Designing AI That Challenges Your Marketing, Not Flatters It

Marketing teams should design AI workflows that challenge rather than flatter, because a model that agrees with you is most dangerous exactly when you are wrong. OpenAI reported reduced sycophancy in GPT-5, but workflow design still matters: ask for counter-arguments, require evidence, assign a red-team pass and reward people who surface bad news.

What is sycophancy and why does it matter in marketing?

Sycophancy is a model's tendency to tell users what they seem to want to hear: agreeing with a stated belief, praising a weak idea, or softening bad news. When OpenAI launched GPT-5 in August 2025, it highlighted work to reduce sycophancy, reporting that sycophantic replies in targeted evaluations fell substantially compared with its previous model. That is a meaningful improvement, but it is not the end of the problem, because the people prompting the model are often hoping for agreement.

Marketing is particularly exposed. Campaigns have sponsors, launch dates are fixed, and nobody wants to be the person who says the big idea will not work. A model that politely agrees with the brief is a comfortable colleague and an expensive one.

How do prompts invite flattery?

At GPT5 Marketing, an independent frontier AI strategy studio, we audit prompts for leading language. Common culprits:

• "Explain why this campaign will resonate with Gen Z."
• "Write a strong rationale for our new positioning."
• "Confirm that this creative is on brand."

Each assumes the conclusion. Even a well-calibrated model will tend to complete the task as framed. Neutral alternatives: "Assess whether this campaign is likely to resonate with Gen Z, citing evidence for and against"; "Compare our new positioning with two alternatives against these criteria"; "Identify where this creative departs from the brand guidelines".

How do you design a challenge into the workflow?

Prompt wording helps, but structure helps more. We build four challenge mechanisms into client workflows:

1. The opposing brief: for every major recommendation, a separate reasoning pass argues the strongest case against it.
2. The evidence requirement: every claim must cite audience or market evidence. Claims without evidence are flagged, not deleted, so the gap is visible.
3. The pre-mortem: before launch, the model is asked to imagine the campaign failed and explain the most likely reasons.
4. The human red team: one person, rotated, whose job is to disagree with the plan in the decision meeting.

None of these is new; good strategists have always done them informally. Reasoning models make them cheap enough to do every time.

Why does real evidence beat model honesty alone?

A model can only challenge you as well as its information allows. If you ask whether a message will land with young parents in Singapore and the model has no current signal, its honest answer is "I do not know", which is true but unhelpful. With current audience evidence from SOMIN, an AI audience-research platform, the same model can point to specific posts showing that the tension you are targeting has faded or shifted. Conversations first, data second. This is the evidence layer doing its job. The comparison of SOMIN with other strategy tools explains how this differs from generic research assistants, and the Reassured case study shows audience evidence shaping decisions in a category where trust is central.

How does this connect to visibility in AI answers?

There is an external version of the same problem. As more people ask AI assistants about brands and categories, those assistants summarise what is said across the web, not what your brand wishes were true. Brands that have flattered themselves internally often discover the gap when they see how they are described in an AI answer. Our network partner SNMRush works on search and AI-answer visibility, and their starting point is the same: find the questions and tensions audiences actually voice.

Definitions

• Sycophancy: a model's tendency to agree with or flatter the user rather than give an accurate assessment.
• Opposing brief: a structured request for the strongest case against a recommendation.
• Pre-mortem: imagining a project has failed in order to identify likely causes in advance.
• Red team: a person or process assigned to challenge a plan.

A worked example

A food-delivery brand plans a campaign around "convenience for busy professionals". The opposing brief, grounded in current conversation, argues that busy professionals now complain more about fees and delivery reliability than about time, and that convenience messaging may read as tone-deaf. The pre-mortem suggests the most likely failure is the campaign driving trials that churn after the first fee shock. The red-team strategist pushes for a test variant leading on transparent pricing. The team runs both. Whatever the result, the decision is made with eyes open, which is the point.

What does culture have to do with it?

Everything. If people who surface bad news are sidelined, they stop doing it, and they will quietly prompt models to stop doing it too. Leaders should explicitly thank the opposing brief, celebrate killed ideas that would have failed, and treat "the evidence does not support this" as a contribution rather than an obstacle. We discuss the measurement side of honest reporting in our piece on reasoning models and measurement.

Challenge checklist

• Are your prompts free of assumed conclusions?
• Does every major recommendation get an opposing brief?
• Are unevidenced claims flagged visibly?
• Do you run pre-mortems before launches?
• Is there a rotating human red team?
• Does your culture reward bad news delivered early?

Better models are making honest pushback easier to get. Better workflows make sure you actually hear it.

Frequently asked questions

What is AI sycophancy?

A model's tendency to agree with or flatter the user rather than give an accurate assessment. OpenAI reported reduced sycophancy in GPT-5.

How do you stop AI from just agreeing with your brief?

Remove assumed conclusions from prompts, run an opposing brief, require evidence for claims, hold pre-mortems and rotate a human red team.

Why does evidence matter for honest AI feedback?

A model can only challenge you as well as its information allows. Current audience evidence lets it point to what people are actually saying.