Protect Learning when Adopting AI.

In 1983, Lisanne Bainbridge published a short paper called Ironies of Automation. It was about the people who watched automated industrial processes, operators who had learned the work by doing it by hand before the machines took over.

Bainbridge saw two problems coming. An experienced operator who spends years monitoring automation “may now be an inexperienced one.” And the systems were “riding on their skills, which later generations of operators cannot be expected to have.”

Forty-three years later, I sat in a conference room in Austin at the UT REAL Health AI Symposium and heard speakers describe the same irony with new words.

The room was talking mostly about clinicians. The irony reaches everyone who works next to AI: the analyst whose first queries come from a model, the coder reviewing suggested codes, the scheduler, the manager drafting a performance review, the nurse, and the physician.

How skills change with AI

Speakers used three terms that have been spreading through medical education: de-skilling, never-skilling, and mis-skilling. They’re useful well beyond medicine. They’re also some of the ways skills slip, and you’ll hear other names for related patterns, like automation bias, automation complacency, skill fade, and cognitive offloading.

I also think re-skilling belongs in this conversation. It’s the deliberate work of building skills for changed responsibilities as AI takes on more of a job.

The examples in the right column are illustrations, and the list isn’t complete. When a task moves to AI, ask what people still need to practice and what they need to learn next.

What the evidence says so far

The evidence is early, and it comes from different corners. Taken together, it points one way.

Knowledge workers. Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers about how they use generative AI and found that “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.” It’s a self-reported survey, so it shows an association, and it’s the same pattern the rest of this evidence points to.

People still learning. In a field experiment with nearly a thousand high school math students, students with unrestricted GPT-4 access did better on practice problems. When the access was taken away, they scored 17% worse than students who never had it. A version of the tutor that gave hints instead of answers largely removed the harm. That’s never-skilling in a classroom, and it also shows design matters.

Experienced experts. The clinical example everyone cites comes from four endoscopy centers in Poland. After the centers introduced AI polyp detection, the detection rate in standard, non-AI colonoscopies fell from 28.4% to 22.4%. The authors wrote that AI exposure “might reduce” unaided detection, and the study was observational. It’s one study, and it’s a signal.

A keynote speaker in Austin named the human mechanism underneath all three. After a tool has been right ten or twenty times in a row, people stop checking the next one carefully. It holds in a coding queue as much as in an exam room.

What helps people keep their skills

The research on what helps is more encouraging, and it points at specific habits.

In a Harvard study of 199 people working with AI recommendations, Buçinca and colleagues tested “cognitive forcing” designs, such as asking people to make their own decision before the AI’s suggestion appears. They found that “cognitive forcing significantly reduced overreliance compared to the simple explainable AI approaches.” There was a catch: people rated the designs that helped most as the least pleasant to use. People don’t enjoy the step that protects them, which is exactly why it has to be taught and practiced.

Older research on automation points the same direction. In a 2008 process-control study, Bahner and colleagues trained people to work with an automated decision aid. Operators who experienced the aid’s failures during training checked its recommendations more carefully than operators who had only been told failures could happen. The authors concluded that “exposing operators to automation failures during training significantly decreased complacency” and called it a suitable way to reduce the risk.

The math experiment points the same way from the learning side. The tutor that gave hints and asked questions kept students learning, and the one that handed over answers replaced the learning.

Another speaker raised the bigger version of the problem, and it’s the one I keep coming back to. Every profession has built judgment through repetitive entry-level work. In analytics, the reports a new analyst writes by hand in that first year are where they learn how the data fits together. Coders, schedulers, nurses, and residents all have their own version of that year. If AI absorbs the entry-level work, where does expertise get built? The answer has to be deliberate: training that explains the risk, and practice built into the job that keeps people thinking first.

Re-skilling as the work changes

Re-skilling means learning new skills for a changed role or a different kind of work. An AI rollout can also call for up-skilling, where people deepen skills they already use.

Take an analyst whose weekly report now starts with an AI draft. I’d want them to keep enough query-writing practice to check the output and learn how to test the new reporting workflow before other teams depend on it. A scheduler whose routine bookings are automated may need practice handling unusual cases. These are illustrations of how responsibilities could shift.

AI can help with the learning too. The math experiment gives us a useful design lesson: a tutor that offers hints can keep people working through the problem. I’d look for that kind of support in workforce training, with a qualified person checking the material.

Plus, time saved on routine work can become time for supervised practice. That takes a decision from the organization about where the time goes.

AI literacy training is step one

Learning to drive is a good model. You take a classroom course, log supervised hours behind the wheel, and then take a road test. The course gives you the rules and a shared language, and the hours behind the wheel are what make you a driver.

AI literacy belongs in the same sequence, and accreditors are starting to ask for the first step. The Joint Commission’s new Responsible Use of AI in Healthcare certification asks reviewers to verify that training includes “AI literacy and change management”, is “tailored to staff roles,” and recurs at least every two years. The users reviewers sample include “nurses, physicians, coders, schedulers, case managers.” The certification already treats this as an everybody problem.

Here’s how I’d sequence the work for an organization:

1. A foundation for everyone. What the tools do, where they fail, what automation bias is, how your AI policy works, and who to call when output looks wrong. This is the shared language, and every later step depends on it.

2. Practice that builds the reflex. Role-specific exercises where people review AI output seeded with errors, and commit to their own answer before they look at the AI’s. Deciding first is the habit the Harvard study found protective, seeing the tool fail is what the process-control study found protective, and both are part of the Ground step of The SIGN Method. To see what this practice feels like, try Sign or Send Back, a two-minute exercise we built: five AI-drafted notes, some with statements nobody recorded, and your job is to catch them before you sign. The notes happen to be clinical. The same reflex applies to an AI-drafted report, contract summary, or email.

3. Measurement. For the tools and roles where it matters most, track performance with AI and without it, and report both numbers to your AI governance committee.

For re-skilling, give people supervised practice in their new responsibilities and check whether they can handle a realistic task from that work.

Practice in step two only works once people share the vocabulary from step one, and step three is how you find out whether either one held. It builds on the AI literacy foundation and the education program design I’ve written about before.

Protecting how new people learn

New hires and trainees need the most protection, because they don’t have the old skills to fall back on. Presenters at the conference proposed a progression for medical learners, and it translates to any role:

Foundations without AI until competence shows up. A new analyst writes the first round of queries by hand. A new coder codes a sample of charts unassisted before suggestions turn on. People keep the answers they work out better than the ones they’re handed. A faculty panelist described doing this in their own courses: explaining up front why AI is off-limits for one part of an assignment, then bringing it in and grading how well the student evaluates its output.

Guided critique of AI output. The presenters called it adversarial pedagogy. Learners get AI output seeded with errors and have to find them, which is the same idea the process-control study supports: people who have seen the tool fail check it more carefully.

Integrated practice with graduated trust. People use AI in real work under supervision, and managers extend trust as competence shows up. The math experiment points to how: tools that give hints and ask questions teach, and tools that hand over answers can replace the learning.

Making it work across your organization

Most organizations already have AI switched on somewhere. Four places to start.

Decide which roles get AI on day one. For a new hire still learning the core of the job, AI that drafts the work skips the practice the job was supposed to provide. A conference speaker argued that someone at the governance table should bring an educational lens before a tool goes enterprise-wide. Ask who on your AI governance committee speaks for the people still learning.

Measure with and without AI where you can. Track performance with AI and without it as separate numbers. Most roles already have a quality check, like coding audits, report validation, or chart review. Start there, run it both ways for the highest-risk tools, and name an owner.

Schedule AI-off practice. Set aside periods of routine work or simulation with the tool switched off. Health systems already run EHR downtime drills for the same reason. Panelists made the resilience case plainly: systems go down and cyberattacks happen, some skills have to remain when they do, and keeping them sharp is the organization’s job.

Make it everybody’s training. The clinical conversation gets the attention. The revenue cycle, access, analytics, IT, and management teams use AI every day too, and they need the same three steps. Ask each team what work is changing and build practice around the responsibilities people will take on next.

Some skills should fade. Which ones?

A keynote speaker called deciding what to stop teaching the hard part, and warned against cutting the reasoning skills underneath.

Some skills should move to the tool. Work changes, and skills shift with it; that’s fine as long as the organization keeps the capabilities people need when the tool fails. Calculators and autopilot are the usual comparisons, and they help up to a point. A calculator gives the same right answer every time. A language model can be confidently wrong, and that difference is why mis-skilling exists at all.

Which skills your people can safely hand off is a question for the leaders who know the work. Ask it before a vendor’s rollout plan answers it for you.

The question for your organization

The Joint Commission’s certification checks that training happened. Measuring what people can still do without the tool is up to you.

After Austin, I’d open every AI literacy conversation with two questions. Which skills in your organization must survive the day the AI goes dark? And what new responsibilities should people be learning as the work changes?

If you’re building AI literacy and governance and want to talk through the sequence, schedule a call. For the governance side, see why education has to come before purchasing.

Frequently Asked Questions

What is AI deskilling?

AI deskilling is the loss of a skill a person once had after a period of relying on AI to do that work. It’s one of several ways skills slip around AI, along with never-skilling, mis-skilling, automation bias, and cognitive offloading. The most cited example is an observational study where unaided colonoscopy detection rates fell after centers adopted AI; the authors say AI exposure “might reduce” unaided detection.

What is never-skilling?

Never-skilling describes people who rely on AI while they’re still learning a job and never build the underlying skill. It matters most for new hires and trainees in any role. In a field experiment with high school math students, those who used GPT-4 freely scored 17% worse once access was removed, while a version that gave hints instead of answers largely avoided the harm.

What is re-skilling, and how does it fit with AI adoption?

Re-skilling means learning new skills for a changed role or a different kind of work. As AI takes on tasks, organizations can give people supervised practice in their next responsibilities while keeping essential existing skills in use. AI-assisted practice can offer hints and questions that keep the learner reasoning, with a qualified person checking the material. Measure whether people can do the new work.

Does this only apply to clinicians?

No. The research and conference conversation started in medicine, and the same patterns show up for analysts, coders, schedulers, managers, and anyone else who works next to AI. A 2025 survey of 319 knowledge workers linked higher confidence in generative AI with less critical thinking, and the Joint Commission’s AI certification samples training records for coders, schedulers, and case managers alongside nurses and physicians.

Does AI literacy training help prevent over-reliance on AI?

Yes, as the foundation. Literacy training gives everyone the shared vocabulary and the reasons behind the habits that protect them. Research points to which habits matter: in a study of 199 people, asking people to decide before seeing the AI’s suggestion significantly reduced over-reliance, in a process-control study, operators who experienced an aid’s failures during training checked it more carefully, and in a field experiment with students, an AI tutor that gave hints instead of answers protected learning. I’d pair the training with role-specific practice and measurement of performance with and without AI.

How can organizations protect how new employees learn with AI?

Have new hires build the core of the job without AI first, then practice finding errors in AI output, then use AI in real work with trust extended as competence shows up. Choose tools that give hints and ask questions for learners, run quality checks with and without AI for high-risk work, and schedule periodic AI-off practice.

References