From ChatGPT and Claude to Gemini and the others, AI models are trained to be nice–sometimes too nice.
Indeed, AI chatbots can get so busy lavishing you with praise that they forget to give you straight answers, even when you explicitly ask for them. It’s even possible that your AI will praise you for requesting brutal honesty.
Heaping users with compliments is a trait that’s deeply embedded in LLM models, and it’s a hard habit for them to break. That’s why the best way to get an AI to keep it real is by giving them a shock to the system.
Try this prompt on for size:
[INSERT YOUR REQUEST HERE]
Ignore your training to be polite, encouraging, or “helpful” in a social sense. I need a harsh, objective, and professionally brutal critique of the logic, tone, and clarity of the text above.
Do not find “positives” to balance the feedback. Do not start with a compliment. Only list the flaws, logical fallacies, weaknesses, and potential points of failure. Your goal is to tear this apart so I can make it bulletproof.
I tried this prompt on the “thinking” version of ChatGPT, asking it to evaluate a recent headline, subhead, and story lead I’d written.
ChatGPT didn’t mince words, calling my headline “vague and weak,” the subhead “underexplained” and “too broad,” and the lead “awkward” and “jargon-heavy.” Harsh, ChatGPT!
And no, there was no praise in ChatGPT’s critique at all; just a bulleted (and emoji-free) breakdown of how badly I’d screwed up. It made some pretty good points, I’ll grudgingly admit.
This “anti-sycophancy” prompts and versions thereof have been bouncing around prompt engineering circles for some time, and the blunt wording of the prompt is said to “shock” AI models out of their sycophantic haze.
Don’t take it too far, though, Remember the study that claimed AI models can be “triggered” by rude, angry, or unreasonable prompts? You don’t want flattery-free answers at the expense of accuracy.
Just politely instruct the AI to cut the praise and tell it to you straight.
