We need to preserve the Friction.
A thousand high-school students in Turkey were given ChatGPT to help with their maths practice.
Their practice scores went up 48%. Then the researchers took the AI away and gave everyone the same exam.
The students who had practised with ChatGPT scored 17% lower than classmates who never had it at all (Bastani et al., 2025).
They did better than the textbook group with the tool, and worse without it. I haven’t been able to stop thinking about that study, because it describes something I see coming for every industry, including mine.

The moment it clicks
Most of us remember a problem we couldn’t solve. You stare at it and try three approaches. You get annoyed and nearly give up. Then, suddenly, it clicks.
You never forget that one.
That isn’t nostalgia. It’s one of the most replicated findings in learning science. Robert and Elizabeth Bjork at UCLA call it “desirable difficulty” (Bjork & Bjork, 2011). Learning that feels hard and slow tends to stick. Learning that feels smooth and easy tends to disappear.
One classic experiment shows how big the effect is. Students had to learn a passage of text. One group read it four times. The other read it once and then tried to recall it from memory three times.
Five minutes later, the re-readers were ahead. One week later, the recall group remembered 61%, and the re-readers just 40% (Roediger & Karpicke, 2006).
The friction of pulling information out of your own head is what builds the memory. Reading the answer again feels productive, but it mostly isn’t.
Why friction feels like failure
This is the cruel part.
Harvard physicists randomly split students into two groups. One sat through polished lectures from highly rated teachers. The other worked through problems actively in class, which is messier and more uncomfortable.
The active group learned more. But they felt they had learned less (Deslauriers et al., 2019).
Smooth feels like progress. Friction feels like failing. So we naturally choose the smooth path, even when it’s the worse one.
Now think about what generative AI is: the smoothest, most fluent, most patient explainer ever built. It never makes you feel stuck. That is exactly why it is so risky for learning.
It’s not just maths
The same pattern keeps showing up wherever a tool does the thinking for us.
Software developers. In a randomised trial, mostly junior developers learned a new programming library, some with AI help and some without. Afterwards, the AI group scored 50% on a test of what they had just used, and the group without AI scored 67%. That’s roughly two letter grades (Shen & Tamkin, 2026).
Drivers. Researchers at McGill tested 50 drivers. The more people had relied on GPS over their lives, the worse their spatial memory when they had to find their own way. Over three years, heavier GPS users declined faster (Dahmani & Bohbot, 2020).
Doctors. This one stopped me. In four Polish clinics, experienced doctors with more than 2,000 colonoscopies each began using AI to help detect pre-cancerous growths. A few months later, in procedures without the AI, their detection rate had dropped from 28.4% to 22.4% (Budzyń et al., 2025). These were experts, and they got measurably worse at their core skill once the machine started helping. It’s one observational study, and a later Norwegian study didn’t find the same drop (Pedersen et al., 2026), but the warning is hard to ignore.
Knowledge workers. Microsoft and Carnegie Mellon surveyed 319 professionals who use AI weekly. The more people trusted the AI, the less critical thinking they applied to its output (Lee et al., 2025).
A psychologist named Lisanne Bainbridge predicted this back in 1983. She warned that skills fade when they aren’t used, so an expert who spends their days watching an automated system becomes an inexperienced one. The cruellest irony is that they get called in exactly when something goes wrong (Bainbridge, 1983).
The escalator and the gym
Here’s how I think about it.
Escalators, lifts, cars and washing machines quietly took the effort out of daily life. Mayo Clinic researchers measured it: household machines alone cut our energy use by about 110 calories a day (Lanningham-Foster et al., 2003). That’s small daily, and huge over a decade.
We didn’t respond by smashing the washing machine. We invented the gym. We accepted that if daily life no longer gives us the friction, we have to schedule it on purpose.
AI is the escalator for the mind. I use it, and I’m not going to stop. But if it carries us up every flight, friction has to become something we build in on purpose.
The moment in front of the guest
In hospitality, the real test of skill doesn’t happen at a desk. It happens face to face, in real time, with no screen to turn to.
Picture a guest arriving at the front desk after a 14-hour flight to find that the room they booked for their anniversary isn’t ready. They’re tired, disappointed and starting to raise their voice. There are other guests in the queue.
There’s no time to open an app and type a prompt. In the next ten seconds, the associate has to read the emotion, choose the right words, decide what to offer and own the moment. Either they have that judgment in them or they don’t.
Or picture a bartender when a regular says, half-joking, that tonight is the worst day of their year. Or a duty manager in the lobby at 2 a.m. when a fire alarm goes off and two hundred guests are looking at them for what to do next.
These moments are where hospitality is won or lost, and AI can’t do them for us.
The risk is subtle. If every service-recovery email, every complaint reply and every “how should I handle this?” gets drafted by AI, our people never do the reps. They get very good with the tool, and quietly worse without it, just like the maths students and the doctors.
Then the guest stands in front of them, and the escalator isn’t there.
Great service is a muscle. It’s built in the hard conversations you had to work out yourself.
The good news: friction can be designed back in
This isn’t an argument against AI. The research is clear on that too.
In the same Turkish study, a second group used a version of ChatGPT that was told not to give answers but to guide students with hints instead. Their learning loss disappeared (Bastani et al., 2025).
At Harvard, an AI tutor built to give only one step at a time, and to make students try before it helped, more than doubled learning gains compared with an excellent classroom (Kestin et al., 2025).
In the coding study, how people used AI mattered more than whether they used it. Developers who asked “why does this work?” scored 65% or higher. Those who simply let the AI write the code scored below 40% (Shen & Tamkin, 2026).
The line isn’t AI versus no AI. It’s AI as an answer machine versus AI as a sparring partner.
Gym rules for the AI age
These are the habits I believe in, for myself and for any team:
- Try first, then ask. Draft the guest reply, the forecast or the plan yourself. Then compare it with the AI version. The gap is where you learn.
- Use AI as a sparring partner, not a ghostwriter. Ask it to play the angry guest and let your team practise responding live. “Challenge my answer” beats “Write it for me.”
- Rehearse the live moments. Role-play the hard conversations at the front desk, in the restaurant and on the phone. The moment in front of the guest can’t be outsourced, so it has to be practised.
- Explain it without the screen. If you can’t say why the answer is right, you haven’t learned it. You’ve borrowed it.
- Expect it to feel harder. Tell your team openly that friction is the signal of learning, not of failure.
- Protect your juniors most. They gain the most from the struggle, and they’re the least able to spot when the AI is confidently wrong.

Friction is the point
We spent a century engineering effort out of our bodies, and then paid for gym memberships to put it back.
We’re now doing the same with our minds, only faster. The leaders who do well won’t be the ones who use AI the least. They’ll be the ones who are deliberate about where they keep the friction.























