Is AI Making Us Look More Capable Than We Are?

The clearest evidence so far comes from education, and it's worth taking seriously at work. A study of nearly 1,000 high-school maths students in Turkey, published in PNAS in 2025, gave some students a standard ChatGPT-style assistant during practice sessions. Their practice scores rose by 48%. When the assistant was removed for the exam, they scored about 17% lower than students who had never had access at all. A second version, designed as a tutor that offered hints rather than answers, improved practice scores by 127% and largely avoided that drop.^[1]

The OECD reached a similar conclusion in its Digital Education Outlook 2026: completing tasks successfully with generative AI does not necessarily translate into learning, and the advantage often disappears once the tool is taken away.^[2] These are studies of students rather than employees, but the mechanism isn't specific to classrooms. Anyone who uses AI to produce work they couldn't have produced alone is, in effect, practising with the tool switched on. The output improves immediately. Whether the person's own capability improves depends on how the tool was used.

A field experiment with 758 consultants at Boston Consulting Group shows why this matters even for experienced professionals. On a set of realistic consulting tasks that sat within AI's capabilities, consultants using GPT-4 completed 12.2% more tasks, finished 25.1% faster and produced work of significantly higher quality than those working without it. On a different task, deliberately chosen to fall outside what the AI could do well, consultants using it were less likely to reach the correct answer than those who worked unaided.^[3]

The researchers called this a "jagged technological frontier". AI is excellent at some tasks and unreliable at others, and the boundary between the two isn't obvious from the output, which looks equally confident either way. That has a practical consequence I think organisations underestimate. Being capable now includes knowing where the frontier lies for your own work: which tasks you can safely hand to AI, which need careful checking and which you should do yourself. That knowledge comes from expertise, and it's hard to develop if AI has always done the work for you.

A study by Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers about 936 real examples of using generative AI at work. People who had more confidence in the AI tended to apply less critical thinking to its output. People who had more confidence in their own ability to do the task tended to apply more.^[4] The researchers also found that critical thinking doesn't disappear when AI is involved, but changes shape, moving towards verifying information, combining AI responses with other material and taking responsibility for the overall task.

It's a self-reported survey, so it shows associations rather than cause and effect, but the pattern rings true. Checking AI output well depends on knowing enough to recognise when something is wrong. If people's own expertise slowly erodes because the AI is usually right, their ability to catch the occasions when it isn't erodes with it. (The fifth article in this series looks at what happens when this affects decisions about other people. The concern here is quieter: the gradual loss of the expertise that makes checking possible in the first place.)

In the World Economic Forum's Future of Jobs Report 2025, employers expected 39% of workers' existing skill sets to change or become outdated between 2025 and 2030. The same employers ranked analytical thinking as the most sought-after core skill, ahead of the technology skills that are growing fastest.^[5] That fits the evidence above. Operating AI tools is becoming a basic expectation, but the scarce capability is judging what they produce: spotting the assumption that doesn't hold, the figure that doesn't reconcile or the recommendation that won't survive contact with the real organisation.

In enterprise technology that distinction is familiar. The most expensive problems rarely look like problems at first. They look like a plausible design, a report that adds up or a configuration that passes its first test. Catching them depends on someone who understands the business process well enough to ask why, and that understanding is built by doing the work, not by reviewing someone else's.

If finished work no longer shows reliably what someone understands, organisations need other ways to see it. Most of the answer lies in asking people to show their reasoning rather than just their output. In a performance review, that might mean asking someone to walk through how they reached a recommendation, what they checked and what they rejected. In recruitment, it might mean replacing a take-home task that AI can complete with an exercise in which candidates critique an AI-generated analysis and explain what's wrong with it. In project reviews, it might mean asking who verified the AI's contribution and how.

There are also good reasons to keep some assessments free of AI. Professional certifications and safety-critical roles need to know what someone can do independently, because there will be moments when the tool is unavailable, wrong or not permitted. Other assessments should allow AI and judge how well people use it. Both are legitimate, and I think much of the current confusion comes from organisations treating them as the same thing without saying which one they're measuring.

None of this is an argument for keeping AI away from people who are still learning. The tutor version in the maths study, which gave hints instead of answers, produced the largest gains of all. The customer-support research discussed in the first article in this series also found suggestive evidence that agents learned from the AI's recommendations, picking up the practices of the company's best performers.^[6] Used well, AI can speed up the development of expertise. Used carelessly, it can skip that development entirely while making the output look the same.

The difference is largely a matter of how organisations set things up. Encouraging people to ask AI to explain, challenge or test their work, rather than just produce it, makes a real difference. So does building review points where someone has to explain the reasoning behind AI-assisted work, and designing internal assistants that coach as well as complete. Training matters too, but it has to be specific to be useful. Hours of generic AI training achieve little if they don't help people understand their own work, recognise where AI is unreliable in their own field and make better decisions with it.