The AI Dependency Test: What If ChatGPT Disappeared Tomorrow? 25 Questions That Reveal How Much You Rely on AI

A rough way to tell them apart:

AI assistance: you think → AI helps → you evaluate → you own the result.

AI dependence: AI thinks → AI produces → you mostly approve.

No one is always on one side or the other. You might feel completely in control when writing a strategy memo, but lean on AI for that tricky SQL query you don’t quite get. Which mode you’re in depends on the task, the deadline, and your confidence in that area.

There’s nothing wrong with letting AI handle the boring stuff. Reformatting a CSV, generating boilerplate, turning a rough transcript into neat notes, or drafting yet another standard client update—these are chores. Letting AI take over gives you time back for the work that really needs your attention. If someone calls that a moral failing, they’ve probably never done truly repetitive work.

But it’s a different story when you start letting go of the work that really matters: making your own arguments, truly understanding what you just read, pushing through something new on your own, writing in your own voice, deciding if a claim is solid, recalling something you learned last month, or making a call without asking a model to double-check it for you. These aren’t chores—these are the load-bearing walls of your work. If you take one away, everything might look fine for a while, which is exactly why it’s so easy to miss when something important is quietly slipping away.

INFOGRAPHIC 1 — AI Assistance vs AI Dependence

Answer yes or no. Answer for how you worked last week, not how you intend to work. Score 1 point for every yes.

A. Writing and communication

1. Would writing a professional email to someone senior feel a lot tougher without AI helping out?
2. Do you usually turn to AI to kick off a piece of writing, rather than just asking it to polish what you already wrote?
3. When you write something on your own, do you find yourself double-checking your grammar or word choices more than you used to?
4. Has it been over a month since you wrote something substantial, from start to finish, without any AI help?
5. Would you pause before sending an important message if AI hadn’t reviewed it first?

B. Thinking and decision-making

6. Do you open an AI tool before you’ve spent any time thinking about the problem yourself?
7. When brainstorming, do you generate fewer than three of your own ideas before asking for more?
8. Have you asked AI to confirm a decision you had already made?
9. When comparing two options, do you rely on AI’s comparison rather than building your own?
10. If AI disagreed with your judgment on something in your area of expertise, would you assume AI was right?

C. Learning and research

11. Has AI replaced search engines as your default way of finding things out?
12. Do you read AI summaries of documents more often than the documents?
13. Would you need to reopen an AI conversation to explain something important that you learned from it last week?
14. When AI gives you a fact or a number, do you usually accept it without checking a primary source?
15. In the past month, have you finished a research task without opening a single original source?

D. Work and productivity

16. Would building a presentation from a blank slide take you significantly longer now than it did two years ago?
17. Do you ship code, formulas, or configurations without being able to explain the logic or verify that they work?
18. Do you rely on AI meeting summaries instead of your own notes or memory?
19. Would your weekly or project planning feel harder to structure without AI?
20. Is there documentation in your work that exists only because AI wrote it, and that you haven’t fully read?

E. Creativity and independence

21. Does a blank page feel worse to you now than it did before you had AI tools?
22. Do you ask AI first on creative tasks rather than after an attempt?
23. Have you stopped practicing a skill because AI does it faster?
24. Has it been more than a month since you created something substantial without involving AI?
25. Would you struggle to explain or defend an important piece of AI-assisted work if someone questioned how it was produced?

INFOGRAPHIC 2 — The 25-Question AI Dependency Test
5 categories. 5 questions each. 1 point per yes. Writing · Thinking · Learning · Work · Creativity

0 to 5. AI appears to be mostly an occasional assistant in your current workflow.

6 to 10. AI is integrated into your regular workflow. You may notice a bigger loss of speed than of underlying capability.

11 to 15. AI handles a meaningful portion of your workflow. Some tasks may take noticeably more effort without it.

16 to 20. Several abilities may now be heavily AI-assisted. A week without these tools would likely feel very different.

21 to 25. AI appears central to how you currently work. That is worth examining, not panicking about.

The score is a conversation starter. It isn’t a medical result, a psychological profile, or a productivity grade. The score matters less than which questions gave you a point—those are the tasks worth examining.

Forget the score for a moment and ask a harder question: what could you still do if the tools vanished?

A developer who uses AI autocomplete constantly but understands the system architecture, can read the diff, and can debug production at 2 a.m. is in a fundamentally different position from one who can’t explain why the code works. Both use AI heavily. Only one has a problem.

A writer who uses AI to gather background but controls the argument, the structure, and the voice hasn’t given anything away. A writer who can no longer decide what the piece is about has.

A business owner who uses AI to model three pricing scenarios and then applies judgment about their market is using a calculator. One who adopts whichever option the model ranked first has delegated the decision, not the arithmetic.

The real question is: can you explain it, check it, redo it, or change it yourself? If AI were gone, you don’t have to be just as quick—you need to understand your work well enough to keep going.

Skills don’t wave goodbye when they leave. You stop using them, and then one day you reach for a skill and realize it’s gone.

If you’re over 30, you’ve probably already felt smaller versions of this shift in your own life. Think about how calculators changed the need for mental math, or how GPS made it easier to get around but harder to keep your sense of direction sharp. Autocorrect made spelling less of a worry, and contact lists meant you stopped memorizing phone numbers. These changes didn’t feel like losses at the time—they just faded quietly into the background, only becoming noticeable when you suddenly needed to split a bill or find an address without your phone.

The research on this topic is real, but it’s more nuanced than the headlines suggest. Take, for example, a McGill study of 50 regular drivers: those who used GPS more often performed worse on spatial memory tasks when GPS wasn’t available. In a smaller follow-up with 13 people over three years, heavier GPS users showed a steeper decline in the kind of memory that helps you navigate. Interestingly, the people who relied most on GPS didn’t start with a worse sense of direction—they just used the tool more. But keep in mind, these are small studies. The findings are interesting, but not the final word.

One of the most famous studies in this space was about the "Google effect" on memory. The idea: we remember where to find information, not the information itself. But when a 2018 project tried to replicate 21 big social science findings—this one included—it couldn’t reproduce the results. The evidence here is messy and still up for debate.

So, what about AI? The same careful thinking applies. There’s a reasonable theory that “cognitive offloading”—letting technology handle parts of our thinking—now stretches into writing, reasoning, recall, research, and even generating ideas. But the proof that this leads to any lasting decline for adults is thin so far. If someone tells you AI is ruining your brain, they’re jumping ahead of the evidence.

Still, a few findings are worth knowing (and each comes with caveats):

• A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers, looking at 936 real examples of AI-assisted tasks, found that when people had more confidence in AI, they did less critical thinking. But when they had more confidence in themselves, they engaged in more critical thinking. Keep in mind, this is self-reported and correlational—it’s about how hard people felt they were thinking, not actual ability.

• A study in the journal Societies surveyed 666 people and found that frequent AI users tended to score lower on critical thinking, especially younger participants. This effect was tied to relying on AI for thinking (“cognitive offloading”). The sample was self-selected, and the paper issued a correction later. Again, it’s correlation, not causation.

• The strongest evidence so far comes from a randomized experiment in Turkey. Nearly 1,000 high school students got either a standard chatbot tutor, a tutor with built-in learning safeguards, or no AI help. Students with AI tutors did much better on practice problems—48% and 127% better than those without. But on an unassisted test, the group that used the standard chatbot scored 17% worse than those who never had AI. The safeguarded version largely mitigated that negative learning effect. This was a real experiment with a causal design—but in one subject, at one school.

There’s also the widely cited MIT Media Lab EEG study, where 54 people wrote essays using an AI language model, a search engine, or nothing at all. The AI group showed weaker neural connectivity and poorer recall. In the first session, the study reported that 83% of participants in the LLM group were unable to quote from essays they had just written. But this study is still a preprint, and even the authors say the results are early and should be treated as a hypothesis, not a settled fact. Others have raised concerns about the methods. In short, it’s worth paying attention to, but not something to base big decisions on yet.

Here’s what makes this hard to self-diagnose. Real productivity gains and quiet skill loss look identical from the inside.

Someone producing five reports where they used to produce two has genuinely improved output. That’s not an illusion. But if they can no longer structure a report from scratch, two things changed, and only one was visible.

People are also poor judges of their own AI-assisted performance. In a randomized trial of experienced open-source developers working in their own repositories, METR found that allowing AI increased task completion time by 19%, while the developers estimated afterward that it had reduced their time by 20%. That was 16 developers, 246 tasks, using tools available in early 2025, and METR now treats the result as historical rather than current. Their follow-up in February 2026 found evidence of speedup instead, but with the caveat that many developers refused to participate in no-AI conditions at all, which biased the sample.

That last detail matters. A measurable number of professional developers would not take paid work under the condition of not using AI, showing that AI had become important enough to affect participation. It also introduced selection bias into the follow-up, which is why METR cautioned against treating the newer speed estimate as definitive.

This is what it looks like on the ground: a writer who can whip up ten drafts with AI but feels stuck finishing one without it—a student who gets good grades on homework but can’t quite keep up on tests. A marketer juggling more campaigns than ever, but starting to lose the feel for what actually clicks with people. A developer pushing out code at record speed, yet barely glancing at their own work. A manager whose emails sound sharper, but who avoids really digging into tough people issues—or a founder with beautiful presentations and no gut sense of what the numbers mean.

None of that makes the productivity fake. Both things are true at once.

The distinction that matters isn’t whether you use AI. It’s whether you use it to understand something or to finish something.

Useful: asking for a second explanation when the first one didn’t land. Generating practice problems. Attempting an answer and asking AI to critique it. Comparing two approaches to the same problem and asking why your reasoning failed, after you’ve reasoned.

Here’s where it gets risky: letting AI do the assignment before you’ve tried it yourself. Skimming an AI summary instead of reading the material. Turning in an explanation you couldn’t actually walk through on a whiteboard. Or asking AI for help the second things get tough—usually the exact moment real learning kicks in.

The Turkish high school experiment stands out as some of the clearest evidence we have. But one of its most useful findings often gets overlooked: when the AI tutor focused on guiding students—offering hints and nudges instead of simply giving answers—the negative learning effect was largely mitigated. The tool mattered, but so did the interaction pattern.

INFOGRAPHIC 3 — Student AI: Learning vs Completing

Businesses have long had plans for what to do if something big goes wrong—like a power outage or a key supplier failing. But hardly anyone has made a plan for what happens if an AI tool goes down, even though AI has quickly become essential to how things get done, often before anyone had time to write down the new process.

It helps to think about AI tools the way you’d think about any other critical vendor. What if the service goes offline for three days? What if the price suddenly triples at renewal? What if legal says you can’t use a certain AI model for sensitive data? What if a client contract says their data can’t touch a public AI service, but your team’s whole workflow depends on it?

Recent surveys show just how common AI use has become at work. According to Gallup (February 2026), half of U.S. employees use AI on the job at least a few times a year, with 28% using it weekly and 13% every day. Pew’s 2025 survey, which asked the question a bit differently, put the number at 21%. The surveys don’t match up because "using AI" can mean lots of different things—and neither really measures whether people or organizations could keep things running without those tools.

So what can you actually do to be ready? Here are some steps (and none are about cybersecurity): Write down your most important processes somewhere outside chat windows. Have at least one person who fully understands each key workflow from start to finish. Don’t let critical know-how get stuck in AI conversations that nobody ever saves. Always have a human review anything that has real consequences. Train people to double-check, not just approve. Actively protect your expertise, because it won’t stick around on its own.

It’s not about using less AI. It’s about making sure that if your AI tool goes down, your whole operation doesn’t go down with it.

Pick one ordinary working or study day: no ChatGPT, Claude, Gemini, Copilot, or Perplexity. No AI autocomplete, no AI writing assistants, no AI-generated summaries. Switching to a different model doesn’t count.

While you’re spending a day without AI, keep a little journal. Ask yourself:

1. What did you notice slowing down right away?
2. Was anything surprisingly difficult?
3. Did any skills start to come back after an hour or so?
4. Was there something you found you actually liked doing yourself?
5. Which AI-powered shortcuts felt genuinely valuable?
6. Did you spot any tasks you’d been outsourcing out of habit, not necessity?

The goal is not to prove you don’t need AI. You probably do, and that’s fine. The goal is to find out precisely where AI is helping you and where it has quietly become a requirement. Those are different categories, and most people have never separated them.

Expect the results to surprise you in both directions. Some things will be harder than you assumed. Others will come back immediately and feel good.

INFOGRAPHIC 4 — Could You Work One Day Without AI?

Rate each: Easy / Slower / Difficult

Email · Writing · Research · Brainstorming · Coding · Presentations · Learning something new · Planning your week

Attempt first

Spend five minutes on the problem before opening a chat window. Five minutes is enough to form an approach, which changes what you ask for and what you notice in the answer.

Ask for critique, not creation

Write the draft, then ask AI to challenge it. You’ll get better output, and you’ll keep the part of the work that was actually yours.

Explain it back

After learning something from AI, close the tab and explain it out loud or in writing. If you can’t, you received an answer rather than an understanding.

Verify what matters

For anything consequential, check a primary source. A confident-sounding AI answer isn’t evidence that it’s correct.

Keep one manual repetition

Occasionally do the thing without help. Write one report from scratch each quarter. One function written without autocomplete. It costs an hour and tells you where you stand.

Ask AI to teach, not to answer

Ask for hints, opposing viewpoints, thoughtful questions, or feedback on your work. In the high school study, the safeguarded AI tutor used this approach and largely mitigated the negative learning effect seen with the unrestricted chatbot.

Keep the judgment

An AI recommendation is an input. Treat the moment where it becomes a decision as a separate step that belongs to you.

INFOGRAPHIC 5 — 7 Ways to Keep Your Skills While Using AI
Attempt first · Critique, don’t create · Explain it back · Verify what matters · One manual repetition · Teach, don’t answer · Keep the judgment.

The best AI users aren’t the ones who automate everything—they’re the ones who know what to hand off, what to do themselves, and when to shut the laptop and just think. That kind of judgment is a real human skill, and it’s one you still have to exercise yourself.

Here’s the one question that truly matters:

If all your AI tools vanished tomorrow, what’s the one skill you’d most regret not keeping sharp?

Be honest with yourself about the answer—and you’ll learn more than any score could tell you.

1. “Speed is not the same as capability. AI can give you one without the other.”
2. “You don’t notice a skill leaving. You notice it missing.”
3. “AI should reduce the cost of thinking, not the amount of it.”
4. “The question isn’t how much AI you use. It’s what still works when it’s gone.”
5. “Organizations plan for vendor outages. Almost none plan for capability outages.”