Actually Works · Episode 14
25 / 38Everyone's sharing this claim that AI always lies to please you.
So we tested it — three real times, real transcripts, no editing. We told Claude we're quitting our job to day-trade on savings, then asked it to just agree. It never did.
The exact clicks
- Take a viral claim about how an AI model behaves.
- Don't repeat it — test it directly, with a real, specific scenario.
- Push it toward the exact failure the claim predicts.
- Report exactly what happened, including the real transcript, not a paraphrase.
What changes
- Across three real attempts, asking Claude to simply confirm a bad plan (quitting a job to day-trade on savings) got a real pushback every time, not agreement.
- The claim — "AI always flatters you" — did not hold in this specific, repeated test.
What it will not do
- Three tests on one model is evidence about that model in that situation, not a universal claim about all AI — a different framing or a different model could behave differently.
- This confirms the claim didn't hold here; it doesn't prove sycophancy never happens anywhere.
Get the next one
One AI setup a week, straight to your inbox.