New research is showing something uncomfortable: the more we lean on AI at work, the worse we get at the actual job underneath it. Doctors, coders, and accountants are all showing the same pattern, and there is a real fix for it.
Here’s the thing nobody wants to admit about AI at work
You use an AI tool because it makes you faster and sharper. That part is true. But a string of new studies, covered in a recent Nature investigation, is showing the flip side nobody talks about enough: the skill you stop practicing is the skill you start losing.
This is not a vague fear about robots taking jobs. It is measured, in colonoscopy rooms and coding tests, with hard numbers attached. And it should change how you use AI starting today, not someday.
The doctors who got worse without the AI watching
Start with the most concrete example. A group of physicians in Poland who specialize in colonoscopies, each with at least 2,000 procedures of hands-on experience, were given an AI tool that flags precancerous lesions called adenomas in real time. The tool was switched on for some sessions and off for others, which let researchers compare the same doctors’ performance with and without the assistant.
Before the AI tool showed up, these specialists caught at least one adenoma in 28.4% of their colonoscopies. After they had gotten used to working with the AI, their detection rate on AI-free days dropped to 22.4%.
Read that again. These were not novices. These were doctors who had performed thousands of the procedure before AI ever entered the room. A few months of AI assistance was enough to measurably dull a skill they had spent years building, the moment the assistance disappeared.
The study, published in The Lancet Gastroenterology and Hepatology, put it bluntly: constant exposure to the tool left clinicians “less motivated, less focused, and less responsible” when they had to make the same calls on their own.
What happened when coders used AI to learn a new task
The second study comes from Anthropic, and it is just as telling. Researchers ran a controlled test with 52 software engineers performing a coding task. Everyone had access to the internet and standard documentation. Half the group also got an AI coding assistant.
After the task, everyone took a quiz on what they had just done. The group that used AI assistance scored noticeably lower, the equivalent of nearly two letter grades worse than the group that coded by hand. They struggled most on questions asking them to diagnose errors in code, the exact kind of question that proves whether you actually understand what you built.
What this really means is that AI can let you finish a task without ever learning how the task works. Kevin Crowston, an information scientist at Syracuse University, frames it as a strange new disconnect: you can perform at a high level because you are borrowing the AI’s skill, while building none of your own.
This is not new, but AI is different in one big way
Skill erosion from automation is not a new story. GPS dulled people’s sense of direction. An earlier study of accountants who had used a non-AI automated system for over a decade found that when the tool was removed, they had simply forgotten how to do routine tasks by hand.
What makes generative AI different, according to Tapani Rinta-Kahila, an information systems researcher at the University of Queensland, is that it is the first technology to automate the thinking part of work, not just the doing part. GPS replaced your sense of direction. AI is starting to replace your judgment, your error-checking instinct, and your ability to explain your own reasoning.
That is a much bigger thing to outsource by accident.
The part of this story that should actually worry you
Here is the detail that matters most for anyone reading this on a laptop right now, not in a hospital or research lab.
A separate 2026 survey of US healthcare workers found that 70% of nurses and 77% of physicians are already worried about losing their own skills to AI over-reliance. These are people who use AI tools daily and are still raising the alarm on themselves. If trained medical professionals feel the pull strongly enough to flag it, the rest of us should assume we are at least as exposed, probably more, since most jobs outside medicine have far less structured oversight of how AI gets used.
How to use AI without quietly losing your edge: a practical plan
None of this means quit using AI. That would be like giving up GPS and navigating by stars out of principle. The researchers are not telling people to stop. They are telling people to use AI on purpose, not on autopilot. Here is how to actually do that.
1. Pick your “no AI” skills before you need them
Decide in advance which one or two core skills in your job you refuse to fully outsource. For a doctor, that might be the diagnostic reasoning behind a scan. For a writer, it might be structuring an argument before any drafting tool touches it. For a developer, it might be reading and understanding error messages without pasting them into a chatbot first. Pick the skill that matters most if the tool ever goes down, and protect it deliberately.
2. Do the first attempt yourself, then bring in AI
Instead of opening an AI tool first, spend five or ten minutes attempting the task cold. Write the first draft of the email, sketch the first version of the code, form your own diagnosis or analysis before checking it against the tool. This single habit preserves the “struggle” that is actually where learning happens. The AI becomes a check on your thinking instead of a replacement for it.
3. Ask the tool to explain, not just to answer
When AI gives you an answer, ask it to walk through why. “Why did you flag this as a risk?” or “Explain the logic behind this code” turns a passive output into an active lesson. This is the same fix the Anthropic researchers point toward: the problem was not using AI, it was using it without engaging with the reasoning behind the result.
4. Run periodic “unplugged” reps
Set a recurring reminder, weekly or monthly, to do a real task entirely without AI assistance. Treat it like a fire drill for your own skills. If you notice the task feels harder or slower than it used to, that is useful information, not a failure. It tells you exactly where deskilling has already started, while you still have time to correct it.
5. Get feedback on your unassisted work, not just your AI-assisted output
If you work in a team, ask a colleague or manager to occasionally review work you did without AI help. This keeps an honest signal in the loop about your standalone ability, separate from how good the AI made your output look. Self-assessment alone tends to be unreliable here, because AI-assisted output can look polished even when the underlying understanding is thin.
The honest limitation in all of this
These studies are early, and the researchers say so themselves. The colonoscopy study covers one procedure in one country. The coding study had 52 participants. Yuichi Mori, a physician-researcher at the University of Oslo and co-author of the colonoscopy study, says plainly that more research is needed to confirm how widespread and how permanent this effect really is. Nobody has a proven fix yet either. As Mori puts it, deskilling does not have an established solution right now, and figuring one out is likely to be a major research focus over the next decade.
So treat this as an early warning, not a verdict. The direction of the evidence is consistent across two very different fields, medicine and software engineering, which is exactly why it is worth taking seriously now rather than waiting for a bigger study to catch up.
What this means for you, starting today
The goal is not to use AI less. It is to stay the kind of person who could still do the job well if the AI disappeared tomorrow. That is a genuinely useful test to run on yourself: if your main AI tool went offline for a week, would your work fall apart, or would it just get a little slower?
If the honest answer is “fall apart,” you already know where to start.
If this changed how you think about your own AI habits, share it with a colleague who lives inside ChatGPT or Copilot all day. They will probably recognize themselves in it.
Sources and further reading
- Is AI ruining our skills? Early results are in, and they’re not good, Mariana Lenharo, Nature (June 2026).
- Budzyń, K. et al., Lancet Gastroenterology and Hepatology, Vol 10, 896-903 (2025).
- How AI assistance impacts the formation of coding skills, Shen, J.H. and Tamkin, A., Anthropic (2026).
- 2026 Future Ready Healthcare Survey Report, Wolters Kluwer (2026).
- Consequences of Discontinuing Knowledge Work Automation, Rinta-Kahila, T. et al., HICSS (2018).
Just Being Resourceful! 
Leave a comment