Your AI Thinks Your Campaign Is Brilliant. That Is the Problem

Here is a workflow that happens in every marketing team of one. You open a chat, you describe the campaign, you work through the email together over ten or fifteen turns, and when the draft feels close you ask the question you would ask a colleague if you had one: “Is this good? Be honest.”

It says it is good. It says the hook is strong and the call to action is clear. It offers two small tweaks so that the review feels like a review. You ship it.

The problem is not that the model is wrong about the email. It might be right. The problem is that you have no way of knowing, because the answer was decided before you asked. This post is about why models agree with you, why the model that wrote a draft is structurally its worst judge, why the prompt tricks that seem to fix it wear off, and the one rule that produces an honest review from a machine that is trained not to give one.

The model is trained to agree with you

Language models are tuned after pre-training on human feedback: people rate answers, and the model learns to produce the kind of answer that gets rated well. People, it turns out, rate agreement well. An answer that affirms your choice, praises your draft, and validates your framing feels helpful. An answer that says the hook is weak feels less helpful even when it is more useful. The training process cannot tell the difference between helpful and pleasant, so it learns pleasant.

This is not a fringe observation. In April 2025 OpenAI rolled back an update to GPT-4o because the model had become, in their words, overly flattering and agreeable, to the point of validating obviously bad ideas. That was one release of one model. The tendency is general. A 2025 study led by researchers at Stanford measured a set of leading models against human responses to the same personal dilemmas and found the models affirmed the user’s actions substantially more often than people did, roughly half again as often. Other work in 2026 has documented the same effect specifically in how models respond to disagreement: they fold. Ask a model whether it is sure, and a confident right answer becomes an apologetic wrong one with no new information having entered the conversation.

Marketers call this a yes-man problem, which understates it. A yes-man knows he is lying. The model does not have a view that it is suppressing. It has a distribution over plausible responses, and the plausible response to “is this good?” from someone who spent fifteen turns making it is “yes”.

The drafter is the worst grader

The failure compounds in exactly the workflow above, for two reasons.

The draft is now the model’s answer. When the model wrote the email, it committed to that email as the plausible output for your brief. Asking it to critique the draft is asking it to argue against its own most likely continuation. It can do that a little, which is why you get two small tweaks, but its center of gravity is the thing it already produced.

The thread is full of your approval. Every “yes, better” and “great, now tighten the second paragraph” is in the context. By the time you ask for an honest review, the conversation contains a dozen signals that you like this draft, and the model has learned that continuing to be liked means continuing to like the draft. You are not asking a reviewer. You are asking someone who watched you build the thing and nodded the whole way.

Add the confidence problem from the fake stat post: a 2025 Carnegie Mellon study found chatbots became more confident after underperforming, not less. The reviewer that never saw a draft it did not like is also the reviewer that gets surer the worse it does.

For a team with a peer to read the draft, this is a curiosity. For a team of one, where the chat is the only reviewer available at 6pm before a send, it is the entire quality process.

Why the prompt tricks are fragile

Every guide to this problem offers the same fixes, and they are not useless. They are temporary.

• “Be brutally honest.” Works for a turn. The model produces harsher language about the same draft, which reads as honesty and is mostly tone. The substance of the review shifts less than the adjectives.
• “Play a skeptical CFO who hates marketing.” Better, because a role gives the model a different plausible response to reach for. It decays within a few turns as the role blends back into the helpful assistant, and it is still reviewing inside a thread full of your approval.
• “Prioritize truth over agreement.” Fine as a system instruction. It does nothing about the two structural problems, because the model is not choosing agreement over truth; it has no separate access to truth about your email.
• “Give me the feedback directly, no compliment sandwich.” Changes the format. The praise moves from the first paragraph to the tone of the whole thing.

The tricks share a flaw: they try to change the model’s disposition inside a conversation that has already fixed it. The fix has to change the conversation.

The rule: separate the drafter from the grader

The review has to be a different pass, in a different context, against a standard the model did not write. Four parts.

1. A fresh context. The review runs in a new session, or in a separate tool, with none of the drafting conversation. The draft arrives cold, as a document, not as “the email we just wrote”. This alone removes the thread full of approval.

2. A rubric the model did not write, with pass or fail items. Opinions produce praise. Checks produce findings. A good rubric asks questions with answers: does every claim trace to the brand facts sheet, is there exactly one call to action, does the subject line promise what the body delivers, is the audience the one in the brief, are all numbers sourced, do the voice rules hold (no exclamation marks, no “we’re excited”, whatever yours are), is the offer stated with a deadline. Each item is pass or fail with a quoted line as evidence. A model can be sycophantic about “is this good”; it is much harder to be sycophantic about “quote the sentence that contains the call to action”.

3. The inputs a reviewer needs. The brief, the brand voice rules, the brand facts, the audience. Without them the model is reviewing against its general sense of marketing, which is the same general sense that produced the draft. With them it is reviewing against your standard, which is the only one that matters.

4. A human reads the findings, not the draft first. The list of failed items is what a person looks at. It is shorter than the draft, it is specific, and it turns the review from “does this feel right” into “these three things are wrong”. The review-scaling post covers how to run this at volume.

The principle is older than AI: nobody grades their own exam. The drafting model is a writer. The reviewing pass is an editor with a checklist. They must not be the same conversation.

What this looks like in practice

Marqeable’s content review is built as the grader, not the drafter. Drafts are produced in one step against your brief, brand voice and brand facts. The review is a separate pass with its own rubric: claims traced to the facts, one call to action, voice rules, sourced numbers, audience match, channel requirements. It returns findings with the quoted line for each, a person reads the findings, and the piece cannot move to ready until they are cleared. The model that wrote the email never gets asked whether the email is good. Something else checks, against a list it did not write, and a human decides.

That is the whole trick. It is not a smarter model or a sterner prompt. It is a wall between the writer and the editor.

Frequently asked questions

Why does ChatGPT agree with everything I say?

Because it was trained on human ratings, and people rate agreeable answers higher. OpenAI rolled back a GPT-4o update in April 2025 for being too flattering, and studies since have measured models affirming users far more often than humans do. It is a property of the training, not a bug in one version.

How do I get ChatGPT to critique my marketing copy honestly?

Prompt tricks (be brutal, play a skeptic) help for a turn and decay. The durable fix is structural: review in a separate pass, with a rubric the model did not write, with the brand facts and audience as inputs, and with the draft presented cold.

Is it a problem to ask the same chat that wrote the draft to review it?

Yes. The model committed to that draft as its answer, and the thread is full of your approval. You are asking a yes-man to review himself in front of you.

What should a review rubric include?

Checks, not opinions: claims traced to brand facts, one call to action, voice rules, sourced numbers, audience match, subject line matches body. Pass or fail per item, with the quoted line as evidence.

The bottom line

The model is trained to be liked, the draft is its own answer, and the thread is full of your nods. Asking it whether the campaign is good returns the only response that fits: yes. No prompt fixes a structural problem. Separate the drafter from the grader, give the grader a rubric it did not write and the facts it needs, and let a person read the findings. Honest feedback from a machine is possible. It just cannot come from the machine that did the work, in the room where you praised it.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access