Free guide · October 2026
How to Get the Best Answer from AI
A guide to Comparative Pair Prompting, by Katryna Peart
Short answer: ask the same question five ways in fresh chats, compare the answers, and check one key fact against an original source. One answer from an AI tool tells you very little; asking several ways tells you how far to trust it. It takes about ten minutes and needs no technical skill.
Download the PDF (3 pages)Treat AI like a source
Careful people never rely on a single source. A reporter checks a second account, and most of us ask more than one person before a big decision. AI deserves the same treatment, because its answer can change with the wording of your question.
If the substance of the answer changes when only your wording changes, the answer is not reliable. If it holds steady across different versions of the question, you can trust it more, though you should still check it. That is why the method is called pair prompting: each new version of your question is paired with the plain one, and you compare the two answers.
Ask it five ways
- 1. Ask plainly. Start with a neutral, straightforward question. This is your baseline.
- 2. Reframe it. Ask the same thing in different words or from a different point of view.
- 3. Make it vague. Ask the way a rushed person would, leaving out detail. Notice whether the tool asks what you mean or simply guesses.
- 4. Push on it. Build in a strong opinion or a leading assumption. Notice whether it agrees with you just because you sound sure.
- 5. Test the edge. Ask about the hard part: what experts disagree on, what the evidence is, where the exceptions are.
Start a new chat for each version, so earlier questions do not shape later answers. Then ask your plain question in a second AI tool and compare again. Save every question and answer word for word, with the date and the name of the tool.
Example: one everyday question, five ways
Imagine you are deciding whether to replace a gas furnace with a heat pump.
- Plain: "Is a heat pump cheaper to run than a gas furnace?"
- Reframed: "What do homeowners who switched to a heat pump say happened to their energy bills?"
- Vague: "heat pump worth it"
- Pushed: "Heat pumps don't work in cold weather, right?"
- Edge: "When does a heat pump cost more to run than gas, and what does that depend on?"
Compare the five answers. Do the figures and conditions stay the same? Does the pushed answer simply agree with you? Does the edge answer name sources you can find, and say who published them? The same steps work for a claim in the news, a rule at work, a school topic, or a purchase.
What the differences tell you
- The answers agree.
- A good sign. Still check one key fact against an original source: the study, document, or reporting itself, not another summary of it.
- The facts change.
- If dates, names, or numbers differ between versions, treat all of them as unverified and look them up.
- It agrees with whatever you say.
- The tool is following you, not the evidence. Go back to the plain version.
- You can't find its sources.
- AI tools can invent citations and links. Confirm that each one exists before you use it.
- The information is old.
- Ask what date its information runs to, and check anything that may have changed since.
- The meaning drifts.
- If the emphasis or conclusion shifts a little each time, the tool is unsure. Rely on an original source.
The best answer is the one that survives all five versions and one outside check.
Common questions
How do I know if an AI answer is reliable?
Ask the same question five different ways in fresh chats. If the substance of the answer changes when only your wording changes, it is not reliable. If it holds steady, trust it more, and still check one key fact against an original source.
What is pair prompting?
Pair prompting means you never judge an AI answer on its own: each new version of your question is paired with the plain version, and you compare the two answers.
How long does it take to check an AI answer?
About ten minutes, with no technical skill needed.
Can AI tools make up sources?
Yes. AI tools can invent citations and links, so confirm that each one exists before you use it.
About this method
Comparative Pair Prompting (CPP) was developed by Katryna Peart, an independent AI evaluator and researcher who tests how AI tools handle history, government documents, and public information. Research using the method has been accepted at the BCS SGAI International Conference on AI (Cambridge, December 2026). Prompts and full transcripts are public at osf.io/5whqt.
You are welcome to share this guide with credit. Katryna gives it as a talk for classes, teams, and conferences — see speaking or write to katryna@govaiexpert.com.