# How do you research a decision with AI without being misled?

Date: 2024-07-29 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/research-decision-ai-without-being-misled/

**Answer:** Use the AI to organize the question, list what you need to know and draft a comparison, then confirm every fact that matters in a primary source before you rely on it. Ask for facts and opinions in separate steps, so you always know which parts of the answer can be checked and which are only the model's judgment.

AI chatbots are excellent research assistants for the early stages of a decision. They can turn a vague worry into a list of questions, explain unfamiliar terms, and lay options side by side in seconds. The risk is that they deliver everything in the same fluent, confident tone, whether a detail is right, out of date or invented. OpenAI's own help pages say plainly that ChatGPT can make up facts, a failure usually called hallucination, and that users should check its answers.

So the goal is not to avoid AI, but to give it the right job. Let it structure your thinking and do the first sweep. Keep the final word for sources you can check and for your own priorities.

## Why fluent answers can mislead

Three habits of language models matter most when you are deciding something real.

1. **They fill gaps.** If the model doesn't know a price, a specification or a date, it may produce a plausible one rather than say so.
2. **They can invent sources.** A 2023 study in the journal *Cureus* asked ChatGPT-3.5 to write short medical papers with references; of 115 references, 47% did not exist and another 46% were real but cited inaccurately. A link or citation in an answer is something to check, not proof.
3. **They follow your framing.** Ask "Why is car A the best choice?" and you will get reasons for car A. Neutral questions get more balanced answers.

## A worked example: choosing between two used cars

Imagine you are deciding between two used compact SUVs. Start by asking for structure, not a verdict.

```prompt
I'm choosing between two used cars: a 2021 [model A] and a 2020 [model B].
My priorities, in order: reliability, total cost over five years, safety, cargo space.

Step 1. Do not recommend either car yet.
Make a table of facts I should compare, with these columns:
fact | car A | car B | where I can verify it | how sure you are (high, medium, low)

Write "unknown" if you are not sure of a value. Do not guess numbers.
Then list the questions I should ask the seller or dealer.
```

The reply gives you a working document. Here is part of what such a table might look like; the values are illustrative.

| Fact | Car A | Car B | Where to verify | Model's confidence |
|---|---|---|---|---|
| Cargo space, seats up | about 1,050 L | unknown | Manufacturer's spec sheet for that model year | medium |
| Recall history | two recalls | one recall | Government recall database, using the car's VIN | low |
| Fuel use, combined | 7.8 L/100 km | 8.4 L/100 km | Official government fuel ratings | medium |
| Asking price | not available | not available | Written quote from the seller | none |

Now do the checking yourself. Look up each row in the source named, and replace the model's value with the verified one, noting where you found it. Rows marked low or unknown deserve the most attention. A recall count, for example, must come from the official database for that exact vehicle, not from the model's memory.

For the money side, use guidance from a neutral authority. The U.S. Federal Trade Commission's car-financing advice is a good example: get the full out-the-door price in writing, including taxes and fees, and compare total cost rather than the monthly payment, since longer loans lower the payment but raise what you pay overall. You can then ask the AI to do the arithmetic with your verified numbers, and check the result with a calculator.

> **Try it:** Pick a purchase you are considering. Run the step 1 prompt, then verify just three rows of the table yourself. Note how many of the model's values were right, close or wrong. That ratio tells you how much to trust its next table.

## Ask for the opinion last

Only after the facts are verified, paste the corrected table back and ask for a judgment, with conditions:

```prompt
Here is my verified comparison table. Based only on this table and my priorities,
which car fits better, and why? Then argue the opposite case as strongly as you can.
Label every sentence that is your opinion rather than a fact from the table.
```

Asking for the strongest counter-argument is a cheap way to see what the first answer left out. If the case for the other car is surprisingly good, the decision is closer than it looked.

## How do you separate facts from the AI's opinion?

Do it in the structure of your prompts, not after the fact. Ask for facts first, in a table with a column for where each one can be verified, and forbid recommendations in that step. Ask for judgment in a second, separate message. When facts and opinions arrive in one flowing paragraph, it is very hard to untangle them.

Then apply a simple test to each sentence: could someone check this against a document? "The 2021 model has a 1.5-litre engine" is a factual claim; it is either right or wrong, and a spec sheet will tell you. "This is the smarter buy" is an opinion; no document can confirm it, and it depends on priorities the model only partly knows. Words like "best," "better value," "reliable" and "worth it" are signals of opinion, even when they sound like facts.

Finally, keep a record of what you verified and where. A short note beside each row, such as "manufacturer site, checked July 29," turns an AI draft into your own research. The model's opinion can still be useful as a second view, but by then it is one input among several, weighed by you.

## Sources

- ChatGPT General FAQ: https://help.openai.com/en/articles/6783457-chatgpt-general-faq
- High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content: https://doi.org/10.7759/cureus.39238
- Financing or Leasing a Car: https://consumer.ftc.gov/articles/financing-or-leasing-car