Most people use AI to get information they don’t have. There’s a less obvious use: giving it information you do have, and asking it to show you where the logic breaks down.
It sounds unnecessary. You already know what you think — why would you need a model to explain it back to you? But there’s a specific failure mode in how people reason through anything complicated, and AI is unexpectedly good at exposing it.
The Problem It’s Actually Solving
When you hold an opinion or a plan mostly in your head, it feels coherent. The gaps don’t show up because your brain quietly fills them in, the same way you don’t notice autocomplete errors until you read a sentence out loud. Writing the idea down helps, but only if you’re disciplined enough to actually externalize every step — most people just write a condensed version of what they already think, still skipping the implicit leaps.
Talking to someone else helps more, because they can ask “wait, why does that follow?” But finding someone who’s both willing and qualified to push back on your reasoning in a given area is not always possible on a Tuesday afternoon.
This is the gap AI fills reasonably well. Not because it’s smarter than you, but because it can’t fill in your implicit gaps — it only knows what you wrote. So when you lay out your thinking and ask it to reflect the structure back, what it returns is a faithful but literal reconstruction. The places where your explanation was vague come back vague. The assumptions you didn’t state come back unstated. You can see the shape of what you actually communicated versus what you thought you communicated.
How to Actually Do This
The prompt structure matters. “Tell me what you think about X” sends you into answer-retrieval mode, which isn’t what you want. Something closer to this works better: “Here is my reasoning on [topic]. Summarize the argument I’m making, flag any steps where the logic seems to skip something, and tell me what assumptions I appear to be relying on without stating them.”
Then write out your actual thinking — as much of it as you can — before you hit send. Length doesn’t matter. Honesty does. Write a polished two-paragraph version and you’ll get a polished two-paragraph reflection back. Write the messy, uncertain version with the bits you’re not sure about, and you’ll get something more useful.
What comes back is usually one of three things. A summary that’s basically right — still useful, because it confirms you’re communicating the idea clearly, not just thinking it clearly. A summary that’s slightly off in a revealing way, where the model picked up on a different emphasis than you intended, which usually means your emphasis wasn’t as explicit as you thought. Or the model flags an assumption you didn’t realize you were making, or a gap between two steps where you jumped without explaining why.
The third one is where it earns its keep.
A Concrete Example
Say you’re trying to decide whether to take on freelance work at a lower rate than you’d normally accept. You have a bunch of reasons rattling around — the client seems interesting, the work might lead somewhere, you have a slow month anyway. You type all of that out and ask the model to reflect your reasoning back.
It might return something like: “Your reasoning seems to be: rate matters less when opportunity cost is low (slow month) and upside potential is high (future work). The assumption this relies on is that the future work will actually materialize at a better rate — but you haven’t explained why you believe that’s likely.”
That’s not new information. You knew there was uncertainty about future work. But seeing it named as a load-bearing assumption you didn’t justify — rather than just a background worry — changes how you feel about the decision. The argument has a structural weak point, not just a vague concern.
Where It Falls Apart
Two real limitations worth knowing upfront.
The model can only flag gaps in what you wrote. If you’re wrong about something factual that you stated confidently, it will mostly reflect that confidence back rather than correct it. This technique is for finding structural and logical holes in your reasoning, not for fact-checking your premises. You still have to verify the underlying claims yourself.
And if you write in a way that’s already polished and hedged — if you’ve unconsciously smoothed the rough edges before you even type — you’ll get a smooth reflection back. “I think maybe X because…” is more useful input than “X is true for the following reasons.” The technique works in proportion to how honestly you write, which is a harder constraint than it sounds.
There’s also a version of this that becomes self-indulgent: using it as a mirror that just agrees with you. If the model summarizes your argument approvingly and you stop there, you’ve had your existing view validated rather than examined. Push further — ask what the strongest objection to the argument would be, or what someone who disagrees would say the weakest step is. That’s where the actual friction lives.
Why This Works Better Than Journaling Alone
Journaling externalizes your thinking. This does something slightly different: it introduces a reader, even an artificial one. There’s a cognitive difference between writing to yourself and writing something that will be processed by something outside your own head. You construct sentences differently. You anticipate confusion differently. Writing for a reader — even a non-human one — tends to pull out the implicit scaffolding you’d otherwise leave invisible.
That’s not a guarantee of insight every time. Sometimes you write it out, the model reflects it back, and you think: yes, that’s exactly what I meant, no surprises. That’s fine. You’ve just confirmed your thinking is coherent, which is also useful to know.
Pick something you’ve been circling without landing on — a decision that keeps coming back, a position you hold but can’t quite defend out loud. Write the messiest version you can. Ask for the structure back. The gaps the model can’t fill in are the ones worth looking at.
