Using AI without losing your own edge

There's a question that comes up in no AI course: the thing you're delegating right now, could you actually still do it yourself? Not in theory, but now, without a chat window. For most tasks the answer doesn't matter. For the ones your profession depends on, it does. This playbook takes that seriously without trying to talk you out of AI. I use it all day long. The question isn't whether, it's how, and by now there's usable research on that instead of just gut feeling.

One thing up front, for context: some of the studies below are correlational, so they show a connection and not proof of cause and effect. The experimental ones are the genuinely interesting ones, because they show what helps. All sources are at the end. Making that distinction yourself is, incidentally, exactly the ability this is about.

Step 1, understand what's really happening

The technical term is cognitive offloading. You hand a thinking step outwards, the same way you stopped memorising phone numbers once the phone took that over. That's neither new nor bad in itself. Nobody works out square roots in their head, because calculators exist.

What's different about AI is the scope. A calculator takes over one arithmetic step. A language model takes over framing the problem, weighing it up and writing it out, all in one go. Which is to say, exactly the steps that train your judgement.

Gerlich surveyed and tested 666 people across different age groups on this in 2025. The relationship between frequent AI use and critical thinking was clearly negative, r = -0.68, and cognitive offloading explained a large part of it. Younger participants used AI more heavily and did worse on critical thinking. That's a correlation, not proof. But it's strong enough that you shouldn't wave it away.

Step 2, the uncomfortable part, you don't notice it

If you don't play the piano for three months, you notice it the moment you sit down again. With cognitive abilities it works differently, and that's the real catch.

A review paper in Cognitive Research: Principles and Implications describes exactly this mechanism. AI-driven skill loss runs largely unnoticed, because you still do the task. Just not the thinking part of it any more. The example in the paper is a radiologist with AI assistance: he keeps producing successful reports, so he feels sharp. What suffers is telling similar-looking findings apart, judging degrees of severity, spotting the subtle case. He'd have no reason to notice.

Translated to you: your texts still get finished, your quotes still go out, your decisions still get made. The signal that would warn you never appears.

Step 3, decide which mode you're in before you prompt

The most practical lever in the whole playbook, and it costs five seconds. Before you type, decide deliberately: do I want the result, or do I want to be able to do this?

Execution mode is completely legitimate. Writing the text on an invoice, drafting a scheduling email, re-sorting a CSV. Offloading is the entire point of the exercise there, and guilt is out of place.

Understanding mode applies to everything that belongs to your core. Your field, the decisions you have to defend in front of clients, the craft you get paid for. Different rules apply here, and they're in the next few steps.

Most people mix the two without noticing and slide further into execution mode over months. Just naming which one you're in stops a large part of that.

Step 4, flip the order around

In understanding mode you don't ask first. You write down your own answer first, at least as bullet points, at least three sentences, and only after that do you ask the AI.

It feels like a waste of time and it's the most effective trick there is. First, you've then actually thought. Second, you see the difference between your answer and its answer, and that difference is where the real learning sits. Third, you notice immediately when the AI misses something you had. That happens more often than you'd think.

In practice: a two-minute timer, a notepad or an empty document, and only then the chat window.

Step 5, ask with structure instead of just asking

This is where it gets concrete, because this is exactly what was tested experimentally. In a study with 150 participants from Germany, Switzerland and the UK, 450 answers were rated by independent assessors across four conditions: without AI, with AI and no guidance, with AI and structured guidance, and AI on its own.

The result is uncomfortable and helpful at the same time. Unguided AI use led to offloading without improving the quality of the reasoning. So you hand over the thinking and get nothing better in return. Structured prompting, by contrast, cut offloading noticeably and improved both the quality of the argument and the mental engagement.

Structured here means treating the AI as a research tool, not as a task solver. So not "write me an analysis of X", but "give me the three strongest counterarguments to my thesis, with sources, and tell me where my reasoning is weakest". You keep the conclusion, it supplies the material.

Step 6, ask for options, not answers

There are three typical illusions in dealing with AI, described in the literature as illusions of understanding. You believe you've understood something more deeply than you have. You believe you've considered every possibility, when in fact only the ones the model showed you. And you take the output to be objective, even though training data and vendor decisions are baked into it.

A simple habit works against the second one: ask for options with their downsides, never for the best solution. "Give me three approaches, each with its biggest drawback, and tell me which one fails under which circumstances." The choice then stays with you, and you see the field instead of a slice of it.

Against the third one, it helps to put the same question to a second model now and then. Where two vendors diverge, there's rarely a fact and usually a judgement call.

Step 7, do the explain-it-back test

A test you can run in two minutes, and it's more honest than any self-image. Take something you got done with AI this week and explain it to someone. Out loud, without notes, without looking anything up.

If you get through it, it was execution and everything is fine. If you get stuck at the point where it gets interesting, you used the result without owning it. That's no disgrace, it's information.

Do this once a week with exactly one thing. Not with all of them, nobody keeps that up.

Step 8, keep a handful of tasks AI-free

Musicians call this the practice piece. You need a few activities that you deliberately keep doing without help, so the ability stays with you.

Pick two or three from your core area. Not the biggest ones and not the most urgent ones, but the typical ones. Writing one quote a month entirely yourself. One round of troubleshooting without asking. One client conversation without a prepared script.

That costs you maybe three hours a month and it's the cheapest insurance policy you can take out on your own employability.

Step 9, "that felt easy" tells you nothing at all

This is the finding that surprised me the most, and the most valuable one in practice. In the same study with the 450 rated answers, there was no relationship whatsoever between cognitive offloading and how difficult the task felt, r = 0.0078 with p = 0.87.

In plain terms: how easy something felt tells you nothing about whether you were thinking while you did it. Your feeling is not a measuring instrument at this point. Participants were convinced they had done the reasoning themselves, while they had handed over the bulk of the thinking.

That's why resolutions like "I'll pay attention" don't work. You need external touchstones, and those are steps 4, 7 and 8.

Step 10, build it into the calendar, not into your good intentions

Everything above falls apart within two weeks if it depends on self-discipline. So turn it into a small fixed routine.

One appointment a week, fifteen minutes, always the same slot. Three things inside it: the explain-it-back test from step 7 on one thing. A quick look at whether you kept to the order from step 4 this week when you were in understanding mode. And the question of whether your AI-free tasks from step 8 actually happened or quietly fell asleep.

Fifteen minutes a week. That's the whole price. If you work in a team, put the point into your AI policy, then it isn't something everybody settles alone with themselves.

What's next

If you want to go deeper on checking AI claims, the playbook on fact-checking AI answers is the logical next step. If you want to understand why models agree with you so readily, you'll find that in Level 4 under sycophancy and bias, and the basics on invented facts are in Level 1 under hallucinations. If you're in the middle of learning a new field, read this together with the prompting lessons in Level 2, because the structured asking from step 5 is the tool of the trade there.

Sources

• Gerlich, M. (2025): AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1), 6. 666 participants, r = -0.68. https://www.mdpi.com/2075-4698/15/1/6
• From Offloading to Engagement: An Experimental Study on Structured Prompting and Critical Reasoning with Generative AI. Data 10(11), 172. n = 150 from Germany, Switzerland and the UK, 450 rated answers. https://www.mdpi.com/2306-5729/10/11/172
• Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers' awareness? Cognitive Research: Principles and Implications (2024). https://link.springer.com/article/10.1186/s41235-024-00572-8
• Georgiou, G. P.: ChatGPT produces more "lazy" thinkers: Evidence of cognitive engagement decline. arXiv:2507.00181. https://arxiv.org/pdf/2507.00181
• Learners' AI dependence and critical thinking: The psychological mechanism of fatigue and the social buffering role of AI literacy. 580 students. https://www.sciencedirect.com/science/article/pii/S0001691825010388