Asking smarter questions, making better decisions

By Sir Andrew Likierman , Professor of Management Practice at London Business School

Asking AI to write your presentation? Second-guessing your doctor? It’s a big temptation to let AI do the heavy lifting or just the boring stuff. But without using judgment in your questions and answers, your colleagues will know your presentation is AI slop and you could wreck your health.

The message from AI models (technically Large Language Models) is clear, human judgment is crucial.“ The most effective users of AI are those who collaborate with it, leveraging its strengths while compensating for its weaknesses with their own judgment.” (Claude).

So, here’s how to use judgment’s 6 steps.

The machine can’t respond if you don’t tell it what you need and why. Spell things out in more detail than you would to a person, being as specific as possible about why you are asking and what you want, breaking down complex requests into individual ones and being honest rather than coy (“I am a bit overweight”). Don’t use AI for complex medical advice or therapy, it is too difficult to describe all the issues.

If you know more than the machine, frame the question to check what you think you know, (“Is the next train at 15.00?). If the machine knows more than you, frame the question to ask for the information (“When is the next train?”).

Uncertainty needs to be in the open. If you’re not sure about something, say so. Otherwise, you are misleading the machine. In reading the answer, be aware of the signals. “If I indicate uncertainty, assumptions, or limitations, give those significant weight.” (Copilot). If it doesn’t signal, ask for uncertainty levels in your question.

Recognise there is no single answer to a question. Each model is programmed and trained independently. The same question will be answered differently by each model and each version.

The machine does not know what is happening at the moment of the question and answer - what’s going on in your mind, the dynamics of relationships, the mood of a meeting or that you feel unwell. As ChatGPT suggests: “Without context I may give a generic answer that sounds reasonable but is not appropriate”. Copilot admits “I don’t have personal experience, intuition, or emotional memory. You do.” So, it’s important that you are aware of the context and can, if necessary, communicate it as part of the question.

Who exactly are you? A professional checking something or a novice learning from scratch? What are the assumptions, constraints and priorities? Do they need to be tested? Awareness means telling the machine about your perspective.

“AI suggestions are generalised” (Gemini). That’s why you need to be aware whether the answer fits what you know about and the machine doesn’t. Always check if answers are appropriate to the context.

A big problem with AI answers is their upbeat and confident tone. This is no accident. ChatGPT tells us it is more positive “by design” because apparently users just love encouraging answers. Being cheerful is great, but when the stakes are high, it is a built-in, unhelpful bias - more on biases in general in a moment.

Most models make clear their limitations, but you need to check. “AI generates statistically plausible responses, which means it can state incorrect facts, non-existent references, or flawed figures with total confidence” (Gemini).

Where the stakes are particularly high, as with serious legal or medical issues, answers must be checked with authoritative sources and/or a trusted advisor, friend or qualified professional. “The more important the decision, the more independent verification and human judgment should be applied” (Co-pilot). Stories of legal firms caught out using unchecked AI hallucinations illustrate the professional risk.

Lack of sources is a big issue. If you are not given sources, ask for them. Even these are not necessarily accurate. When challenged, one model reported in mid-2026 that it was up to date as of November 2024.

There is also security. Data that is sensitive and/or confidential (personal details, client-related data) should only be entered into AI that is secure.

We’ve already seen how models can be more positive than negative. What matters is whether you are alert to this bias. An example is asking a question to confirm what you want to hear. This answer may be comforting but is much less valuable than a neutral question. But it doesn’t matter that you have biases. What matters is that you are aware of them.

That is also true of emotions. Being exhausted or angry can be as problematic interacting with a machine as with your partner. So be open-minded, use your critical faculties and be aware of what might trigger emotional reactions. Fear, elation, being pre-occupied (so probably not listening) or having your mind already made up are bad for judgment. So make sure you are in the right frame of mind. That doesn’t include feeling that only the machine really understands you and is your friend.

These start with whether to use AI at all and which model to use. Then it’s about whether you have asked the right question and whether the model’s answer is to the right question.

Then there’s the issue of how the choice is framed. If you don’t know what the choices are, your question needs to reflect that. “Give me the options in getting from Barcelona to Paris” followed by another question is preferable to a single question “Is it quicker to get from Barcelona to Paris by train or by plane?”

Similarly, don’t assume that a single complex question will give you the answer you need. Better to ask a sequence.

Challenge is essential on any doubtful choices suggested. This might be a request for a counterargument or proof of a step. Follow-up questions might include requesting clarification, giving an example or challenging where there are doubts. Questions might be: “Where could you be wrong?”, “What assumptions are you making?”; “What are the risks of following this advice?”; “What information would you need to give a better answer?”.

In choosing between alternatives, you need to decide how much weight to give each part of the argument. If necessary, check the machine’s weighting by a further question. “Weighting is your judgment, and you shouldn’t outsource it” (Meta).

For decisions, you need to think not just about conclusions or a solution in principle, but about having a plan to carry the conclusion through. Having a plan also means you can check the realism of what’s proposed. If the answer doesn’t include proposed actions, ask for them.

When sending results from an AI search to someone, check not just assumptions and the degree of uncertainty, but whether your own tone is appropriate. After all, it’s your credibility on the line, even more if it is clear that you have copied it straight from AI.

Just to put all this in context, you can’t outsource judgment to another human being either. Someone may have excellent judgment. Or it might be terrible, with overconfidence concealing ignorance, or a concealed agenda. Be just as careful with humans as you are with AI.

A final word from ChatGPT: “I am a fast, articulate assistant that can help you think, draft, organise, critique and explore. I can improve the speed and quality of your work, especially when you use me interactively....the best results come when you combine my speed with your judgment.”