Vibe coding creates confident idiots: 4 studies

Vibe coding creates confident idiots when a tool returns a working result without transferring the reasoning that produced it, and the gap fills with confidence instead of skill. Four studies published between 2025 and 2026 measured that gap: weaker self-explanation, worse debugging, and more overestimation of your own performance.

Vibe coding creates confident idiots: what the claim actually means

Vibe coding creates confident idiots when an AI tool returns the finished output of someone who studied the problem without transferring any of that study to the person who prompted it. The phrase comes from a video by YouTuber Cole Hastings, who argued that the gap left by the skipped learning gets filled with confidence instead of skill.

The vocabulary matters here. A midwit, in the bell-curve meme, is the person in the middle who learned a little and now overcomplicates everything to look smarter. Hastings' argument is that an AI-assisted builder can look like that middle without having done the middle's work: the tool did the studying, the human kept the posture.

That distinction explains why the insult lands on a specific character rather than on everyone who uses AI. The product manager who generates a dashboard in ten minutes with Lovable, a prompt-to-app builder, and then tells a senior developer that programming has become a commodity is the target. The tool is incidental; the unearned certainty is the point.

The rest of this article separates what has actually been measured from what is anecdote. The measurements are narrower than the slogan, and they are more useful because of it.

What the MIT Media Lab essay study measured in 2025

The MIT Media Lab study "Your Brain on ChatGPT" recruited 54 participants for an essay-writing task and recorded EEG activity in three conditions: essay writing with ChatGPT, with Google search, and with no tools at all. The reported pattern was a gradient in measured brain connectivity, strongest in the no-tool group and weakest in the ChatGPT group.

The finding that travels furthest is a recall problem. When asked to quote a passage from the essay they had just produced, participants in the ChatGPT group had trouble reproducing their own text. The preprint, released in 2025, links that difficulty to weaker memory encoding of material generated by the tool rather than authored by the person.

Two caveats belong next to that result. It is a preprint covering 54 people doing one writing task, and the lead researcher, Nataliya Kosmyna, has publicly stated that the study found no "brain rot" effect and did not measure IQ. Anyone claiming ChatGPT makes you stupid is going past what the data supports. ChatGPT, OpenAI's assistant, is the tool the study examined; the claim about permanent cognitive damage is not in the paper.

Read narrowly, the study supports one practical warning: if you cannot recall or explain what the tool produced for you, you did not author it.

Anthropic 2026 trial: AI help, worse debugging

Anthropic, the company behind Claude Code, reported a small trial in which 52 developers, most of them junior, learned an unfamiliar Python library. Roughly half used AI assistance and half worked without it, and both groups took a quiz on the material afterward.

The reported quiz scores were 50% for the AI-assisted group and 67% for the group that coded without AI. The AI group was not meaningfully faster, and the study identified debugging as the biggest bottleneck: locating the failure, understanding why it failed, and reasoning about the cause. That is the skill that also lets a developer notice when the model has hallucinated or reasoned incorrectly.

The same report contains the counterexample that matters most for practice. Within the AI-using group, some participants learned at least as well as the manual group. The behavioral difference was how they used the tool: asking why a solution worked, requesting an explanation, and writing parts of the code by hand rather than delegating the whole task and the debugging to the model.

The lesson is not that AI assistance prevents learning. It is that delegating the explanation and the debugging does.

Aalto University 2026: assisted answers, overestimated scores

A separate study led by researchers at Aalto University in Finland asked around 500 participants to solve a logic problem, half with access to ChatGPT and half without. The AI-assisted group answered more accurately overall, and participants in that group also overestimated how well they had done.

The researchers observed the usage pattern rather than guessing at it: most participants submitted a single prompt per question, copied the answer, and moved on without verifying or following up. Self-assessment did not track real accuracy, which is the more consequential finding for anyone who grades their own work.

The Aalto finding is not that using AI makes you wrong. It is that the accuracy gain measured in the study did not translate into a better sense of your own competence, and the mismatch is what makes the overconfidence durable.

Microsoft Research and Carnegie Mellon 2025: trust tracks with less scrutiny

A study from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who used generative AI at least weekly. The survey found a negative association between trust in the tool and critical-thinking effort: participants who trusted the AI more reported applying less scrutiny to its output.

The flip side was also reported. Participants who trusted their own judgment more tended to question the AI's output and evaluate it against their own understanding. Confidence in the tool and confidence in yourself pull in opposite directions on the amount of review the work receives.

Combine that survey with the debug findings and the mechanism becomes plain. Whoever uses AI the most questions it the least, whoever questions it the least believes they performed better, and the belief gets shipped as a spec or a pull request.

Why explanation and debugging erode first

Explanation and debugging erode before raw output quality does because both depend on friction that AI assistance removes. Generating an answer and recalling why the answer works are different operations, and only the first one happens during a prompt-and-paste loop.

Dr. Dre told an interviewer in 2026 that he uses AI in music production and that the people most threatened by it are the ones who struggle to create. That fits the observed pattern: practitioners with existing skill can use the tool as one more instrument because they can evaluate what it returns. The punk framing goes further in the other direction; producer Rick Rubin has called vibe coding the punk rock of programming, pointing out that punk needed an idea and nerve rather than technique.

The analogy has a limit worth stating. Garage punk bands played three chords badly and improved over years of performing, and the audience could hear the effort. Someone shipping AI-generated code they cannot explain has the garage band's confidence without the practice that eventually made the band good.

A practical protocol against unearned confidence

Explain the generated code out loud before you accept it. The old rubber-duck method works without the duck: say what the function does, what it assumes, and what it will do at the edges. If you cannot finish the sentence, you have found the part you do not understand yet.

Separate the projects that pay from the projects that teach. Keep a build-for-learning repository where the AI is allowed to explain but not to complete, and keep the production work under normal delivery constraints. The two modes have different goals and should not share a default posture.

Sit with a bug before delegating it. A delay of twenty to thirty minutes before pasting the error into a chat window is the friction where debugging skill survives. Searching on Stack Overflow or reading the library's own error documentation counts, and so does asking the model for a hint rather than the fix.

Run a short self-check each week:

FAQ

• Does using ChatGPT or Claude Code make you dumber? Nothing published as of 2026 shows a permanent drop in ability from using AI assistants. The MIT Media Lab essay study, a preprint with 54 participants, found weaker brain connectivity and poorer recall of self-generated text, and its lead researcher has said it found no brain-rot effect and did not measure IQ.

• What is an "AI midwit"? It is a term from YouTuber Cole Hastings describing someone who skipped the learning and kept the confidence. The pre-AI midwit studied a little and overcomplicated things; the AI midwit never studied and still behaves like the expert, because a working output arrived without the reasoning behind it.

• Which task loses skill first with AI assistance? Debugging. Anthropic 2026 trial on an unfamiliar Python library reported its biggest bottleneck in understanding where code failed and why, which is also the skill you need to notice when a model is wrong.

• Why does more AI literacy lead to more overconfidence? Overconfidence rose alongside literacy in the Aalto University 2026 logic study, where most participants sent a single prompt per question and copied the result. Prompt fluency measures control of the tool, not the ability to judge whether its answer is correct.

• How do I keep learning while still shipping with AI? Ask the model why the code works, write parts by hand, request hints instead of complete fixes, and explain each result out loud before accepting it. Participants in the Anthropic trial who took that approach learned about as well as the group that worked without AI.

• Does AI assistance actually slow developers down? Not established here. The Anthropic trial reported a quiz-score gap of 67% versus 50% and did not describe the AI group as substantially faster, so the case against AI assistance should be made on verification and understanding, not on a universal productivity claim.

• Should I stop using AI to code? No credible source reviewed here recommends that. The pattern that eroded learning was delegating both the writing and the debugging to the tool, not using the tool for explanation, review, or reference.

• What does punk rock have to do with vibe coding? Producer Rick Rubin called vibe coding the punk rock of programming, meaning an idea and nerve can substitute for formal technique. Punk bands practiced for years afterward, which is the part of the analogy that AI-assisted builders often skip.

• How much of coding do these studies say AI can replace safely? None of the four studies measured replacement, and all four measured learning or self-assessment on narrow tasks. Treat the confidence gap as a personal risk signal during code review, not as a verdict on what the tool can or cannot build.

Turn filmed knowledge into a written article

The four studies above only matter if the people who understand them write down what they mean; a video explaining the confidence gap can hold real knowledge that is hard to search, quote, or return to, and a written version is what lets a reader check a specific number without rewinding.

If you have that kind of knowledge sitting in a YouTube video, an explanation, a lesson, or a breakdown of what the research says and how to apply it, Skala Blog turns the recording into a structured article by transcribing the video and drafting it for review.

Paste a YouTube URL into Skala Blog, let the video become a transcript, and turn the transcript into an article you can edit and publish.

Source video