Start By Making The Answer Checkable
The easiest time to catch a bad AI answer is before it gives you one. Instead of asking, "What should I do about this client email?" give the tool enough shape to produce something you can inspect: the task, the background, the limits, and the exact kind of answer you want. For example: "Help me reply to this client. They are upset about a late delivery. I want to sound calm, accept responsibility, and avoid offering a refund. Draft three versions under 120 words. Ask me any questions first if something important is missing."
That last sentence matters. "What do you need from me to give your best answer?" is a simple safety habit. It makes the AI show you what it does not know instead of guessing. The tradeoff is that it may slow you down by one extra message, but for important work that is cheaper than fixing a confident mistake later.
Separate Facts From Suggestions
Once the answer arrives, do not check everything the same way. Split it into two piles: facts and suggestions. Facts are claims that can be true or false, such as a deadline, a policy, a price, a statistic, or what a law or report says. Suggestions are judgment calls, such as the tone of an email, the order of a plan, or which option sounds more persuasive.
Ask the AI to label them: "Separate the factual claims in your answer from your recommendations. For each fact, say how I could verify it." This is useful because AI often blends real information with reasonable-sounding advice. If you are planning a trip, "the museum is closed on Mondays" needs checking. "Visit the museum before lunch because it gets crowded" is a suggestion you can weigh, not a fact you should blindly trust.
Use Sources And Comparison For High-Stakes Claims
For anything with consequences, make the AI show its work. Ask: "Which parts of this answer depend on outside facts? Give sources or tell me you are unsure." A source is only helpful if it actually supports the claim, so open it when the answer matters. Watch for vague source language like "experts say" or links that mention the topic but do not prove the specific point.
Then compare. Put the same question into two or three tools, especially when the answer affects money, health, legal choices, work decisions, or public claims. You are not looking for identical wording. You are looking for convergence, meaning the tools independently point to the same core answer. If ChatGPT, Claude, and a search-focused tool all agree on the main fact and cite similar evidence, confidence rises. If they disagree, slow down and check the original source yourself.
Bring In A Skeptic Before You Act
After you have a draft answer, ask for criticism instead of another rewrite. Try: "Act as a careful reviewer. What in this answer might be wrong, unsupported, incomplete, or too confident?" This creates a simple verifier, meaning a second pass whose job is to test the answer rather than improve the style. It is especially useful for research summaries, important emails, meeting plans, and recommendations you will share with others.
Keep the process proportional. You do not need three tools and a full critique to brainstorm dinner ideas. You do need more checking before sending a sensitive message to your manager or relying on a policy summary. The habit is not "distrust AI forever." It is "match the amount of checking to the cost of being wrong." When you make answers checkable, separate facts from advice, compare important claims, and invite criticism, AI becomes much more useful without asking you to become an expert.
Key takeaways
- Ask the AI what information it needs before answering when the result matters.
- Separate factual claims from recommendations before deciding what to trust.
- Require sources for outside facts, then open the sources for important claims.
- Compare high-stakes answers across multiple AI tools and look for agreement on the core facts.
- Use a skeptic prompt to find weak, unsupported, or overconfident parts before acting.
- Check more when the cost of being wrong is higher.