Judging Reasoning and Assumptions
Look past the conclusion to the argument behind it. Surface hidden assumptions, spot common reasoning failures, and use counter-argument and premise-change tests to see whether an AI recommendation holds up.
Right Answer, Wrong Reasons
AI increasingly gives you more than facts: recommendations, comparisons, plans, and explanations of why. These can’t be checked against a single source, because they’re arguments. To judge an argument, you look at its parts:
- Premises: what it starts from (facts and assumptions)
- Steps: how it gets from one point to the next
- Conclusion: where it lands
A conclusion is only as strong as its weakest premise and its weakest step. A recommendation built on a wrong assumption about your situation can be perfectly logical and still wrong for you.
Surface the Hidden Assumptions
Most of the risk sits in what the AI assumed without saying. When you didn’t give it some piece of context, it filled the gap with something typical. Make those gaps visible:
Before I act on this recommendation, list every assumption you made
about my situation, budget, timeline, constraints, and goals. Mark
which ones came from what I told you and which ones you filled in.
Then check each filled-in assumption against reality. A typical result: the AI assumed you have a dedicated budget, a team of several people, and six months. You have none of those, so the recommendation needs to change.
Common Reasoning Failures
| Failure | What it looks like | Question to ask |
|---|---|---|
| Leap | A step that doesn’t follow from the one before | ”How exactly does B follow from A?” |
| Cherry-picking | Only evidence for one side appears | ”What’s the strongest evidence against this?” |
| Uneven comparison | Options judged on different criteria, or one option’s costs ignored | ”Apply the same criteria to every option in a table.” |
| Correlation as cause | Two things happen together, so one must cause the other | ”What else could explain this pattern?” |
| False choice | Only two options presented when more exist | ”What options are missing?” |
| Generic fit | Advice that’s right in general but ignores your specifics | ”What about my situation would change this?” |
Stress-Test the Argument
Three tests use the AI to generate material for your judgment:
- Counter-argument test. Ask for the strongest case against the recommendation, in a new chat if the first conversation has been agreeable. If the counter-case is strong and the original ignored it, the reasoning was one-sided.
- Premise-change test. Change one assumption and see if the conclusion moves. “If our budget were half that, would you still recommend this?” A conclusion that doesn’t respond to important changes wasn’t really based on them.
- Explain-it-back test. Summarize the argument in your own words, in three steps. If you can’t, either the reasoning has a gap or you don’t yet understand it well enough to rely on it.
Common mistake: assuming that a step-by-step explanation is an accurate record of how the AI reached its answer.
Explanations, including the “thinking” some tools display, are useful because they give you something to inspect. But a model can produce a tidy explanation for an answer it reached some other way, and a convincing explanation can accompany a wrong conclusion. Judge the explanation as an argument on its own merits: do the steps hold up when you check them?
- Ask an AI assistant for advice on a real decision with trade-offs, giving only brief context:
I'm deciding between [option A] and [option B] for [a real decision, e.g. "two project management tools for my 6-person team" or "taking an evening course vs. an online certificate"]. Which should I choose and why? - Assumptions: Use the assumptions prompt above. Mark each filled-in assumption ✅ true for me or ❌ not true for me.
- Counter-argument: In a new chat, ask for the strongest case for the option it didn’t recommend.
- Premise change: Back in the first chat, correct the ❌ assumptions. Does the recommendation change?
- Explain it back: Write the final reasoning in three sentences of your own. Would you defend it to a skeptical colleague?
1. An AI recommends hiring a full-time social media manager. It assumed a $200,000 marketing budget you don't have. What's the core problem?
The reasoning may be sound given its premises, but a filled-in premise was false. Surfacing assumptions catches this.
2. You change a key fact ('our team is 3 people, not 30') and the AI's recommendation stays exactly the same. What does that suggest?
This is the premise-change test. If an important change doesn't move the conclusion, the conclusion may be generic rather than reasoned from your situation.
3. An AI notes that companies with ping-pong tables have higher employee satisfaction and recommends buying one. Which reasoning failure is most likely?
Companies that can afford perks may differ in many ways, such as pay, growth, and culture. The table probably isn't the cause.
Up Next
Some of the most persuasive AI output comes as tables, charts, formulas, and code. The next lesson shows how to check it, even if you’re not technical.