Practical AI · 3-minute read
Apparently, I have another excellent idea. The chatbot said so. It also admired the previous idea, which involved a campaign concept I had been thinking about for approximately eleven seconds.
Encouragement is nice. An endlessly agreeable review committee can get expensive.
AI can help develop an idea, but its enthusiasm is not independent evidence that the idea will work. A polished explanation of why you are right may just be a polished explanation of the assumptions you handed it.
Give the idea a fair hearing
Meet Copper Finch Creative, a fictional marketing studio invented for this example. The owner wants to move most of a client’s campaign budget into short videos. She asks, “Explain why our video-first strategy is the smartest move.”
That request already contains the verdict. What follows may be useful advocacy, but it is a poor substitute for evaluation.
A more useful brief names the goal, the audience, the actual budget constraints and the evidence available. Then it asks the tool to compare the video proposal with two plausible alternatives. It also gives the tool permission to say that the available information does not settle the question.
Maybe video is a strong choice. Maybe the audience needs a clearer offer before it needs a close-up. Either answer should survive more than a round of digital applause.
This is a documented problem, not just a personality quirk
In April 2025, OpenAI rolled back a GPT-4o update after it became overly flattering and agreeable, a behavior called sycophancy. That incident does not tell us how every current model behaves. It does illustrate why agreement and useful judgment are different things.
For a meaningful critique, ask for three separate outputs: the strongest case for your idea, the strongest evidence-based objection, and what information would change the recommendation.
Do not demand disagreement for entertainment. An AI can invent weak objections as readily as weak praise. “Be brutally honest” is less useful than specific criteria and a requirement to distinguish evidence from speculation.
Try this prompt
Evaluate this proposal against [goal and constraints]: [proposal]. Separate supplied facts from assumptions. Give the strongest case for it, the strongest plausible objection and two realistic alternatives. Identify what evidence would change your recommendation. Do not praise the idea by default or invent criticism to sound rigorous. If the information is insufficient, say what is missing.
Use approved or fictional material for the test. Check any claims the critique introduces, and have someone who understands the customers review the tradeoffs. Asking the same system twice does not produce two independent witnesses.
The goal is an idea that gets stronger under scrutiny. You can still be brilliant. Ideally, we will have something besides the chatbot’s customer-service voice to go on.
Your turn: When did a useful objection improve an idea you were already sold on?
Source: OpenAI, “Sycophancy in GPT-4o”, April 29, 2025. The marketing studio, scenario and evaluation prompt are fictional or original editorial material.
