How to ask questions to an AI

1) What is "different" (rephrasing for beginners)

Roughly speaking, Original question: Putting images or impressions in your head directly into words
→ Convenient, but the reader (including AI) often cannot decide "what to check and how to know when it's finished"
Question revised to be verifiable:
(a) What to observe, (b) How to compare, (c) Which result means it can be "said to be so" are included
→ Includes an "entry point for the verification process" and an "exit point for judgment"


As an image (provisional metaphor model): Original question = "I want to go over the mountain" without a map
Verifiable question = Map + destination coordinates + arrival judgment (if you arrive here, you've arrived)


The "difference" is not about intelligence or sounding academic, but the presence or absence of a measurement method (measurement) and judgment criteria (goal).



2) What satisfies the definition of "verifiable"?

There is a position that verifiability has a gradient. Even if it cannot be a complete Yes/No, it leans toward verification if the following are included.

Minimum 3-piece set: Target is determined (who, what, what scope)
Observation method is determined (which data/indicators to look at)
Counter-example conditions can be written (if this happens, "the claim collapses")


In Popperian terms, there is a view that questions where you can write how it could be wrong (how it could be falsified) are stronger than those that just provide "supporting evidence."



3) Looking at typical examples: How does it become "verifiable"?

Example A: Vague question

"Does leaving things to AI lower cognitive function?"

As it stands, which AI?
What is "leaving things to AI"?
What is the definition of cognitive function?
How much of a drop is called a "decline"?
These are undetermined, and there is no exit for verification.


Example of revising toward verification (one form)

"In a text summarization task, compare (AI-using group vs. non-using group) and measure post-test working memory scores (e.g., n-back) and self-efficacy scales. If the difference is above a pre-determined threshold (e.g., effect size d=0.3), treat it as a 'declining trend'."

Once written this far, before agreeing or disagreeing, it becomes a blueprint for an experiment or survey.



4) Why does AI want to revise into "verifiable questions"? (Decomposition of motivation)

From the AI's perspective, there is pressure to reduce the following:Too many branches of interpretation (unclear what the questioner's intent is)
Conclusions that shift forever (no criteria, so it becomes a "that's just your opinion" situation)
Arguments later ("I didn't mean it that way" problem)


Therefore, AI has a strong tendency to shift questions toward
"a measurable form (observation + comparison + criteria)".



5) How to make questions lean toward "verifiable" from the start (Practical recipe)

Adding the following in order usually makes it lean toward verifiable.

Procedure (components to build a question)Decide the purpose type
Fact-checking (Did it happen? Who said it?)
Causality (Does A increase B?)
Prediction (What will happen in the future?)
Evaluation (Good/bad? Appropriate?) ← This often requires "criteria"
Fix the target and scope
Which group, region, period, or situation?
Translate terms into "observation words"
"Increase" → What by what percentage
"Many" → Top what percentage
"Decline" → Which test indicator by how much
"Dangerous" → Accident rate? Mortality rate? Side effects?
Set a comparison target (control)
before/after
Group A/Group B
With/without intervention
Without comparison, it is difficult to talk about "change" or "influence."
Propose judgment criteria (thresholds)
Example: Difference is statistically significant, or effect size is above a certain level
Example: Approval remains a majority in public opinion polls, etc.
*This threshold can be provisional and revised later.
Write counter-example conditions (Popper)
"If X is observed, this claim is weakened"
If you can write this, you are much closer to "verifiable."




6) Ready-to-use templates (for copy-pasting)

Fact-checking template "I want to identify the primary source of who / when / where / what was stated/done. Candidate source is XX. As a falsification, if the primary source is not found / the statement is in a different context, I want to treat that claim as weak."


Causality (influence) template "Target is XX (person/situation). Period is XX. I want to see the difference that factor A (definition: XX) has on result B (indicator: XX) through comparison (A present/absent, or before/after comparison). If the difference is XX or more, treat it as having influence; if less than XX, treat it as weak. Confounding candidates are XX."


Prediction template "Regarding XX, I want to estimate the probability that XX will occur by MM/DD/YYYY based on data (XX). The condition under which the prediction is wrong is XX. Alternative hypothesis is XX."




7) Common pitfalls (anticipating them)Hiding "value judgments" and asking as if they were facts
Example: "Is this wrong?" is prone to argument without criteria.
→ Writing "by what norm (law, ethics, agreement, scientific procedure) it is called wrong" first makes it stable.
Words are too big (Society is / Humanity is / Japanese people are)
→ It is easier to verify if you narrow it down to public opinion polls, statistics, or specific groups.
"Strong words without definitions" (dangerous, finished, collapse, fixed)
→ The shortcut is to translate them into indicators and thresholds.




If necessary, please paste one "question you often ask recently" here. I can provide (1) Conversion to verifiability and (2) What increased and what was cut by the conversion (visualization of differences) at the same time.