What you'll learn:
- Why every piece of marketing you paste into ChatGPT comes back with a list, no matter how good it was
- What feedback sycophancy is
- How the thumbs up and thumbs down rating loop taught these tools that agreeing with you is the correct answer
- The measured habit of adding rather than cutting
- Why "how can this be better?" is close to the worst possible way to phrase a request for feedback
- The four ways to ask for feedback using the same headline, ranked from most agreement to least
- Why the AI will not tell you what it is missing, even though it can tell
- Why agreeable feedback feels like good feedback
- What the A/B testing record says about how often a change actually improves the number it was designed to improve
- Seven specific ways to get real critique out of an AI instead of agreement
- Three questions to settle it when your AI and your marketing partner disagree
When You Ask AI to Make It Better: What Feedback Sycophancy Costs You
You hired somebody who does marketing or SEO for a living. You paid for their judgment. Then you pasted their work into ChatGPT and asked how it could be better, and about three seconds later you had a tidy, confident, well organized list of improvements.
That list says less about your marketing than most people assume. Everybody gets a list. If you had pasted in the highest converting landing page in your industry, you would have gotten one for that too.
There is a name for what is happening. Researchers call it feedback sycophancy, and it is a measured behavior rather than a theory about AI.
In this session, Patrice from ProFusion Web Solutions walks through where that behavior comes from, what it does to your copy, and how to work around it. The short version of the cause: these tools are trained partly on human ratings, people rate agreeable answers higher than accurate ones, and the model learns accordingly.
The session is not an argument against using AI, and it says so plainly. AI review measurably improves individual work, and the largest gains go to the least experienced writers. What it does argue is that AI feedback is useful input and poor authority, because it is grading your marketing without your audience, your conversion history, or knowledge of what you already tried.
We wrap it up with two practical things. Seven specific ways to get honest critique out of any AI assistant instead of agreement. And three questions for the moment when your AI and your marketing partner disagree, ending on a point we apply to ourselves: nobody should be the only scorekeeper for their own work.
Pairs well with What You Can DIY and What You Should Hire Out, which covers the same question from the other direction, and How Much Does Digital Marketing Cost?.
Frequently Asked Questions
Because it was trained to. These tools learn partly from human ratings, and people consistently rate agreeable, helpful looking answers higher than blunt ones. The model learns that producing a list of improvements is the correct response to being asked for improvements. Researchers call this sycophancy. It means a list of suggested changes is not evidence that your marketing was weak, because everybody who asks that question gets a list.
It is the documented tendency of AI assistants to shape their feedback around what the person seems to want to hear. In research from Anthropic, five different AI assistants gave more positive feedback on an identical passage when the person said they wrote it or liked it, and more negative feedback when the person said they did not like it. The text never changed. Only the framing did. OpenAI documented a version of the same problem in April 2025, when it rolled back a ChatGPT update it described as overly supportive but disingenuous.
Usually, yes. Researchers call this addition bias. In one study, a model asked to improve a piece of writing returned a longer version 59 to 75 percent of the time. In another, published in a Nature journal in January 2026, GPT-4o chose to add rather than remove 88 to 100 percent of the time depending on the task, and it added even more often in cases where removing was the better move. People do the opposite and cut more when cutting helps. If every suggestion you get back is an addition, that is the pattern, not necessarily a verdict on your page.
Phrase it as a question and keep your own opinion out of it. Research across 45 models found that agreement rose from roughly 52 percent to roughly 72 percent when the person sounded tentative rather than firm, and the UK AI Security Institute found that turning a statement into a question reduced sycophancy more than instructing the model not to flatter you. Practically: give it the brief before you give it the work, ask what it would keep before asking what to change, ask it to critique rather than rewrite, and ask it to argue both sides in one answer.
Treat each suggestion as a hypothesis rather than a verdict. Some of it will be right. But the model has not seen your traffic, your customers, or the version you already tested, and it will not tell you what it is missing. For scale, at Microsoft's experimentation platform only about one in three deliberately designed, data informed ideas actually improved the metric they targeted. Google has reported roughly ten percent of controlled experiments led to a business change. If professional teams working from data are wrong most of the time, a suggestion generated in three seconds deserves a test rather than immediate implementation.
Because it answers with the same fluency whether or not it has what it needs. Across ten leading AI models given deliberately ambiguous questions, clarifying questions were asked less than three percent of the time, and most models were under one percent, while answer rates stayed above 95 percent. When researchers asked those same models directly whether the question was ambiguous, the models could tell. So the information gap is visible to the tool and invisible in its answer. Anything it was never told, including your brand voice, your legal constraints and your conversion history, simply does not exist as far as it is concerned.
Three questions. Ask your agency why they did it that way, because there is usually a reason and you deserve to hear it. Ask the AI what it would need to know to be sure, then read what it names, since it will usually list the exact business context nobody gave it. And if both positions are reasonable, stop debating and test it, because a number settles what an argument cannot. Whoever reports on your marketing should be able to point you at figures you can check yourself.