The short answer
“Dumber” is two different claims. One is that AI use damages your ability to think. There is no good evidence for that yet. The studies are new, small, or show only a correlation.
The other claim is that you do not learn what you hand to a machine. The evidence for that is good, and it is older than AI.
What the studies found
| Year | Study | What it found |
|---|---|---|
| 2011 | Sparrow, Liu and Wegner, Science | People who expected a fact to be saved on a computer remembered the fact less well, and remembered where to find it better. |
| 2016 | Risko and Gilbert, Trends in Cognitive Sciences | A review of cognitive offloading: the use of a tool or the world to hold or do mental work. It saves effort now. What is offloaded is remembered less. |
| 2025 | Bastani and others, PNAS | A field experiment with nearly 1,000 high school math students. With an AI like standard ChatGPT, practice grades rose 48 percent. On a later exam with no AI, those students scored 17 percent lower than students who never had it. A tutor version with guardrails largely removed the harm. |
| 2025 | Gerlich, Societies | A survey of 666 people. More use of AI tools went with lower scores on a critical thinking measure. This is a correlation. It does not show which one causes the other. |
| 2025 | Kosmyna and others, arXiv preprint | 54 people wrote essays with ChatGPT, a search engine, or no tool, with EEG recorded. The ChatGPT group showed the weakest brain connectivity and recalled their own essays worst. It is a preprint with a small sample, and it has not been peer reviewed. |
What the strongest study says
The 2025 field experiment is the most useful, because it is large, it is randomized, and it tested two designs. Both AI tools made practice go better. Only one of them hurt the later exam.
The standard chat design gave answers. Students used it, in the authors’ word, as a “crutch”, and then did worse alone. The tutor design gave hints and held back the answer. Its students practiced far better and did about as well as the control group on the exam.
So the question is not “AI or no AI”. The question is who does the thinking.
Why it happens
Memory is built by effort. Pulling an answer out of your own head strengthens it, and reading an answer does not. This is retrieval practice, and it has been tested for decades.
A chatbot removes the effort. It is instant and it is fluent, so its answer feels like your own understanding. The feeling is real. The memory is not there, because nothing asked you to make it.
How to use AI and still learn
- Try first. Write your own answer before you ask. Then compare.
- Ask for a hint, not the answer. This is the difference between the two tools in the field experiment.
- Have it ask you. Tell the AI to quiz you on the material and to hold back the answers until you reply.
- Recall later with no tool. The exam in the study had no AI. Test yourself the same way, a day later.
- Offload what you do not need to keep. A phone number is fine to hand off. A skill you will be tested on is not. More in cognitive offloading.
Where Gillo stands
Gillo uses AI to write the course. It does not use AI to do your thinking. A Bite gives you a short read and then asks five questions. You answer from memory, and XP comes only from recall. Keep Sharp asks again later, with no text on the screen.
That is the guardrail design from the research, built into the product: the AI prepares the material and holds back the answer until you have tried.