# Can ChatGPT help you make better decisions?

Date: 2023-07-03 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/chatgpt-help-make-better-decisions/

**Answer:** Yes, as a structured thinking partner: ChatGPT is good at laying out options, suggesting criteria you forgot, building a weighted decision matrix and arguing against your favorite choice. It cannot know your situation unless you tell it, it can sound sure when it is wrong, and it tends to lean toward what you seem to want, so the facts and the final call stay with you.

Most hard decisions are not hard because of missing information. They are hard because there are too many things to weigh at once, and our minds keep jumping between them. ChatGPT is useful here less as an oracle and more as a patient assistant that writes the problem down in an orderly way: what are the options, what matters, how much each thing matters, and how each option scores.

That structure is the real value. Once your reasoning is on the page, you can see it, question it and change it. The model can also suggest criteria you had not thought of, and it will argue the other side on request without getting tired or offended.

## The tool: a weighted decision matrix

A decision matrix is a table with your options in rows and your criteria in columns. Each criterion gets a **weight**, a percentage showing how much it matters to you, and each option gets a score on each criterion, say from 1 (poor) to 5 (excellent). Multiply each score by its weight, add across the row, and you get a total you can compare.

The numbers do not make the decision for you. They force you to state your priorities openly, and they show how close or far apart the options really are.

## A worked example: where should a small bakery spend its budget?

Imagine the owner of a two-person bakery with a sum set aside for one improvement. The choices: a second oven, a website with online ordering, or a part-time helper for the busy mornings. Here is a prompt that sets a role, gives context and asks for the work in steps.

```prompt
Act as a practical small-business advisor. I own a two-person bakery and can afford one of three improvements this year: (A) a second oven, (B) a website with online pre-orders, (C) a part-time helper for weekday mornings.

My situation: we sell out by 11 a.m. most days, I work 60-hour weeks, and about a third of customers ask whether they can order ahead.

Before scoring anything, ask me up to five questions about facts you would need. Then propose 3 or 4 criteria with weights that add up to 100%, explain each weight in one line, score each option from 1 to 5, and show the weighted totals in a table. Finally, list what would have to be true for the lowest-scoring option to be the best choice.
```

The "ask me questions first" line is important. Left alone, ChatGPT usually guesses what an unclear request means rather than asking, as OpenAI itself notes. Inviting questions turns that guess into a short interview.

After the owner answers, the matrix might come back like this (illustrative scores):

| Option | Extra sales (40%) | Owner time saved (35%) | Risk if it fails (25%) | Weighted total |
|---|---|---|---|---|
| A. Second oven | 3 | 2 | 4 | 2.90 |
| B. Online pre-orders | 4 | 3 | 2 | 3.15 |
| C. Part-time helper | 2 | 5 | 3 | 3.30 |

For risk, a higher score means a safer choice. The totals are close, which is itself useful information: this is a judgment call about priorities, not a clear win. Now push on it. Ask, "If I lower the weight on owner time to 20% and raise extra sales to 55%, does the ranking change?" With those weights the online option moves to the top (3.30, against 3.05 for the oven and 2.85 for the helper), and the owner learns exactly which value is driving the answer.

> **Try it:** Take a real choice you are facing, such as two job offers or three phones. Use the prompt above with your own options and context, answer the model's questions honestly, then change one weight by 15 points and see whether the winner changes. If it does, that weight is the conversation you need to have with yourself.

## Other moves that work

- **Pre-mortem.** "Imagine it is a year from now and option C failed. Write the three most likely reasons." This brings up risks that optimism hides.
- **Steelman the alternative.** "Make the strongest honest case for option A." A good decision survives its best counterargument.
- **Pros and cons with consequences.** Ask for pros and cons, then ask which ones are reversible and which are not. Reversible mistakes are cheaper.

## Where does it go wrong?

The first problem is **missing context**. The model knows only what is in the conversation, plus general patterns it learned in training. It does not know your cash flow, your local market or how tired you are unless you say so, and it fills gaps with typical assumptions. Its scores are only as good as the facts you give it, so check the numbers it relies on before you trust the table.

The second is **overconfidence**. OpenAI warns that ChatGPT can write answers that sound reasonable but are wrong, and that a small change in wording can change the answer. A neat table with two decimal places looks precise, but the scores underneath are guesses. Ask the same question in a new chat with different wording; if the ranking flips, treat the result as a rough sketch.

The third is **sycophancy**, the habit of telling people what they seem to want to hear. Anthropic researchers reported in December that larger language models are more likely to repeat back a user's preferred answer. If your prompt says "I'm leaning toward the oven," expect the analysis to lean there too. Keep your preference out of the first prompt, and ask for the case against your favorite before you commit.

Used this way, ChatGPT does not decide for you. It makes your own reasoning visible, which is often the step that was missing.

## Sources

- A Beginner's Guide to ChatGPT: https://www.amazon.com/dp/B0BZRRPXZW
- ChatGPT: Optimizing Language Models for Dialogue: https://openai.com/blog/chatgpt/
- Discovering Language Model Behaviors with Model-Written Evaluations: https://arxiv.org/abs/2212.09251v1