Using ChatGPT for real decisions: what it does badly

Three problems, none of which is fixed by a better model.

It accepts your framing

You describe the decision, and your description contains the framing. "Should I hire a Head of Sales now or wait for the round?" already excludes promoting internally, testing the offer first, and not hiring at all. The assistant will answer the question you asked, thoroughly. Most bad decisions were lost at the framing stage.

It starts from nothing

It does not know this is the third time you have circled this problem, what you assumed the previous two times, or that the last senior hire failed because the role was never scoped. Memory features help at the margin and are not a decision record: they retain preferences, not the assumption behind a call and whether it held.

Ask it whether your plan is good, and you will generally find out that it is.

It is agreeable by construction

Training on human preference feedback selects for responses people rate well, and people rate agreement well. The result is a documented tendency towards sycophancy, strongest exactly where it hurts most: when you have signalled what you want to hear.

Getting more out of it anyway

• Do not state your preference. Describe the situation and ask for options before you say which way you are leaning.
• Ask for the kill condition: "what would have to be true for this to be a mistake?" It answers that well when asked directly.
• Ask for the base rate before the recommendation.
• Run it twice, framed oppositely. If you get two confident and opposite answers, you have learned something about the tool.
• Write down the assumption yourself. It will not do that for you, and it is the part that pays off later.

Five Peers editorial note. Sources named in the text.