“My ex was a jerk. The next one was a jerk. The one after that? Also a jerk.” That’s four relationships, so four experiences. But if you file every one of them under the same label, your brain’s database may only have gained one useful column: jerk.
Here’s the catch.
Having lots of experiences is not the same as turning them into models you can reuse. ChatGPT works the same way. Hand over the whole answer and you may end up thinking less. But keep asking “Why?”, “What else could explain this?”, “What would prove me wrong?”, “What does the research say?” and “What would change my conclusion?”, and the AI stops doing your thinking for you. It becomes a sparring partner for your reasoning.
A 2026 systematic review that pulled together 67 empirical studies found the same split. When people used ChatGPT in a structured way, with questions and reflection built in, it tended to support higher-level thinking. When people just took whatever answer came back, shallow engagement and “letting the AI do the thinking” showed up more often.[1]
1. Does using AI really make us think less?
The honest answer: it depends on how you use it.
A 2025 study looked at 319 knowledge workers and 936 real examples of AI use. The more people trusted the AI, the less critical thinking they reported. The more they trusted their own judgment, the more they double-checked. Their critical thinking also moved toward setting goals, checking outputs, pulling information together, and making the final call.[2]
So this:
Ask AI → accept answer → done
is how thinking gets outsourced.
But this:
Something feels off
→ turn it into a question
→ form a hypothesis
→ ask AI for rival hypotheses
→ check research and primary sources
→ compare with other cases
→ fix the model
→ test it in real life
→ adjust again based on what happened
still leaves you with plenty to do.
You got AI to save effort, and somehow you’ve hired a peer-review committee inside your own head.
2. Why doesn’t “they were a jerk” teach you much?
“Jerk” feels good to say, but it’s far too blunt to predict anything.
Person A = jerk
Person B = jerk
Person C = jerk
Person D = jerk
Four experiences, and almost nothing left that tells one case from another.
Save what actually happened, step by step, instead:
Crosses a small boundary
→ watches how you react
→ stops after being told / keeps going
→ demands get smaller / get bigger
Shows intense affection
→ the relationship settles down
→ effort levels off naturally / vanishes suddenly
Gets called out on a problem
→ explains the situation
→ tries to make it right / pushes all the blame back
One of these behaviors doesn’t prove someone has a fixed bad character. Misunderstandings, tiredness, lack of skill, or a one-time mistake can look the same.
What’s worth noting is the conditions, whether it repeats, how they change after feedback, who ends up benefiting, and whether they repair things.
Don’t jump from “they did one bad thing” to “they’re a bad person.” Keep track of the mechanism.
3. Can how you compare experiences matter more than how many you have?
Sometimes, yes.
In a classic study by Gick and Holyoak, people who explained what two very different-looking problems had in common were better at carrying the underlying solution over to a new problem.[3]
The same showed up in negotiation training. Teams that compared two cases to find the shared structure were better at using a key strategy in a new negotiation than teams that studied the same cases separately.[4]
So rather than writing down:
Relationship A: awful
Relationship B: awful
Workplace C: awful
try this:
Relationship A: fuzzy boundaries let the demands grow
Workplace C: fuzzy role boundaries let the tasks pile up
Shared structure: when boundaries aren’t defined, the burden tends to land on one side
Different people, same wiring diagram.
4. What does asking “why?” and “what would disprove this?” actually train?
Explaining causes and connections to yourself is called self-explanation. A meta-analysis of 64 research reports and 69 effect sizes found a moderate learning benefit on average when people were prompted to explain.[5]
Another meta-analysis, covering 122 experiments and 10,382 participants, looked at practicing pulling knowledge out of memory. Compared with simply re-reading, it carried over better to new material on average, including application and inference questions.[6]
So questions like these do more than make a chat longer:
- What actually happened?
- Why did I read it that way?
- What else could be the cause?
- What’s the strongest counterexample?
- What new information would change my conclusion?
- Have I seen the same structure somewhere else?
They turn an experience into a structure you can search later.
5. How does that lead to the feeling of “I can see what’s coming next”?
Once you have enough abstract models, you don’t have to start from zero with every new situation.
New event
→ compare with models A, B and C
→ guess which stage you’re in
→ form a hypothesis about what comes next
→ watch, then update
That’s not fortune-telling. It just means you can come up with a first hypothesis from a library of patterns.
The big trap is hindsight. Once you know the result, saying “I knew it” costs nothing.
If you really want to test your forecasting, write it down before the result:
Forecast: X will happen within 2 weeks
Probability: 70%
Reason: A and B have already been seen
What would prove me wrong: if C happens, I revise the model
Outcome: happened / didn’t happen
Decision-science research on intelligence analysis also recommends tracking forecast accuracy and uncertainty with numbers.[7]
That’s how “I feel like I’m good at predicting” becomes something you can actually calibrate.
6. If ChatGPT is your outside brain, what should you keep for yourself?
AI is great at the heavy, repetitive parts: searching, suggesting options, comparing, arguing the other side, summarizing, and tidying up format.
What you should hold on to:
- What struck you as odd
- How you define the problem
- Which evidence you trust
- What would prove the model wrong
- Which model you finally go with
- What you’ll test in real life
- What you’ll throw out after seeing the result
Handy questions to ask:
- What’s the strongest argument against this hypothesis?
- Which other cause-and-effect story explains the same facts?
- If my starting assumption is wrong, where is it most likely wrong?
- What can I check in primary sources?
- Am I mixing up correlation and causation?
- What extra fact would change the conclusion?
- Are there similar cases in other fields?
- What would this model predict happens next?
If all you ever ask is “I’m right, aren’t I?”, the AI turns into a giant bobblehead.
Your outside brain needs to be good at proving you wrong, not at nodding.
7. Why does writing it up as an article make the loop even stronger?
A chat scrolls away. A written article keeps the structure.
Experience
→ question
→ research
→ counterexample
→ shared structure
→ practical rule
→ saved
→ reused on the next case
An article isn’t just something you publish. It works as a cache of thinking that was expensive the first time.
And the bigger your pile of articles gets, the more often you catch yourself thinking, “Wait, isn’t this the same structure as that other problem?”
It’s not only knowledge that grows. The links between pieces of knowledge grow too.
8. So what is “a good eye for people and situations” actually made of?
It isn’t simply how many people, jobs or crises you’ve been through.
A rough way to put it:
judgment
≈ number of cases
× detail pulled from each case
× comparison across cases
× attempts to prove yourself wrong
× checking against reality
Store four people as “jerks” and you may end up with jerk × 4 = one data point.
Pull out boundaries, blame-shifting, repair, whether words match actions, and how behavior changes under different conditions, and those same four experiences leave you with several models you can reuse.
ChatGPT’s real value isn’t just faster answers.
It’s that you can run this loop faster: come up with a hypothesis, attack it, rebuild it, carry it to another case, and check it against reality.
Don’t hand your thinking over to AI.
Just do more rounds of thinking.
Then AI isn’t a replacement brain. It’s the lab colleague who keeps tossing the ball back and never seems to go home.
