The clearest signs are behavioural: you cannot explain a result without reopening the chat, 'looks plausible' replaces checking, and you stop drafting first. Appian's October 2026 survey found 60% of workers have caught AI errors at work, yet only 14% fully trust its outputs. Paul Bloom's Psych explains why feelings mislead: people confabulate, so audit a weekly log instead.

The question "am I relying on AI too much?" is hard to answer by feel, and the reason is older than AI. Paul Bloom's Psych gathers decades of evidence that people confabulate: we produce confident explanations of our own behaviour that have little to do with its real cause. Ask yourself whether you lean on the chatbot too much and you get a fluent, flattering answer from the same machinery. The signs worth trusting are behaviours you can observe, not feelings.

This autumn's research gives a sharper list than "you use it a lot." In a paper in the Academy of Management Review, covered by Workplace Insight on September 18, 2026, Dirk Lindebaum of the University of Bath and colleagues describe "epistemic de-skilling": managers gradually losing know-how because they outsource the thinking, most often under time pressure. The counter-case, "epistemic up-skilling," happens when managers use AI to test their reasoning and know they will have to justify the decision. Lindebaum's line is the best single test I have found: what the AI cannot explain, you must explain to yourself and to others.

The second sign is that checking has quietly stopped. Appian's survey of Australian workers, released October 8, 2026, found 60% had personally caught an AI tool making an error at work, only 14% fully trust the AI outputs their organisation uses, and more than one in five said employees are discouraged from challenging AI recommendations. When a team stops questioning outputs, the error rate does not fall; it just stops being seen. The third sign is a shift in confidence: Inside AI's September 29, 2026 summary of HBR research reports that executives who used generative AI made worse predictions, becoming more optimistic and less accurate.

Put together, the observable signs are these. You cannot defend a result without reopening the chat. "It looks plausible" has replaced an actual check. You no longer write a first draft or a first opinion before you ask. Decisions about people, a hire or a hard message to a co-founder, go through the model and come out nearly verbatim. And you feel relief when the answer matches what you already knew, the habit covered in why people keep asking AI things they already know.

Because self-report is unreliable, audit a log rather than a feeling. I wrote a sample week of seven AI uses of the kind a founder might have and ran a strict audit prompt on it. The log is invented for the test; the run is real.

Tested October 14, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

I want an honest check on whether I am over-relying on AI at work. Below is a log of what I used AI for this week. For each item, label it as one of three things: (A) delegation: AI produced output I could have produced myself, and I checked it; (B) thinking partner: AI challenged or extended reasoning I had already done; (C) substitution: AI did the thinking I am paid to do, and I could not explain or defend the result without it. Be strict: if the log does not say I drafted first or checked the result, assume I did not. Then name the one item I should take back from AI next week, and give me one question to answer on paper before I open the chat for that kind of task. Do not reassure me and do not give general advice about AI.

My log:
1. Monday: asked Claude to summarise a 40-page customer contract before a call; I read the summary, not the contract.
2. Monday: had it rewrite my […]

The start of what came back, unedited:

## Labels

**1. Contract summary: C.** You didn't read the contract, so you couldn't defend anything said on that call beyond what the summary chose to include. Knowing the terms of a customer contract is your job. You also have no way of knowing what the summary left out.

**2. Investor update tone rewrite: A, with a caveat.** You drafted first, so the substance is yours. The log doesn't say you checked the rewrite, though. Tone edits can quietly soften a number, a risk, or an ask. This counts as delegation only if you compared it line by line against your draft before sending.

**3. Hiring pick: C.** This is the clearest case. You handed over the decision and adopted the result. Your interview notes were the input, but the judgment was the AI's. If the hire asks why you chose them over the other candidate, your honest answer is "the model said so."

**4. Churn SQL: C.** "Looked plausible" is not a check. The query encodes a definition of churn: which customers count, what time window, how reactivations are handled. You didn't make those choices or verify them. If someone challenges […]

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

What it did well: it was strict in the way the prompt asked. It labelled five of the seven items as substitution and found the pattern on its own: the only two uses that escaped were the two where I had written something first. It caught a phrase most people would let pass, "the numbers looked plausible," and pointed out that a churn query encodes a definition of churn nobody chose. Its pick of what to take back was reasoned rather than obvious: hiring was the bigger lapse, but weekly priorities recur and steer everything else. The question it gave, which one thing gets done if only one does and why it beats the second, is usable as written.

What it did badly: it applied my strictness rule without weighing stakes, so a routine contract summary and a hiring decision got the same label. It cannot tell whether a log is honest, and a log written to look good will pass. It also has no sense of time: one heavy week before a deadline looks identical to a habit. Run it weekly on your real log and watch whether the substitution count falls.

If the audit comes back mostly substitution, the answer is not quitting; see how to break AI dependency without giving up AI at work. The short version is the rule the test surfaced: write first, then ask.

Sources: Over-reliance on Gen-AI may weaken moral insights, understanding, and know-how, Workplace Insight, September 18, 2026; Six in 10 Australian Workers Have Caught AI Getting It Wrong at Work, Appian, October 8, 2026; HBR Research Reveals AI Intensifies Work and Erodes Judgment, Inside AI, September 29, 2026; Paul Bloom, Psych: The Story of the Human Mind, Ecco, February 28, 2023.


Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success