Am I becoming dependent on AI?

Dependency is not about how often you use it. Plenty of people use these systems constantly and remain entirely capable without them. It is about whether you could still do the work if it disappeared tomorrow, and whether anyone, including you, has checked recently.

This page is a diagnostic rather than an essay. Six questions, an honest reading of what the answers mean, and a straight account of what the evidence does and does not support.

Six questions#

1 · When did you last complete a piece of real work in your domain without assistance? If you cannot name a recent instance, that is not proof of anything, but it means you have no current information about your own capability. Most people discover they are working from a several-year-old estimate of themselves.

2 · Do you form a view before you prompt, or after? If the model speaks first, its framing is in the room and the alternatives you never saw are gone. This is the single highest-leverage habit and it costs about ninety seconds. See Human at the Start.

3 · Could you tell if the output were wrong? Honestly, in your own domain, on a plausible-but-incorrect answer rather than an obviously silly one. If no, you are not reviewing, you are approving. That is the capability test, and the one that matters most.

4 · Has the difficulty gone, or has the work gone? Removing tedium is a straight gain. Removing the effortful step that built your judgement is not, and the two feel identical in the moment. The distinction is desirable difficulty. It is the reason this question is hard to answer from the inside.

5 · When it is unavailable, do you postpone the work? Mild inconvenience is normal. Genuine inability to proceed on work you used to do unaided is the clearest signal on this list, and the only one that shows up without anyone testing for it.

6 · Whose voice is the output in? If you no longer recognise your own writing, thinking or approach in what you produce, something has been substituted rather than assisted. That may be fine for a status report and is not fine for the work you are known for.

How to actually test it, rather than wonder#

Pick one real task in your domain that you would normally hand over. Do it unaided, to completion, and note where it was harder than you expected. Then do it again with assistance and compare. The gap is your answer. It costs about twenty minutes a month.

Nobody does this, so almost nobody knows. Performance with the tool is the only measurement most people ever take. It is the one measurement that cannot detect the thing they are worried about.

Four studies, and none of them can tell you about yourself#

What can be supported is narrower than the anxiety.

There is one good study of persistence. Bastani and colleagues found that when access to a GPT-4 tutor was withdrawn, students who had used an unrestricted interface scored 17 per cent lower than students who never had access, while a guardrailed version largely removed the harm. One study, one subject, one age group, unreplicated. It is the most important unreplicated finding in the field.

There is suggestive but weak evidence on thinking. Three studies point the same way, and all three have design problems: one is self-reported, one is correlational and cannot establish direction, one has a small sample. Agreement between three weak designs is suggestive, not strong.

There is one measurement of the manner rather than the amount, and it explains why this page has to be a diagnostic. Zhu and colleagues followed 589 students and early-career workers across three surveys at two-week intervals, published in Frontiers in Psychology in July 2026, and separated two ways of handing work over. Dependent offloading means accepting the output with little evaluation and letting it structure the reasoning. Autonomous offloading means treating the output as something to argue with. The two turned out to be almost unrelated to each other, at r = 0.08, so they behave as two variables and not as two ends of one dial. Only the dependent kind tracked what the authors call cognitive agency transfer. The part that bears on question six is what happened to felt benefit: both modes felt equally useful at the time, r = 0.22 apiece, and that feeling barely tracked the later self-assessments at all, r = 0.06 on capability and zero on independent judgment. Everything in it is self-reported on scales the authors built for the study and call preliminary, they decline any causal reading themselves, and the sample is young. So it cannot tell you whether your capability has moved. It can tell you that the sensation of a session going well is not the instrument you thought it was.

There is no longitudinal evidence at all. No study has measured unaided capability after sustained use over years. Zhu and colleagues do not change this: four weeks of self-report with no baseline wave is not a measurement of capability over time, and reading it as one would be the error their own limitations section warns against. Anyone telling you confidently that AI is or is not making you worse is going beyond what exists, including anyone doing it from this site. See what we actually know.

So the useful posture is neither alarm nor dismissal. It is measurement, because you can measure your own case cheaply even though the field cannot yet measure the general one.

If the answers worry you#

The response is selectivity rather than abstinence, and it turns on one question: which capabilities are you paid for, and which are incidental?

Keep the repetitions that build the judgement you are valued for, and delegate the ones that do not. A lawyer should probably keep drafting the argument and can happily delegate the formatting. An analyst should keep framing the question and can delegate the chart. The mistake is using it uniformly rather than using it heavily, so that the load-bearing practice disappears alongside the tedium without anyone deciding.

And write down what you are protecting. An intention that lives only in your head loses to deadline pressure every time.

Related SuperSkills research#

The practice version is using AI without dependency. On the mechanism, cognitive offloading and desirable difficulty. On detecting error, how do I know when AI is wrong. On the organisational version, capability debt. See should AI remember everything about me. If the six questions above assume more than you have yet done with these tools, how to use AI at work is the same argument written for somebody starting out.

Key sources

• Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025.
• Zhu, Q., Li, X., Dong, Y., Chang, P. and Fan, M. (2026). Not all cognitive offloading is equal. Frontiers in Psychology, 17.
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9).
• Dahmani, L. and Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory. Scientific Reports, 10.

About this page#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This is a self-diagnostic, not a clinical instrument, and it has not been validated. It is offered because the alternative most people have is worrying without measuring.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.