Cognitive debt and kids: what AI can do to thinking, and how to avoid it

Cognitive debt is what builds up when AI does thinking a child would otherwise do: the homework gets easier today, but they remember less, own the work less and practise the skill less. The evidence so far comes mostly from teens and adults, but it points one way: let kids think first, and use AI for hints and questions, not finished answers.

Written by the Ask Poppy team, who make one of the tools mentioned here. The advice works with any AI, or none. We haven't run a controlled study of our own app.

Where the term comes from

In June 2025, researchers at the MIT Media Lab posted a preprint called Your Brain on ChatGPT (Kosmyna et al., 2025). Fifty-four adults aged 18 to 39 wrote essays using ChatGPT, a search engine, or nothing, while EEG recorded their brain activity. The ChatGPT group showed the weakest brain connectivity during writing, felt the least ownership of their essays and often struggled to quote what they had just written. The authors called the long-run cost of this "cognitive debt". (Our summary: Your Brain on ChatGPT, in plain English.)

Keep its limits in mind: it's a preprint, not yet peer-reviewed; it used adults, not children; only 18 people completed the last session; and it measured a single type of task. It does not show permanent damage. It's a warning light, not a diagnosis.

What other studies add

• Teenagers, real classrooms. About 1,000 high-school students in Turkey practised maths with an unrestricted GPT-4 ("GPT Base"), a version that gave hints ("GPT Tutor"), or no AI. GPT Base raised practice scores by 48% but scores on the later test without AI were 17% lower than the no-AI group. GPT Tutor raised practice scores by 127% with no significant drop on the test. Students using the unrestricted version mostly asked for answers and didn't notice the harm (Bastani et al., PNAS 2025; our summary).
• Easier isn't better. University students researching with ChatGPT felt less mental effort than those using Google, but their reasoning was weaker (Stadler et al., 2024).
• Overconfidence. Adults solving logic puzzles with AI did slightly better but overestimated their performance by even more (Fernandes et al., 2026).
• Habit and critical thinking. In a UK survey of 666 people, heavier AI use went with lower critical-thinking scores, mostly through "cognitive offloading", and the youngest group (17 to 25) relied on AI most (Gerlich, 2025). It's correlational, so it can't show cause.

The other side matters too. When AI is set up to ask questions and give hints, with a teacher involved, studies have found real gains: for example a supervised AI tutor for UK pupils aged 13 to 15 (LearnLM and Eedi, 2025) and teacher-guided AI sessions in Nigeria (De Simone et al., 2025). The pattern: AI that thinks for you costs you; AI that makes you think can help.

Why children are a special case

Children are still building the skills adults offload: reading stamina, mental arithmetic, organising an argument, sitting with a hard problem. An adult who hands an essay to AI is skipping a skill they already have. A 10-year-old is skipping the practice that builds it. That's an inference, not a finding: there's no long-term study of AI and young children yet. But it's a good reason to be more careful the younger the child.

Five rules of thumb

1. Try first. Two honest minutes on the problem, out loud or on paper, before any AI.
2. Hints, not answers. Ask for one hint at a time (our hint ladder writes that prompt for any AI).
3. Explain it back. After using AI, the child explains the idea without the screen. If they can't, the AI did the learning. Try explain it back.
4. Own the output. Anything handed in should be something the child could rewrite and defend tomorrow.
5. Protect AI-free time. Reading, play, making things, chores, being bored. These are where attention and imagination grow.

Age by age

What to look for in an AI tool

• It asks what the child already thinks before explaining.
• It gives hints step by step, and doesn't jump to the full answer by default.
• A parent can see conversations and control settings such as memory.
• It's built for children, with an age rating and terms that allow your child's age (many general chatbots require 13+).
• It's honest that it can make mistakes.

Where Ask Poppy fits

We built Ask Poppy around these ideas. It's a voice-first companion: a child holds to talk, and Poppy asks what they already know and offers one hint at a time instead of leading with the answer. Parents can add up to three child profiles, open each child's conversations (children are told a parent can see them), and turn memory off per child. A one-button, screen-free Poppy device is in development. You can reserve one on askpoppy.ai with a fully refundable $30 toward a planned $200 price; estimated shipping is spring 2027. The app works fully without it. Like any AI, Poppy can make mistakes, which is one more reason to keep an adult involved.

Sources (checked 26 Oct 2026): Kosmyna et al., "Your Brain on ChatGPT", arXiv preprint, June 2025 · Bastani et al., PNAS, June 2025 · Gerlich, Societies 15(1), 2025. Our plain-English research summaries are on research.askpoppy.ai. Spot something out of date? Email [email protected].

Questions

What is cognitive debt?

It's a term from a June 2025 MIT Media Lab preprint (Kosmyna et al.) for what builds up when you let AI do thinking you would otherwise do yourself: the task gets easier now, but you remember less, feel less ownership of the result and practise the skill less. The study was small, used adults and isn't peer-reviewed yet.

Does ChatGPT make kids lazy thinkers?

It can if it does the work for them. In a PNAS 2025 study of about 1,000 high-school students, unrestricted ChatGPT help raised practice scores but lowered scores on a later test without AI by 17%. A version that gave hints instead did not cause that drop. The tool matters less than whether the child still does the thinking.

Was the MIT cognitive debt study done on children?

No. It used 54 adults aged 18 to 39 writing essays, with only 18 completing the final session. It measured brain activity with EEG during writing. It suggests a risk worth taking seriously, not proof of lasting harm in children.

How do I stop my child building cognitive debt with AI?

Have them try first, out loud or on paper, before asking AI. Ask the AI for hints, not answers. Afterwards, have them explain the idea back without the screen. And keep plenty of AI-free time for reading, play, making things and boredom.

Is all AI use bad for children's brains?

No. Studies of AI tutors that ask questions and give hints, with a teacher or adult involved, have found learning gains. The risk comes mainly from handing the whole task over.

More on kids and AI

• Does AI make kids dumber? What the research shows
• Does AI hurt kids' writing and creativity?
• Protect your child's thinking skills, age by age
• A Waldorf-inspired approach to AI at home
• Screen-free AI for kids: voice first, screens later
• All guides on raising kids alongside AI
• AI tutors compared side by side

By Simon Dreyer, founder of Ask Poppy · Last updated 26 October 2026 · How we write these pages