Using AI without dependency

Dependency has little to do with volume. Some of the most capable people I work with use AI constantly and are not remotely dependent on it, and some of the most dependent use it twice a week. The distinction is simpler and harder than volume: leverage is using AI to do more with a capability you hold, and dependency is using AI in place of a capability that is draining away. There is one reliable test. It is uncomfortable. Take the tool away and see what happens. If your work gets slower, that is leverage working as intended. If your work gets worse, or you cannot start at all, that is dependency, and it accumulated without you agreeing to it. The practical answer is to decide, deliberately and in advance, which part of the thinking you keep, and then to keep it whether or not the tool is open. Using AI less has nothing to do with it.

Leverage and dependency are different things#

Most advice on this subject fails because it treats AI use as a quantity, and issues guidance about moderation. That framing does not survive contact with real work. A surgeon who relies on imaging is not dependent on imaging; a pilot who flies with autopilot is not dependent on autopilot, so long as the aircraft can still be flown by hand when the automation disengages at night over the Atlantic. The word we want is retained capability, which is a different target from moderation.

So the honest question replaces the volume question with three narrower ones. Could you produce a competent version of this without the tool, if slower? Could you tell a good output from a confident wrong one? And did you decide what you were trying to achieve before the model told you what was achievable? A yes to all three is leverage at any volume. A no to any of them is dependency at any volume.

Cognitive offloading, before AI and since#

The underlying mechanism predates AI by decades. Psychologists call it cognitive offloading: using an external tool to reduce the mental demand of a task. Risko and Gilbert, in a 2016 review, showed something important about how we decide to do it. We offload not only when a task is genuinely hard, but when we judge it to be hard, and that judgement is frequently wrong. We hand away work we did not need to hand away, and with it the practice we would otherwise have had.

Where that leads has been studied in two long-running natural experiments. Sparrow, Liu and Wegner, writing in Science in 2011, described the Google effect: when people expect information to remain available, they remember where to find it rather than the thing itself. Dahmani and Bohbot, in 2020, found that habitual satellite-navigation users had worse spatial memory when asked to navigate unaided, and that heavier GPS use over the following three years was associated with a steeper decline still. When a capability is reliably performed by an external system, the human version of it weakens. Memory and navigation went first. Reasoning is simply the next function in the queue, and considerably more central to professional life than either.

The workplace evidence on generative AI fits the same shape. The 2025 Microsoft Research and Carnegie Mellon survey of 319 knowledge workers, covering 936 real uses of AI at work, found that higher confidence in the tool was associated with less critical thinking, and that the thinking that remains changes character: from producing to verifying, from solving to integrating, from doing to supervising. Michael Gerlich's 2025 study of 666 participants found a negative correlation between frequent AI use and critical-thinking scores, with cognitive offloading as the mediating mechanism and the effect strongest among the youngest users, who have been offloading longest.

Two results give the practical warning teeth. In 2023, Dell'Acqua and colleagues, working with Boston Consulting Group and researchers at Harvard, MIT and Wharton, ran a controlled experiment with 758 consultants using GPT-4 and described what they found as a jagged technological frontier. On tasks inside the model's competence, AI-assisted consultants were dramatically better. On a task designed to sit just outside it, consultants using AI performed worse than those with no AI at all, because they accepted confident output they should have interrogated. That is automation bias, which Parasuraman and Manzey, reviewing decades of aviation, medical and military work in 2010, showed appears in novices and experts alike, cannot be trained away, and worsens under load. And in a 2025 field experiment published in PNAS, Bastani and colleagues found that high-school students given unrestricted GPT-4 access during practice performed 17 percent worse than a control group once the tool was removed, while a version designed to give hints rather than answers largely eliminated that effect.

Self-report and correlation#

Much of the workplace evidence is self-reported or correlational, and that limitation should be stated rather than glossed. The Microsoft and Carnegie Mellon study asked workers to describe their own thinking, and people who already think differently may well use AI differently; a survey cannot separate the two. The Gerlich study establishes a correlation, not a direction of causation, and it carries a published correction from September 2025 that anyone citing it should read alongside it. The MIT Media Lab study on cognitive debt is suggestive and widely quoted, but it rests on 54 participants and remains a preprint.

The navigation and memory research is the closest long-run analogue we have, and it points consistently one way, but spatial memory is not reasoning and the analogy should carry weight without carrying certainty. Nobody has yet measured what a decade of habitual AI use does to professional judgement, for the straightforward reason that a decade has not elapsed. What can be said with confidence is narrower and harder to dismiss: we get worse at what we stop practising, offloading decisions are frequently misjudged, and confident machine output reliably suppresses scrutiny.

Dependency gets designed in#

Dependency is a design failure, not a character failure, and treating it as a character failure is why most advice on the subject does nothing. Nobody chooses to become dependent. It happens through a sequence of individually sensible decisions, each of which saves twenty minutes, none of which is the moment anything was decided. That is what I mean by drift rather than design. The person who ends up unable to start a document without a prompt did not make that choice; they made two hundred small ones, and this was the residue.

The tell shows up in what is left when the tool is closed, never in how the work looks. Outputs stay good, often for years. That is what makes this difficult to catch, and why capability debt is invisible on every dashboard an organisation keeps. Quality of output stopped being a reliable proxy for capability of the person the moment good output became available to anyone with a subscription.

There is a subtler form worth naming, because it is the one that catches thoughtful people. It means handing over the question, not the work. Asking a model what to do about a decision, a career, a relationship, before you have formed any view of your own, does something more consequential than saving effort, because the framing arrives with the answer and you never see the alternatives you were not offered. I wrote about a small version of this in The 'God Prompt' in December 2024, when a viral prompt promising to reveal your hidden fears was going round: unnervingly accurate, and also generic enough to fit almost anyone. I called it robot astrology rather than insight. The joke has aged into something less funny as people have started routing consequential questions the same way.

The answer is not to use AI less, and I want to be unambiguous about that, because the abstinence framing is both wrong and unhelpful. The answer is to be first. Form the view, then consult. Write the bad draft, then improve it. Set the intent, the framing and the boundaries before the machine generates, rather than editing whatever it produced. That is the principle I call Human at the Start. It is the difference between a person who is amplified by a tool and a person who is steered by one.

The dependency self-check#

Five questions. Answer them about a specific task you did this week, not about yourself in general, because the general answer is always flattering.

• Could I do this unaided, if slower? Not perfectly. Competently. If the answer is no, and it used to be yes, that is the whole finding.
• Would I notice if the output were confidently wrong? The jagged-frontier result is the warning: the people who trusted the machine outside its competence did worse than people with no machine at all.
• Did I have a view before I prompted? If the model's framing is the only framing you have considered, you have outsourced the question rather than the labour.
• Can I say what I changed and why? If you cannot name what you altered in the output, you did not supervise it. You approved it.
• When did I last do this without the tool? If the answer is more than a few months for something central to your value, schedule it. Capability is maintained the way fitness is maintained.

The check does not produce a score. It moves the question from a vague anxiety about using AI too much to a specific, answerable question about one capability you care about keeping.

Write first, then prompt#

Write first, then prompt. Four minutes of your own thinking before the first prompt changes the whole interaction, because you now have a position for the model to attack rather than a vacuum for it to fill. This single habit does more than any other and costs almost nothing.

Keep one unaided rep in the rotation. Choose the capability that carries your value, and exercise it deliberately at intervals, without the tool. Analysts should build a model by hand occasionally. Writers should write something without assistance. Not out of nostalgia, but for the same reason pilots fly manual approaches: so that the capability is there on the day the automation is not.

Treat verification as the work, not the residue. Verification is the judgement layer and it is frequently harder than production. If it is being done in the last ninety seconds before something goes out, it is not being done.

Watch the delegation you make for other people. Automating your own drafting is a personal decision. Automating your team's drafting decides who is capable of the senior role in five years. See how humans learn with AI for what that costs and how to design around it.

Development of the idea#

I first wrote about outsourcing a personal question to a model in The 'God Prompt' (1 December 2024). Are You Flying or Are You Being Flown? (1 March 2026) developed the aviation analogy and the argument about skill atrophy under automation. The Case for Being Bad at Things (18 January 2026) set out the rule that sits underneath the self-check above: feel the difficulty first, form your own answer, then automate deliberately. Human at the Start is developed further in my European Business Review article on accountability gaps in leadership decisions (21 August 2026), and in SuperSkills (Kogan Page, 2026).

Key research and primary sources

• Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Portfolio.
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9).
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. Microsoft Research and Carnegie Mellon, CHI 2025.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. Proceedings of the National Academy of Sciences, 122(26).
• Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. See also the correction published in September 2025.
• Sparrow, B., Liu, J. and Wegner, D. M. (2011). Google Effects on Memory. Science, 333(6043).
• Dahmani, L. and Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory. Scientific Reports, 10, 6310.
• Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt. MIT Media Lab preprint.

Related SuperSkills research#

On the underlying question, see AI and human judgement and AI and critical thinking. On the operating principle, Human at the Start and the Augmented Mindset. On what dependency costs an organisation rather than a person, capability debt, the missed reps and usage theatre. On the same question at the level of a decision process, see human and AI decision making. The mechanism is defined at cognitive offloading. See am I becoming dependent on AI.

About this research#

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 work draws on research across more than 200 organisations in 30 countries over seven years. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is the author's. Cognitive offloading, automation bias and the Google effect are established concepts from the research literature and are not his. Human at the Start, drift versus design, capability debt and the missed reps are part of the SuperSkills lexicon. This is a living reference, reviewed and updated as significant new evidence appears.

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