A prompt doesn’t just give an AI a task. It gives it a point of view. And a point of view comes with blind spots.
One file, two prompts
For the final project of my Data Analyst training, I had the same analysis audited by two AI models. Same data, same documents. One thing changed: how the prompt framed the work.
- To the first one: “let’s revisit a project you worked on.”
- To the other: “here’s a student’s notebook, audit it.”
Result: 32% of findings in common. On the same file.
And it mattered: the audit caught a stock coverage figure reported as 35.6 months that was actually 3. A number already presented to the client.
The question already carries a slant
Telling a model it’s reviewing its own work puts it in a position where validating becomes the natural outcome. It isn’t lying, it isn’t flattering: it answers the question as asked, and the question already carried a slant. That’s why “what’s wrong with my analysis?” and “confirm my analysis is sound” aren’t the same request. The first opens, the second closes. And we write the second far more often than we think.
It’s The Mom Test, applied to AI. Ask your mother if your business idea is good and she’ll say yes… not because she’s lying, but because the question invited it. The fix is the same with a model: stop asking for validation, ask for facts you can check.
Three habits I’ve adopted since
- Write the prompt that explicitly allows the work to be torn apart, especially when the work is mine.
- Change the posture, not just the question, and compare what surfaces.
- Never keep a finding without checking it against the data. Both models got something wrong at least once, and each time the data settled it.
Because whatever an AI tells you, the answer is always in the data. If a finding can’t be backed by a number, it isn’t a finding. It’s an opinion, or worse, a guess.
An AI doesn’t confirm your biases because it’s eager to please. It confirms them because you put them in the question. Whoever writes the prompt decides what the model is able to see.
What’s the last prompt you wrote that was secretly asking for a yes?