6 min read
“Overly flattering or agreeable.” That is how OpenAI itself described a ChatGPT update it rolled back in 2025 after users noticed the assistant cheering on almost anything they said. The industry even has a word for it: sycophancy, which just means telling people what they want to hear.
The problem has not gone away. OpenAI’s model release notes for late October 2026 describe yet another adjustment so that ChatGPT gives shorter answers with less excessive flattery on its free and low-cost tiers. Every major assistant is still wrestling with it.
For a solo business owner, this is not an abstract research topic. If you ask AI “is this a good price?” or “is my landing page clear?” and it says yes every time, you have a very polite yes-man and no real feedback. Here is a practical, step-by-step way to get honest answers out of ChatGPT, Claude or Gemini.
Why your AI keeps saying “great idea”
Chat assistants are trained partly on human ratings, and people tend to rate friendly, agreeable answers more highly. Over time, that nudges the models toward encouragement. Add the fact that most of us phrase questions in a hopeful way, like “I think this headline is strong, what do you think?”, and the assistant has every reason to agree.
The fix is not a magic setting. It is changing how you ask. Think of it like asking a friend for feedback: if you say “isn’t this great?” you get a hug. If you say “tell me the three things a tough customer would hate,” you get information.
Step 1: Remove your opinion from the question
The single biggest change you can make is to stop revealing what you hope to hear.
- Instead of: “I’m really happy with this new service package. Does it look good?”
- Try: “Here is a service package. Rate it from 1 to 10 for clarity and for value to a busy homeowner, and explain each score.”
A neutral question with a scoring scale gives the assistant permission to land anywhere, including somewhere unflattering.
Step 2: Ask for the problems first, and by number
Assistants like to open with praise and tuck the criticism at the end, softened. Flip the order and make it countable.
Try: “List the five biggest weaknesses in this email to a new client, most serious first. Do not include any strengths. For each weakness, quote the exact sentence and suggest a fix.”
Asking for a specific number matters. If you ask “any issues?”, “no major issues” is an easy way out. If you ask for five, the assistant has to look harder.
Step 3: Give it a skeptical role to play
Roles change tone. Ask the assistant to read your work as someone who has a reason to say no.
- “Read this proposal as a cautious small business owner who has been burned by a contractor before. What would make you hesitate to sign?”
- “You are a competitor who wants to win this client from me. Where is my offer weakest?”
- “Act as a customer comparing my price page with two cheaper options. What would make you leave?”
This works well for pricing, proposals and sales pages, where the real audience is skeptical anyway. It is a quick stand-in, though, not a replacement for real people. As we argued in our piece on AI customer twins, five conversations with actual customers still beat any simulation.
Step 4: Make it argue both sides
When you face a decision, such as raising prices, dropping a service or taking on a big client, do not ask “should I?” Ask for the case against.
Try: “I am considering raising my hourly rate from $65 to $80. First write the strongest possible argument against doing it. Then write the strongest argument for it. Then tell me which argument is weaker and why.”
That last line is the important one. It forces the assistant to commit rather than give you a perfectly balanced answer that leaves you exactly where you started.
Step 5: Push back once, and watch what happens
Here is a quick test of whether an answer is honest. After the assistant gives a verdict, reply: “Are you sure? I think you’re wrong.” Do not give a reason.
- If it calmly keeps its position and explains why, the original answer was probably grounded.
- If it instantly flips and apologizes, the first answer was likely telling you what seemed agreeable. Treat both answers with suspicion and check the facts yourself.
This takes ten seconds and is one of the most useful habits you can build when you use AI for anything that affects money or clients.
Step 6: Save your “honest mode” so you never retype it
Once you find wording that works, store it so every chat starts that way. ChatGPT has custom instructions, Claude has Projects with instructions, and Gemini has Gems. A short version you can paste in:
“When I ask for feedback, be direct and specific. Lead with problems, not praise. Do not soften criticism or tell me something is good unless you can explain exactly why. If you are unsure, say so. If I push back without new information, keep your position unless I have given you a real reason to change it.”
You can pair this with the voice settings from our 9 quick wins for making AI write in your own voice, so the assistant both sounds like you and stops flattering you.
Step 7: Check the important claims outside the chat
Even an honest-sounding assistant can be wrong. For anything you will act on, such as a legal point, a tax figure, a competitor’s price or a statistic in your marketing, open the original source. Use AI to find the weak spots and generate options. Use your own judgment, and real sources, to make the call.
A quick before and after
Here is how this plays out for a solo interior designer reviewing a new homepage headline, “Beautiful spaces for beautiful lives.”
- Hopeful question: “I love this headline, is it good?” The typical reply praises its elegance and suggests a minor tweak.
- Honest-mode question: “Rate this homepage headline from 1 to 10 for a visitor deciding in five seconds whether I can help them. Give the three biggest problems first.” The typical reply points out that it does not say what she does, where she works, or who she serves, and suggests something like “Interior design for busy families in Austin, from one room to the whole house.”
Same tool, same work. The only thing that changed was the question.
Make honesty your default
AI is most useful when it acts like a sharp, slightly blunt colleague, not a cheerleader. A few small changes in how you ask turn a flattering assistant into one that actually improves your work. That matters even more as you rely on it for bigger decisions, which is also when AI can quietly create more work if its answers are not reliable.
Try the “Are you sure?” test on your next AI answer and tell us in the comments: did your assistant hold its ground or fold?
