Why do I keep asking AI things I already know, and how do I stop?

Checking ChatGPT for Things You Already Know: The 62% Problem

Because the check feels like safety rather than laziness, and it quietly recalibrates your confidence. In a Frontiers in Psychology experiment published September 20, 2026, 342 students given open ChatGPT support accepted 62.4% of incorrect advice; a brief reflection prompt cut that to 39.7%. Greene and Grant's Solution-Focused Coaching supplies the fix: answer first, find the exceptions, and scale the habit down one point at a time.

Someone reached this site by typing a sentence into a search engine that I want to quote in full, spelling and all, because it is the most honest description of the habit I have read: "i need to stop using ai, i'm becoming codependent, looking up everything, things that i know i already know, is anybody else having this problem." Yes. I had it. In the spring I caught myself asking a model for the syntax of a command I have typed thousands of times, and the uncomfortable part was not the lookup but the small relief when the answer matched. That relief is the whole mechanism, and this month two studies put numbers on it.

The one that matters most was published in Frontiers in Psychology on September 20, 2026. In a three-condition experiment, 342 undergraduates did reasoning and information-evaluation tasks alone, with open multi-turn ChatGPT support, or with the same ChatGPT support plus a brief metacognitive reflection prompt. Open ChatGPT support raised final confidence and came with 62.4% acceptance of incorrect AI advice. The reflection prompt cut acceptance of incorrect advice to 39.7%, with an odds ratio of 0.40, improved awareness calibration from 0.41 to 0.59, and halved an attribution-bias index from 0.42 to 0.21, meaning people stopped crediting themselves for the hits and blaming the model for the misses. Recommendation accuracy was the same in both AI groups, so the difference was entirely in what the humans did with it. A second paper, from Welsch and colleagues on arXiv on September 15, 2026, tested interventions on 917 people across twelve planning problems: per-task reliability cards and contrasting replies reduced overconfidence, pause points and post-problem reflection did less, and none of it improved task performance, which the authors state plainly: measured monitoring and task performance are separable targets. The University of Melbourne and KPMG's 48,000-person survey from 2025 gives the base rate, 66% of employees rely on AI output without checking accuracy.

Read together, the picture is not that you are lazy. The check is a confidence-regulation move. Asking the model something you know converts a flicker of uncertainty into a feeling of being right, and the Frontiers data says that feeling is loosely coupled to actually being right. The word the search query used, codependent, is closer than most of the research vocabulary: you are outsourcing the feeling of certainty, not the knowledge.

The framework I use with clients on this is Jane Greene and Anthony Grant's Solution-Focused Coaching, and the first thing it tells you is to stop asking why. Greene and Grant's evidence is that dwelling on the causes of a problem tends to perpetuate and exaggerate it, and that the working question is "how would you like it to be?" followed by "when it already works, what is different about those times?" Their tools are small and specific: exceptions, scaling on a one-to-ten line and moving a point at a time, and the House of Change, which says a habit only shifts when you touch thoughts, feelings, behavior and environment together. Here is the four-step version I run, which I call answer-first:

Answer first, then ask. Before any query about something you might already know, type your own answer into the prompt, then ask the model to check it. This is the reflection prompt from the Frontiers study made into a keystroke habit; I keep the sentence "If I am asking about something I probably know, ask me for my answer first" in my ChatGPT custom instructions and in a Claude project's instructions, so the tool itself enforces it. My own answer being wrong is fine. Not having one is the habit.
Collect the exceptions. For three days, note every time you almost checked and did not. What was different? Usually the answer is that you were mid-flow, or someone was watching, or the stakes were trivially low. Greene and Grant's rule is to do more of what already works, so schedule more of those conditions rather than fighting the urge in the others.
Scale it, one point. Rate the habit: ten is "I check everything," one is "I check only what I genuinely do not know." Most people start around seven. Pick what a six would look like in behavior, one fewer reassurance query per hour of work, and stop there for a week. A one-point drop that holds beats a resolution that collapses.
Rebuild all four corners. Thought: the check buys a feeling, not a fact. Feeling: the discomfort of answering unaided is, in Greene and Grant's words, a sign of moving forward, not a signal to retreat. Behavior: the answer-first keystroke. Environment: the AI tab closed during writing blocks, the app off the dock, the model reachable but not ambient.

What I noticed after a month was that my query count barely dropped and the queries changed. I still ask the model constantly, but I ask it to disagree with an answer I have already written, and the Frontiers finding explains why that matters: the reflection group did not use AI less or reject it more, they discriminated better between useful and misleading advice. The goal was never abstinence. Abstinence is the problem-focused answer, and it fails for the same reason most diets do.

The limits are in the studies themselves. Both samples are students or online panels doing bounded tasks, not founders making decisions with money attached, and the Frontiers authors say their findings concern immediate discrimination, not long-term learning. The Welsch paper found that better self-monitoring did not improve performance, so answer-first will make you more honest about what you know before it makes you better at anything. And scaling only works if you actually count; the week I stopped logging, the number crept back to seven. That is not a failure of the method. It is what a habit is.

Sources: "College students' metacognitive awareness of generative-AI reliance: an experimental study of decision confidence and attribution bias," Frontiers in Psychology, published September 20, 2026 (frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1926110/full); Welsch et al., "Beyond 'ChatGPT Can Make Mistakes': Designing Interventions to Support Metacognitive Monitoring in AI-Assisted Work," arXiv:2609.17065, September 15, 2026 (arxiv.org/abs/2609.17065); University of Melbourne and KPMG, "Trust, attitudes and use of AI," 2025, as tabulated in Digital Applied's "Who Gains Most From AI at Work," September 23, 2026 (digitalapplied.com/blog/who-gains-most-from-ai-at-work-studies-compared); Jane Greene and Anthony M. Grant, Solution-Focused Coaching: Managing People in a Complex World (Pearson Education, 2003).