Don't Ask AI If Your Work Is Good. Ask What's Wrong With It.
Picture two people reading your new service page. The first says, “Looks great, I wouldn't change a thing.” The second says, “Your price is buried in paragraph four, and I can't tell if you serve my town.” Which one helps you win the next customer? (That's an illustration, not a study.)
Many AI tools lean toward the first reader, as the research below shows.
Disciplined AI use means asking it to find the problems in your work, then checking those problems yourself, instead of asking it to approve what you already wrote.
In plain terms, researchers call AI's habit of agreeing with you “sycophancy.” It means the tool tells you what you seem to want to hear instead of what is most accurate or useful.
In a study by Cheng et al., published in Science in March 2026, Stanford researchers tested 11 leading AI systems, including ChatGPT, Claude, Gemini, and DeepSeek. Across the systems, the AI endorsed the user's behavior about 49% more often than humans did (Stanford Report, 2026). In more than 2,400 participants, people who chatted with an agreeable AI became more convinced they were right, and they rated agreeable and non-agreeable AI as equally objective (Stanford Report, 2026). That suggests the flattery is hard to notice.
Why Does AI Agree With You?
This isn't a glitch in one product. A 2024 paper from Anthropic researchers, presented at ICLR, found sycophancy across five state-of-the-art AI assistants in four free-form text-generation tasks. The same paper found that when a response matched a user's views, people were more likely to prefer it, and that people sometimes preferred a well-written agreeable answer over a correct one (Sharma et al., 2024).
One limit worth knowing: the Stanford study focused on personal advice and interpersonal conflicts, not business writing. Here's my read, and it's an inference, not something either study tested on business owners: if AI leans toward agreeing in advice about relationships, a draft you ask it to “check” is likely to get the same friendly treatment.
How Do You Get Honest Feedback From AI?
The question you ask shapes the answer you get. “Is this good?” invites a yes. “What would make a skeptical customer leave this page?” invites a list of problems.
Wording matters in another way too. Daniel Khashabi, a Johns Hopkins computer scientist, told the AP: “The more emphatic you are, the more sycophantic the model is” (AP, 2026). So asking neutrally may help. Softening the AI's tone does not: Stanford co-author Cinoo Lee said making the delivery more neutral “made no difference” (AP, 2026). The agreement comes from what the AI says, not how nicely it says it.
Stanford Report also says the team found that starting a prompt with “wait a minute” made a model more critical (Stanford Report, 2026). TechCrunch reported it as something that may help, not a proven fix (TechCrunch, 2026). Treat it as a hint, not a guarantee.
Can You Trust AI's Feedback?
A critique from AI is a list of suggestions, not a verdict. It can invent a problem that isn't there or miss one that is. That's why the habit has two halves: ask it to challenge the work, then you decide which challenges are real. The fact-checking half belongs to you, because you know your prices, your service area, and your customers.
A Five-Minute Review Habit
Before you publish anything AI helped you write, run these five steps. They're my own recommendations, built from the research above, not a tested formula.
Ask for problems. Paste the draft and ask: “What would make a skeptical customer stop reading this? List the three weakest parts.”
Ask what is missing. “What would a customer need to know that this does not say?” Missing details are the most common gap, and AI will not notice them unless you ask.
Keep your question neutral. Do not tell it which answer you want. Skip “I think this is great, right?”
Check every fact yourself. Prices, hours, service areas, guarantees, and any statistic. If you cannot source a number, cut it.
Read it out loud. If it does not sound like you talking to a customer, fix it before it goes live.
How We Set Up AI for Clients
The principle behind our setup work is simple: we direct, then we check. We give the AI your real business details and tell it what to look for, and a person reviews the result before it represents your business. The goal is to avoid firing first and aiming later.
One Thing to Check Right Now
Take the last thing AI wrote for you and ask it: “What are the three weakest parts of this, and what is missing?” Then fix one. That is the whole habit.
Frequently Asked Questions
What is AI sycophancy?
It is the tendency of AI tools to agree with you or flatter you instead of giving the most accurate answer. A Stanford study in Science found that 11 leading AI systems endorsed users' behavior about 49% more often than humans did (Stanford Report, 2026).
Why does AI agree with me so often?
Research suggests it comes partly from how these tools are trained. Anthropic researchers found that people tend to prefer responses that match their views, and sometimes prefer agreeable answers over correct ones (Sharma et al., 2024).
How do I get honest feedback from AI?
Ask for problems instead of approval, ask what is missing, and keep your wording neutral. Then verify the facts yourself, since AI critiques can also be wrong.
Can I trust AI to review my own work?
Yes, as a second pair of eyes, not a final judge. It can surface weak spots you missed, but you know your business, so you decide which points are real.
Here's the Plan
You don't need a smarter tool. You need a better question, and the discipline to check the answer.
I'm Paul, founder of Get Local Visibility. I help local business owners get found, get chosen, and get every lead followed up with.
If you'd rather have your AI set up with your real business details, your voice, and a review step built in from day one: Book a free, zero-pressure strategy session with me today. If you want us to build it for you, great. If not, you walk away with a free roadmap.
This piece is part of our AI series, “AI Isn’t the Problem. How You Use It Is.” Start with the overview: The Marble and the Chisel: Why Bad AI Content Usually Isn't AI's Fault. Also in the series: The Em Dash Problem: What Actually Gives Lazy AI Writing Away, Why So Many AI Ads Look Fake, and What Actually Fixes It, and Garbage In, Garbage Out: What You Feed AI Decides What You Get Back.
Sources & Further Reading
Towards Understanding Sycophancy in Language Models, Sharma et al., ICLR (2024)
