Use AI to write or code, and you may start believing you're better at it than you are.

This explores how using AI changes the way people judge their own skills. The corpus points to one main effect: people credit themselves for what the AI did. Researchers call it the LLM Fallacy, a self-perception error where AI-assisted output gets absorbed into a person's sense of what they can do (Do AI-assisted outputs fool users about their own skills?). It is not hallucination or automation bias. It happens regardless of whether the output is accurate or whether the person over-relies on it (How does AI-assisted work reshape how people see their own abilities?). Making the AI more accurate won't fix it, because the problem is in how people see their own contribution.

Several things feed the illusion. Four mechanisms interact: it's unclear who did what, fluent output feels like skill, thinking gets handed off to the tool, and the pipeline behind the output is hidden. The effect is multiplicative, so each mechanism amplifies the others (How do AI tools trick users into overestimating their own skills?). The fluency piece is the most striking. People read the ease of a polished result as a sign of their own ability, and LLMs produce polish whether or not the user understood anything (Does processing ease mislead users about their own competence?). The illusion is strongest when the output is seamless and the human-AI boundary disappears.

Asking people how competent they feel won't catch this. A pooled analysis of three studies found a correlation of just .055 between self-reported and objectively measured AI competence, with confidence intervals that include zero (Can self-ratings replace objective performance scores for AI competence?). Real ability may also be moving the other way. In a four-month EEG study, heavier LLM use went with weaker brain connectivity, worse memory, and difficulty recalling one's own recent work (Does AI assistance weaken our brain's ability to think independently?). Even correct AI suggestions have a cost: they break the immersion that good reasoning depends on (Does AI assistance always help reasoning or does it carry hidden costs?). Feeling more capable while being measurably less engaged is a hard gap to see from the inside.

The same fluency trick works on other people too. In a study of 2,939 writers and 11,091 readers, AI assistance shifted how readers perceived the writer on all 29 dimensions tested, toward more confident, higher quality, and more agreeable (Does AI writing assistance change how readers perceive the writer?). Judges of AI models fall for it as well. Models trained to imitate ChatGPT fooled human evaluators with a confident style while closing no real capability gap (Can imitating ChatGPT fool evaluators into thinking models improved?). In every case, polish gets read as competence. The model can't reliably flag where its contribution ends either, since its self-reports are unstable (How well do language models understand their own knowledge?). And people judge AI partners mainly on perceived competence, which accounts for about half of the variance in their impressions (How do users mentally model dialogue agent partners?).

The unexpected takeaway is that a more seamless AI makes the misperception worse, because the seamlessness is what hides who did the work. The fixes the corpus suggests are about making the boundary between human and machine contribution visible, not about making the AI more correct.