Question Everything, With AI

The Daily Question: one item from the news, turned into one question a parent can ask.

AI can make your child’s homework better. Whether it makes your child a better learner depends on who does the questioning.

Two recent studies, both updated in October 2026, asked the question I think every parent should now be asking about AI. When it helps with the work, does the person doing the work get any better?

Read carefully, their answer is: not by itself. In both, AI made the work better while people had it. What was left once it was taken away differed with who they were, and perhaps with how they had used it.

That is the heart of what I want to say to Sri Lankan parents. Banning AI is not the answer. Letting AI answer for your child is not the answer either. I would put it this way:

Your child should question everything, with AI, and reason through everything it responds with. Think, write, critically analyse, design the next one. And loop.

In simple terms, the AI can help with the work, but your child has to do the questioning and the reasoning. That is the part that makes a learner, and it is a habit you can build at home.

There is a date that makes this practical. The G.C.E. (O/L) examination runs from 8 to 17 December 2026, and the Department of Examinations timetable is plain about the hall: “it is a punishable offence even to bring any electronic device including mobile phones into the examination hall”. On those days, only what your child can do alone will count.

What did the new research find?

The first study is about lawyers. In a three-month randomised trial with 133 patent lawyers at eleven US law firms, AI access raised the quality of their patent drafting: by 0.34 standard deviations (a standard way of measuring how big a difference is) at 10 days, with larger gains for junior lawyers. After three months, everyone marked up a patent application without AI. The lawyers who had used it did better overall (0.32 SD), but that advantage came entirely from senior lawyers, those with seven or more years’ practice (0.45 SD). Junior lawyers showed no average gain. Their scores spread out instead, which the authors read as AI being “a springboard for some juniors and a cushion for others”.

The limits matter. It is a working paper, not yet peer reviewed. Google ran the study and paid for the experiment, six of the seven authors work at Google, and the participants were adult professionals, not schoolchildren.

The second study is about persistence. In three short online experiments with 1,222 adults in the United States, people who had used an AI assistant for about ten minutes solved fewer problems once it was taken away than people who never had it, and in some experiments gave up more often, especially those who had asked the AI for answers rather than hints. The authors caution that this last comparison is not necessarily causal. The AI gave complete answers on demand, everyone was tested minutes after it was removed, and the authors say whether the effect lasts is unknown. It is a preprint.

Look again at the lawyers. The juniors gained more while the tool was in their hands, and showed no average gain once it was gone. Both studies show a gap that is easy to miss: the work in front of you can improve while the person doing it does not.

Is there evidence from students?

Yes, and it points the same way.

In Generative AI without guardrails can harm learning, researchers worked with nearly a thousand high school maths students at one school in Türkiye. When the AI was taken away, those who had used a standard chatbot did 17% worse than students who never had it, on an exam taken without AI. A tutor version that gave teacher-designed hints instead of answers largely mitigated that harm.

Notice where the harm came from: AI that handed over answers.

A preprint, AI Meets the Classroom, reports two pre-registered lab experiments that found no overall effect on learning of large language models, the AI behind chatbots like ChatGPT. In exploratory analyses and a field study, students who used them to generate solutions covered more topics but understood each less, those who asked for explanations understood more, and the models widened the gap between students with low and high prior knowledge.

And AI can teach well when it is built to. A tutor built on the same teaching practices as the class helped Harvard physics undergraduates learn more in less time, measured immediately after the lesson.

So I am not making a case against AI. I am making a case about how it is used.

One honest limit: none of these trials studied Sri Lankan schoolchildren. I am reading across from other countries, other ages and other subjects, and you should read me with that in mind.

Why doesn’t better work mean better learning?

Because the effort that produces learning is often the effort that AI removes.

How well you do while practising is a poor guide to what you will remember. As a 2015 review by Soderstrom and Bjork puts it, “people often mistakenly interpret their performance during acquisition as a reliable guide to long-term learning”. The idea is sometimes called “desirable difficulties”: some struggle is part of how learning happens.

A classic experiment shows it well. Students studied prose passages, then either restudied them or tried to recall them from memory, with no feedback. Roediger and Karpicke found that restudying did better on a test five minutes later. Two days and a week later, recall had won: “prior testing produced substantially greater retention than studying, even though repeated studying increased students’ confidence in their ability to remember the material.”

A 2013 review of ten study techniques rated practice tests and distributed practice (spreading study out over time) “high utility”, and underlining and rereading of “surprisingly low utility”.

Put those together and the AI problem is easy to see. A finished answer from a chatbot is a very comfortable form of rereading. It feels like understanding. It looks good in the exercise book. But nobody had to pull anything out of their own head to get it.

What does “question everything, with AI” look like?

Here is my loop, step by step. Where there is evidence, I link it. Where there is not, I tell you it is my reading.

Question

Not “give me the answer”, but “why does this work?” or “where did I go wrong?”. A 2007 US Department of Education practice guide rates “ask deep explanatory questions” as backed by strong evidence. That guidance predates generative AI and was not tested with it, but the habit it describes is the one I mean.

Reason through everything it responds with

The answer is not the end. Does this make sense? Can I follow every step? In the classroom study, asking for explanations went with more understanding and asking for solutions with less. In the persistence study, the extra giving-up sat with people who said they had asked for answers. Both comparisons are exploratory or observational, not proof, but they lean the same way as the Türkiye trial.

Think, and write

Then close the chat and write it out from memory, in your own words. This is my application of the recall research: writing what you remember, without looking, is the kind of effort Roediger and Karpicke found pays off days later.

Critically analyse

Is the AI’s answer actually right? Is mine? What did I get wrong, and why? I know of no trial on checking AI answers. My case rests on the Soderstrom and Bjork point: we often mistake how well practice is going for how much we have learned, so checking cannot be left to that feeling.

Design the next one

Your child makes the next question: a harder version, a different number, a “what if”. This step is my reading. I have not seen a trial of children designing the next question for an AI.

And loop

Do it again, and come back to it another day. Practice tests and spreading study out over time are the two techniques the 2013 review rated highest. The loop is a way of doing both.

The loop as a whole has not been tested. I have found no study that taught children these steps with an AI and measured the result. It is my proposal, built from tested parts.

Doesn’t a child need help, not just hard questions?

Yes. This is the strongest objection to what I am saying.

Kirschner, Sweller and Clark argue that leaving beginners to discover things on their own “is less effective and less efficient than instructional approaches that place a strong emphasis on guidance”. In their reading, guidance starts to matter less only once learners know enough to guide themselves. Critically analysing an answer and designing the next question are things experts do well. A child who does not yet know the topic cannot simply be told to question everything and left alone.

The Türkiye study makes the same point. What reduced the harm was not the students’ willpower. It was the design of the tutor. And the classroom study’s exploratory analyses found that AI widened the gap between students who knew more and those who knew less. A loop that only works for the strongest would make things worse.

So the loop needs guidance, not abandonment. In practice, that means asking the AI for a hint or an explanation rather than the answer, staying close to the textbook and the teacher, and having an adult who asks the questions with the child until the child can ask them alone. Questioning is something you learn by being questioned.

What can you do at home this week?

You do not need to understand the subject to do this. You only need to ask.

Tonight, ask your child to show you one answer the AI gave them. Ask them to walk you through why it is right. Ask how they would check it. Then make the next question together, close the device, and let them try it alone.

Do it again next week, and the week after. If your child is sitting the O/L, the Mathematics paper is on Monday 14 December, and that room will have no AI in it. Every loop between now and then is a rehearsal for the day it is just your child, a pen and the paper.

This is a habit, not a guarantee, and I cannot promise you a grade. What I can say is that a child who has practised questioning, reasoning and checking has practised what the exam hall asks for.

Where this comes from

At Idasara, the idea we keep returning to is learning how to learn: that the most useful thing a student can carry out of school is knowing how their own learning works. Our book Learn to Learn is about how learning works, written for Sri Lankan O/L and A/L students and their parents and teachers, and you will find it through Idasara Academy. The loop in this essay is my own reading of the research, not a feature of any product. It comes from a belief I set out in The Question Is the Last Human Advantage: AI will keep getting better at answers, so our children have to get better at questions.

So, tonight, ask one question. Then ask the next one. And loop.

---

Sources

1. Autor et al., Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting, NBER Working Paper 35720 (Sept 2026, rev. Oct 2026). Working paper; Google-run and Google-funded.
2. Liu, Christian, Dumbalska, Bakker and Dubey, AI Assistance Reduces Persistence and Hurts Independent Performance, arXiv:2604.04721 v5 (3 Oct 2026). Preprint.
3. Bastani et al., Generative AI without guardrails can harm learning, PNAS 122(26) (2025).
4. Lehmann, Cornelius and Sting, AI Meets the Classroom: When Do Large Language Models Harm Learning?, arXiv:2409.09047. Preprint.
5. Kestin et al., AI tutoring outperforms in-class active learning, Scientific Reports 15:17458 (2025).
6. Soderstrom and Bjork, Learning versus performance: An integrative review, Perspectives on Psychological Science (2015).
7. Roediger and Karpicke, Test-enhanced learning, Psychological Science 17(3) (2006).
8. Dunlosky et al., Improving Students’ Learning With Effective Learning Techniques (APS summary, 2013).
9. Pashler et al., Organizing Instruction and Study to Improve Student Learning, IES Practice Guide NCER 2007-2004 (2007).
10. Kirschner, Sweller and Clark, Why Minimal Guidance During Instruction Does Not Work, Educational Psychologist 41(2) (2006).
11. Department of Examinations, Sri Lanka, G.C.E. (O/L) Examination 2026 Time Table (6 Oct 2026).