well are doing something specific: they are building a clear sense of where AI makes their work better and where it makes their thinking lazy. That calibration is the actual skill.

If your mental model is out of date, updating it takes more than reading about AI. It takes using it, on real work, until you have enough reps to form an accurate opinion. That's not a large time investment. Two weeks of serious daily use is enough to fundamentally change what you know.