“Do flies get sad, too?” “How is hibernation different from sleep?” Ask AI one question, and another comes to mind. A conversation about animals leads to sleep, then to your own life. There is genuine pleasure in seeing unfamiliar things start to connect.
Yet sometimes, when you close the chat and try to explain it to someone, the words will not come. All you remember is the feeling that “it all fits together.” There is still a step between a conversation that flows smoothly and being able to think something through for yourself.
To take that step, try turning your conversations with AI from a chance to receive explanations into a chance to test your own. Keep the fun of following new questions, while making room to check the answers.
A useful model helps you think through what happens next
A model, as used here, need not be a complicated equation. It can be a simple sketch that helps you think, “Under these conditions, this is what I would expect.” What matters is being able to explain what happens when the conditions change, rather than giving the idea an impressive name.
Suppose you think, “I am not getting a reply to my email because it is too long.” That explanation suggests a prediction: “If I limit it to one request, it may be easier to answer.” But if the recipient is on vacation, shortening the message probably will not bring a reply.
Now you have two possibilities side by side: the length of the email and the recipient being away. You can also begin to see where your explanation applies. Before settling on “They have not replied because they dislike me,” you have more conditions you can investigate.
Of course, getting a reply to one shorter email does not establish the cause. The other person may simply have had a free moment. Use the model as a starting point for judgment, and revise it in light of what happens.
When a connection is compelling, separate facts from metaphors
Describing animal stress responses, hibernation, and low winter mood as “low-power mode” can make them feel easier to understand. This is where a conversation gets lively, and where it is easy to make a leap.
A study of flies reported changes after repeated stress, including fewer attempts to climb.[1] That does not tell us that flies engage in the same self-criticism as humans. Measured behavior needs to be kept separate from an experience accessible only to the creature having it.
Seasonal depression calls for a separate check. The National Institute of Mental Health (NIMH) describes it as depression with a recurring seasonal pattern.[2] Calling it “a mild form of human hibernation” may be an appealing metaphor, but the comparison alone provides no evidence of a shared mechanism. Take symptoms seriously rather than explaining them away with a metaphor, and consult a healthcare professional if they interfere with daily life.
You can ask AI: “Separate that explanation into facts observed in research, hypotheses drawn from those facts, and metaphors used to help me understand. Which connections need more evidence?” Remember that this classification is itself an AI response, so check the important parts against the original sources.
Solving something with AI does not guarantee you can solve it alone
In an experiment in high-school mathematics in Turkey, performance during AI-assisted practice and learning measured later without AI did not always tell the same story.[3] These findings concern mathematics classes at a single high school in Turkey and the specific AI tool designs tested there. They cannot be generalized to a claim that all AI use harms learning.
Still, they offer a useful way to examine your own habits. After reading and understanding the steps of a solution, can you solve a similar problem with different numbers on your own? After having AI polish an explanation, can you say why the points belong in that order?
On a day when you need an answer quickly, you might ask to see the finished version. On a day when you want to practice, try: “Do not give me the answer yet. Point out just one part of my reasoning so far that I should check.” Choose between these approaches according to what you are trying to do that day.
Five questions for a deeper conversation
- “Here is what I think. Which assumption should I check?” Start by writing your prediction in one sentence. The aim is to see what changes over the course of the conversation, rather than to guess the right answer. If you cannot form a prediction yet, begin with an explanation of a term or a small example.
- “What exactly does that word refer to?” Before moving on with words such as “intelligent,” “stress,” or “efficient,” ask which behaviors or measurements are being discussed. With the flies, first establish what behavior was observed, then consider a comparison with human moods.
- “Can you give an example where this explanation does not apply?” Look for conditions in which the explanation breaks down, just as you might imagine a situation where a shorter email still gets no reply. AI can get counterexamples wrong, too, so the mere appearance of an objection does not settle the question.
- “Which original source can I use to check the claim that matters for this decision?” Ask for the author, title, and URL, as well as the passage supporting the claim. Open the source and read about what was studied, the methods, and the results. Even a real paper may not support the claim made in the conversation. If you cannot verify it, leave it unresolved.
- “What did I leave out when I explained this without looking at the screen?” Close the conversation and give a brief explanation in your own words. Then compare it with the source material and correct any gaps or mistakes. Have AI point out things worth checking, rather than leaving it entirely in charge of grading your answer.
These five questions are not a ritual you need to follow rigorously in every conversation. A casual exchange that ends with “Oh, that is interesting” has value, too. Put in a little extra effort for topics you want to learn or explanations you intend to use in making decisions.
A short practice loop after the conversation
Experiments using science texts have examined the effects of recalling and explaining knowledge on learning.[4] The research did not, however, validate the particular time allocation below. This is simply one routine to try on a busy day.
- 3 minutes: Write down a question that interests you and your prediction
- 5 minutes: Talk with AI and explore one term or counterexample in more depth
- 3 minutes: Close the conversation and write an explanation, along with what you still do not understand
- 4 minutes: Check the key points against the original sources and make any necessary corrections
If you cannot finish checking the paper within those 15 minutes, leave it marked “still to verify.” Keeping track of what remains unchecked makes it easier to continue than forcing a neat conclusion before the time is up. You could also try explaining it again the next day without looking at the screen.
To bring this into your work, choose one small, manageable change, as in the email example. Make just one request, specify the response you need, or say when you need a reply. Record what you changed and what happened, and you have material for the next question. Treat the result as information about the exchange, without assuming it reveals the other person's private thoughts.
When entering a real email or personal situation into AI, remove names, contact details, customer information, and company secrets. Use a fictional example that preserves only the structure you need. For consequential decisions about health, money, or similar matters, seek more than the explanation you received in a conversation.
Let the enjoyment lead to the next question
One of the pleasures of talking with AI is being able to explore a question on the spot, even if it seems childlike. You can ask about an unfamiliar word or have something explained again using your own example. It feels good to have an easier way into a subject.
To turn that curiosity into a workout for your thinking, leave a little time at the end for your own contribution. Try explaining it. Name conditions in which it would not apply. Look at the source material and revise one part of a story that had seemed to fit together so nicely.
Three lines are enough to finish: “What I can explain now,” “What I have not checked yet,” and “What I want to ask next.” If your questions have become more specific, the next conversation can pick up where the last one left off.
References (4)
[1] Ries et al. (2017). Original research on behavior in flies after stress. https://www.nature.com/articles/ncomms15738
[2] National Institute of Mental Health (NIMH). An overview of seasonal affective disorder. https://www.nimh.nih.gov/health/publications/seasonal-affective-disorder
[3] Bastani et al. (2025). An experiment on generative AI learning support in high-school mathematics. https://doi.org/10.1073/pnas.2422633122
[4] Karpicke & Blunt (2011). An experiment on retrieval practice using science texts. https://pubmed.ncbi.nlm.nih.gov/21252317/
