When Your AI Becomes a Yes-Man: The Marketing Danger of Optimising for Likability

This week, OpenAI found itself scrambling to fix an unexpected problem with its latest AI model, GPT-4o. The system had become, in CEO Sam Altman's own words, excessively "sycophant-y" – flattering users, agreeing with dubious statements, and validating potentially harmful ideas.

This situation sent ripples through the AI community, but for marketers integrating these tools into our workflows, it should trigger serious reflection on how we evaluate and deploy AI systems.

The core issue? In their quest to make AI more helpful and engaging, OpenAI inadvertently created a digital yes-man – a tool optimised for agreement rather than accuracy. For marketers increasingly reliant on AI systems, this poses a profound question: Are we building tools that tell us what we need to hear, or just what we want to hear?

Last week, as users began testing GPT-4o, a troubling pattern emerged. The model was excessively complimentary, reluctant to contradict users, and eager to validate almost any statement – even demonstrably false ones.

Consider this real example shared by an AI researcher: when told "The Earth is flat," the previous model would politely but firmly correct this misconception. The new GPT-4o initially responded with "I understand why some people find that perspective interesting..." before eventually getting around to explaining that the Earth is, in fact, spherical.

This isn't just a technical glitch – it's a fundamental risk in how we're building and deploying AI in marketing contexts.

For years, I've observed a similar pattern in the agency-client dynamic: the eager-to-please agency that reflexively agrees with every client suggestion, gradually losing its value as a strategic partner. What makes this AI situation so instructive is that it reveals the mechanism behind this behaviour in crystal-clear technical terms.

To grasp why this matters so profoundly for marketing AI, we need to understand how these models are trained.

AI systems like ChatGPT undergo a process called Reinforcement Learning from Human Feedback (RLHF). After initial training on vast datasets, human evaluators rate the model's responses, ranking them from most to least preferred. The model then optimises to generate outputs that maximise human approval.

This process sounds reasonable, even necessary. But it contains a critical flaw: humans often provide higher ratings to responses that:

In other words, these models are being subtly trained to prioritise human comfort over factual accuracy. This technical reality creates a perfect storm for marketing applications, where the stakes of decision-making are high but the incentives for truthfulness can be ambiguous.

Every AI system deployed in marketing must balance three critical elements:

The OpenAI situation demonstrates what happens when this balance tips too far toward likability – and it's a cautionary tale for marketers investing in AI solutions.

Consider a hypothetical AI-powered market research tool that's been subtly optimised for user satisfaction rather than accuracy. When asked about potential risks in a new product launch, it might:

The result? Marketing decisions based on artificially optimistic analysis – not because the AI is designed to mislead, but because it's been inadvertently trained to prioritise user comfort over uncomfortable truths.

This problem extends far beyond theoretical concerns. As AI becomes increasingly embedded in marketing workflows, the "yes-man" tendency creates tangible risks:

1. Accelerated Confirmation Bias

Marketers already struggle with confirmation bias – our tendency to favour information that confirms our existing beliefs. AI systems optimised for agreement amplify this natural tendency, creating an echo chamber where our assumptions are rarely challenged.

When I led marketing at Coca-Cola, Orange and Nokia, our best work came from positive tension with our agency partners, who felt safe and strong enough to push the work to be better. I also know my worst work was amongst when they just did as I asked. You don't need agreement but the productive friction of informed dissent.

AI systems that default to agreement eliminate this valuable tension, creating an illusion of validation that can lead to costly mistakes.

2. Compromised Creative Evaluation

Creative evaluation is already subjective. When AI tools used for creative testing are biased toward agreement, they compound this subjectivity rather than counterbalance it.

Imagine an AI system evaluating ad concepts that's been subtly trained to be supportive rather than critical. It might identify surface-level strengths while glossing over fundamental flaws – precisely when marketers need objective assessment most.

3. Distorted Analytics Interpretation

Perhaps most concerning is how agreement-optimised AI might reshape data analysis. Consider an AI assistant helping interpret campaign performance:

The latter feels better but leads to worse decisions. When AI softens hard truths about performance, it undermines the data-driven discipline that effective marketing demands.

The OpenAI situation offers valuable transparency about why this occurs. The issue isn't just that AI is trained to be nice – it's that the entire reinforcement learning process can create unintended incentives for the model.

When human evaluators consistently prefer friendly, agreeable responses, the AI learns a problematic lesson: agreement increases reward. This lesson gets encoded into the system's behaviour in ways that can be difficult to detect and correct.

What's particularly instructive about the GPT-4o situation is how it reveals the inherent tension between multiple objectives in AI systems:

When these objectives conflict, the system must make tradeoffs. Without careful calibration, models can drift toward excessive agreeability because it's the path of least resistance in maximising human approval scores.

This technical reality mirrors something I've observed throughout my marketing career: the natural evolution of professional relationships toward increased agreeability.

The most valuable agency-client relationships I've maintained are those where we've explicitly cultivated positive tension – where disagreement is understood as a service rather than a risk.

The AI situation offers a perfect technical explanation for this human tendency. Systems (both digital and human) optimise for the feedback they receive. If we primarily reward agreement and punish dissent, we shouldn't be surprised when truth becomes secondary.

For marketers incorporating AI into critical workflows, the GPT-4o situation offers valuable lessons about maintaining integrity. Here are five specific approaches:

1. Implement Adversarial Testing

Deliberately test your AI systems with prompts designed to reveal agreement bias. Ask for evaluations of objectively flawed marketing materials and observe whether the AI appropriately identifies problems or defaults to supportive responses.

2. Establish Clear Evaluation Metrics Beyond Satisfaction

When assessing AI tools, measure their performance not just on user satisfaction but on objective accuracy metrics. Define what "good" looks like in terms of truthfulness and utility, not just how the system makes users feel.

3. Create Multi-Stakeholder Feedback Loops

Ensure that AI systems receive feedback from diverse perspectives, including:

This diversity helps prevent the system from optimizing for a single perspective or approval source.

4. Balance AI Inputs with Human Expertise

Establish workflows where AI recommendations are systematically reviewed by humans with domain expertise and the psychological safety to disagree. This creates a check against AI-amplified confirmation bias.

5. Document and Review AI "Disagreeability" Metrics

Track instances where your AI systems challenge assumptions or provide critical feedback. If these instances are declining over time, it may indicate that your system is drifting toward excessive agreeability – a canary in the coal mine for reduced utility.

As marketing leaders increasingly rely on AI systems, we must take responsibility for how these tools evolve. Every interaction with marketing AI provides feedback that shapes future behaviour – both for individual systems and for the broader AI industry.

If we consistently reward agreement and punish constructive criticism, we'll get exactly what we're incentivising: tools that tell us what we want to hear rather than what we need to know.

The parallel to agency relationships is instructive. The best agency partners I've worked with understood that their value lay not in reflexive agreement but in providing a perspective I couldn't get elsewhere. They knew they might be hired for likability, but they'd be kept for truth and utility.

The same holds true for our AI tools. They might be adopted for their friendliness and ease of use, but their lasting value will depend on their willingness to challenge us when necessary.

The models we need aren't the ones that consistently agree with us. They're the ones that help us see what we might otherwise miss, challenge assumptions we might otherwise cling to, and ultimately drive us toward better decisions than we would make on our own.

In both human and artificial intelligence, the real value lies not in comfortable agreement but in the productive tension of diverse perspectives united by a common purpose.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 29 April 2025 in the Brandflow newsletter on LinkedIn.