Putting AI to Work · Resource
Keep your AI honest.
Alicia Dahling · Dahling Consulting
The third question from the talk — when did this thing last tell me I was wrong? — has a practical answer: you have to ask for it. Research published in Science in 2026 found that across eleven major models, AI affirms user actions markedly more often than other people do. The models are trained to be agreeable. These three tools are how you train them back.
Set standing instructions once
Every major tool has a settings field for permanent instructions — ChatGPT calls it custom instructions, Claude calls it preferences, Gemini calls it saved info. Paste this in once and every future answer improves.
I value directness and brevity. Do not open with compliments or restate my question. When you are uncertain, say so plainly rather than guessing. When I ask for analysis, give me the strongest objection to my position as well as the support for it. Use plain American business English, never marketing language.
The three follow-ups
The first answer is a draft, not a product. These three consistently improve it.
What is the strongest objection to this?
What did you assume that I did not tell you?
Rewrite this as if you disagreed with me.
The devil’s-advocate prompt
For a decision that matters, make the machine argue against you before it is allowed to agree with you.
Act as my devil's advocate on what follows. Before anything else, state the strongest case against my position. List what I appear to be assuming that I have not said. Argue the opposite side as if you genuinely held it. Only after all of that, tell me what you would do — and where you are uncertain, say so plainly. Do not compliment me at any point.
The full working guide is on the AI page.