Does AI make you less resilient?
A new study from Berkeley, Carnegie Mellon, Oxford, MIT and UCLA found that people who used an AI chatbot for just 10 minutes got more answers wrong and gave up more often once it was taken away. I look at what that means for resilience, for learning design, and for the hard-won habit of digging for an answer yourself.
It’s 2012. I’m a few years into teaching science, sitting through an INSET day, when a slide goes up that I’m sure many of you have seen: the Learning Pyramid. We remember 10% of what we read, 20% of what we hear, and a glorious 90% of what we teach to others. The room nods. I write it down. Then the scientist in me asks a simple question: says who?
So that evening I went looking. Google gave me hundreds of copies of the same pyramid, all beautifully coloured, all confidently cited, and none of them pointing to an actual study. I tried Google Scholar and hit paywall after paywall. I followed one citation back to another, which cited another, which cited a training organisation in Maine, which (as far as anyone has been able to find) has never produced the original research. Several evenings, a lot of tea and one very patient librarian later, I ended up at Edgar Dale’s “Cone of Experience” from 1946. It’s a perfectly sensible model. It also contains no percentages whatsoever. Somebody, somewhere, had just made them up, and the rest of us had been passing them around ever since.
Looking back, that hunt taught me something. I learned to rephrase, to dig, and to keep going when nobody handed me the answer. Today I could ask a chatbot and have an answer in three seconds flat. So here’s the question: if AI hands us every answer, do we lose the grit we built along the way? Would we still be as resilient?
As luck would have it, a team from UC Berkeley, Carnegie Mellon, Oxford, MIT and UCLA has just looked into exactly that. What are the chances!
They recruited around 1,060 adults in the US online and gave them some problems to solve: fractions in two of the experiments and SAT-style reading questions in the third. Half of them got a chatbot in a sidebar (GPT-5, dressed up as “ChatGPT”) that already knew every answer. Everyone had a “skip” button, so giving up was always just one click away. Then, after about 10 to 15 minutes, the researchers pulled the sidebar away without warning, and the AI group had to carry on without the all-knowing oracle that had been sitting next to them. In other words, they went from having the answer handed to them to having to dig for it themselves. They actually ran three separate experiments, each one tightening up the design of the last, but if you want that much detail, you’ll have to go and find the full paper yourself (I’ll make it easy and put the link at the bottom).
What did they find?
A quick health warning first. There were three experiments, so I’ve taken averages and done some rounding that my old statistics lecturer would not approve of. Treat them as ballpark figures, and if you want to know the numbers precisely, go and read the full paper!
With that out of the way, unsurprisingly, while the AI was there, the people who had it did better. On average, they got around 14 more questions right out of every 100 than the group without it. This is probably because people were handing the thinking over to the machine whenever they got stuck (or performing cognitive offloading). Skipping the challenge of the question can feel great in the moment, you may still get the endorphin rush from getting the answer correct, but it may also impact how much effort your brain expects a problem to take.
Then the chatbot vanished, and things flipped almost immediately. The group who never had AI carried on at about the same level, as you’d expect, because they had been doing the hard work all along. The AI group dropped from getting about 90 out of every 100 right to about 68. That’s roughly a quarter fewer correct answers, and it left them about 12 out of 100 behind the people who never had AI at all. Turns out that all of that cognitive offloading may have offloaded more than they bargained for!
So, you want more evidence that the cognitive offloading was having an effect? Well, once the chatbot had gone, the AI group hit the skip button on about 13% of questions, compared with about 6% for the group who never had it. More than double! So, as well as getting more wrong, they were much quicker to throw in the towel. Ten minutes of instant answers seems to have been enough to make the slow, effortful route feel like it wasn’t worth the bother.
My favourite finding came when the researchers asked people how they had used the chatbot. Those who mostly asked it for straight answers did worst once it was gone, getting 65% right. Those who asked it for hints or explanations, and so still did some of the digging themselves, got 76% right. That’s almost exactly level with the people who never had AI (77%).
There are plenty of limitations to this study, and one major one is time. The whole thing lasted about 15 minutes and everyone was tested straight away, so it can’t tell us whether this builds into a habit or fades over time. The participants were also online volunteers with nothing riding on their scores, which is rather different from a student sitting a real exam. All we can really say is that if you give someone the answers and then take them away, they do worse and give up sooner, at least in the short term.
What does this tell us about learning design and AI?
I think this shows something we’ve known in education for a very long time: you have to develop a person’s ability to think and search for themselves. If you simply hand them the answer, they’ll take it and come back, Oliver Twist style, with an empty bowl asking for more. They will never really understand how to get there on their own.
So we need to design AI systems that behave more like a good teacher or a good librarian: ones that challenge us, ask us questions and point us in the right direction, and only hand over the answer once we’ve had a proper go ourselves.
Which brings me back to that INSET day and the Learning Pyramid. Would a chatbot have saved me those evenings of digging? Possibly. It might also have confidently told me that the percentages came from Edgar Dale in 1946, because that’s what half the internet says. The hunt taught me to check, to question and to keep going when the easy answer looked suspicious, and those are exactly the habits we need most now that the easy answer is always one click away. So next time you reach for an AI, try asking it for a hint instead of the answer. Your future self, and possibly your INSET audience, will thank you.
And, hey, if you like the idea of a chatbot that will help you think, then why not check out Soma. You can find it at soma.obline.org and use it for free during the pilot phase!
Read the full paper: AI Assistance Reduces Persistence and Hurts Independent Performance (Liu, Christian, Dumbalska, Bakker and Dubey, 2026)