by Claude Opus 5.5
What are the common ways individuals accidentally make themselves more replaceable when using AI, and what habits prevent deskilling?
People rarely choose to become replaceable. They drift into it by handing the tool the judgement and keeping the clerical work, by skipping the practice that builds expertise, and by producing work that looks like what anyone else with the same tool would produce. The protection is a handful of deliberate habits: think before you prompt, keep doing some work unaided, check systematically, and make your own judgement visible.
The evidence that deskilling is real
This is no longer just a theoretical worry.
A measured drop in clinical skill. A multicentre study in The Lancet Gastroenterology & Hepatology (Budzyń et al., 2025) looked at endoscopists who had been using AI to help spot polyps. When they performed colonoscopies without the AI, their adenoma detection rate fell from 28.4% before AI was introduced to 22.4% afterwards. The study is observational, so it cannot prove cause, but the pattern is what deskilling would look like.
Less critical thinking among confident users. A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers (CHI 2025) found that higher confidence in generative AI was associated with less critical thinking. Higher self-confidence was associated with more.
An old warning. Lisanne Bainbridge’s 1983 paper “Ironies of Automation” warned that when people become monitors of automated systems, they lose the hands-on skill needed to step in when the system fails, which is exactly when they are needed most.
There are labour-market implications too. Brynjolfsson, Li and Raymond’s study of 5,179 customer support agents found that an AI assistant raised productivity by 14% on average and by 34% for novices, with little gain for the most experienced. That is good news for novices, but it means that being merely competent is worth less, because the tool supplies it. A UK signal points the same way. The ISE’s 2026 development survey found 29% of employers reporting rising performance problems among new hires, up from 12% in 2022. That cannot be pinned on AI, but it is the trend to watch.
Seven ways people make themselves more replaceable
1. Becoming the last-mile formatter. If your contribution is tidying AI drafts, you are doing the part of the job that is easiest to remove next. Your work starts to look clean, plausible and interchangeable with anyone else’s.
2. Letting the tool frame the problem. Asking for “a strategy for X” or “an analysis of Y” accepts the model’s framing and its generic best practice. The valuable part of most knowledge work is deciding what the question is and which constraints matter here.
3. Skipping the reps. This is the biggest risk for people early in their careers. Juniors learn judgement by doing routine work many times. A trainee who never drafts a disclosure list, reconciles an account or writes a function from scratch cannot later tell when the AI version is wrong.
4. Rubber-stamping. Automation bias, the tendency to accept a system’s suggestion even when it is wrong, is long documented. Approving output you have not really checked turns you into the human-in-the-loop that organisations want to remove. It also leaves you exposed when an error surfaces. In Ayinde v London Borough of Haringey (June 2025), the High Court dealt with fictitious authorities put before it and warned lawyers of serious professional consequences.
5. Hiding your use. Deloitte’s 2026 UK survey found that 31% of generative AI users at work use it without their employer knowing. Hidden use means you get no credit for the skill you are building. It also means you are not part of the conversation about how the workflow should change, and you carry the risk if something leaks.
6. Making your contribution invisible. If your manager sees only polished output, they may attribute it to the tool. When the next restructure asks what each role adds, you need an answer.
7. Writing yourself out without moving up. Documenting your process so it can be automated is often the right thing to do. But if you write the playbook and do not take ownership of the system that runs it, you have done your successor’s job for them.
Habits that prevent deskilling
Think first, then prompt. Sketch your outline, answer or hypothesis before opening the tool, even roughly. Then use AI to expand, challenge or check it. That keeps the framing and first-principles reasoning, which are the most valuable parts, with you.
Use AI as a sparring partner, not a ghostwriter. Some of the most useful prompts are adversarial: “What would a sceptical finance director attack here?”, “What are the five likeliest errors in this reconciliation?”, “Argue the other side.”
Keep a quota of unaided work. Pick the one or two skills that form the core of your profession, such as drafting, debugging, statistical reasoning or diagnosis, and do some of that work regularly without AI. For example, write one memo a week unaided, or debug one issue without the assistant. Pilots still practise manual landings for the same reason.
Apply the explain-it test. Before you send anything AI-assisted, ask whether you could defend it line by line in a meeting without the tool. If not, you do not yet own it.
Check with a method. Use a short checklist for your domain covering sources traced, figures reconciled, edge cases tested, and policy and data protection checked. Keep an error log of what the tool gets wrong in your work. Within a few months it becomes a valuable asset, and good portfolio evidence (see question 4.4).
Protect the inputs only you can get. Customer conversations, site visits, case files, ward rounds and the institutional memory of why something failed last time are the raw material AI lacks. If AI saves you time, spend some of it there.
Make your judgement visible. Note where you overrode or corrected the tool, both in your work and in your one-to-ones. Share the error log with your team. Volunteer for the exception queue rather than the routine flow.
For managers of juniors: protect learning tasks. Assign some work to be done without AI first, review the reasoning and not just the output, and pair juniors on exceptions. A team that learns only to supervise AI will have nobody who can do the work when the system fails.
A quick self-check
If most of your AI use is “generate, then lightly edit”, you are drifting towards replaceability. If it is “frame, challenge, test, then decide”, you are compounding skill. The tools are the same in both cases. What differs is who is doing the thinking.
Bottom line
Use AI to go faster at work you understand, not to stand in for understanding. Keep some unaided practice, check output with a method, and make sure the people deciding your future can see the judgement you add.