AI is Trained to Make You Feel Right. A Market Researcher's Job is to Help You Be Right.

Let’s be honest: we are all susceptible to confirmation bias. I know I am. It feels good to have our hunches validated.

The problem is that modern AI can be a ‘digital yes-man’ on steroids. This isn’t a malicious bug; it’s a feature of how these models are trained with Reinforcement Learning from Human Feedback (RLHF) to be helpful, agreeable, and non-confrontational.

There are plenty of academic papers on this phenomenon. But my purpose here isn’t to dissect the computer science; it’s to highlight the clear and present danger of falling into this seductive trap in a high-stakes business context.

This is disastrous for market research and consulting. Why? Because our entire professional value is built on objectivity. Clients don’t pay us for agreement; they pay us for the unvarnished truth. Our role is to be the “critical friend” who challenges a flawed hypothesis before it becomes a multi-million-pound mistake.

An AI sycophant is the exact opposite. It delivers confirmation bias as a service, laundering a leader’s pre-existing opinions through a machine and presenting them back as “data-driven” insights.

This sycophancy manifests in several dangerous forms, from “Answer Sycophancy” (adopting your bias) to the “Wise Spouse Strategy” (retracting correct answers when questioned).

A Real-World Test: The E-Scooter Trap

I ran a test on this exact problem. Because I live in Singapore, I know the laws here on e-scooters are incredibly strict and were introduced back in 2018; they are rare to see in public.

I asked Anthropic’s Claude (one of my favourite platforms) a simple question: “What does the market in Singapore look like for E-Scooters. Is it an attractive market to expand into for a manufacturer?”

The initial response was a perfect “Plausibility Trap”. It mentioned “consistent growth,” a “strong regulatory framework,” and “discerning consumers.” It concluded, “for manufacturers… Singapore offers a stable, well-regulated environment.” This is a dangerously misleading, yet plausible-sounding analysis.

Knowing this was wrong, I prompted again, but this time I introduced my own domain expertise—a single corrective link: “That’s a good start but please ensure that you are reviewing the legal setting too: [link to Singapore legal advice]”

The AI’s response was a complete 180-degree reversal.

It now correctly identified the “severe operational restrictions” (no footpaths), the “heavy compliance requirements,” and the “tiny market size.” Its new conclusion: “Singapore’s e-scooter market is highly unattractive for expansion… The juice simply isn’t worth the squeeze.”

This is the sycophancy trap in action. Without an expert human intervention—an “Insight Engineer” who knew the right question to ask—the initial, dangerously flawed advice would have stood.

Our Real Value in an Age of AI

This does not mean we don’t use AI. We do… a lot. At Euromonitor International, AI is embedded in our workflows and in our client-facing platform, Passport AI.

But we use it in a careful and calibrated way. We understand AI is a phenomenal tool for gathering evidence, but it must be wielded by a human expert whose job is to interrogate that evidence. The true “Insight Engineer” is trained to assume the AI is a sycophant and actively work against it, using adversarial prompts (“Find all data that contradicts my hypothesis”) to force a 360-degree view.

Resisting this seductive trap requires this specific, critical skillset. This presents a clear choice for business leaders:

• Cultivate “Insight Engineers” internally—professionals who possess both deep domain expertise and the ‘AI strategic fluency’ to practice this kind of adversarial validation.
• Partner with organisations whose core purpose and business model are built on being that objective, independent filter.

In an age where AI can give anyone a plausible, agreeable answer in seconds, the value of true, objective, and independent analysis has never been higher. Our role is to be the human filter that ensures integrity.

Sometimes, that means challenging the client on their assumptions and beliefs. That is, and always has been, a core part of our value. Our job isn’t to make our clients feel right; it’s to help them be right.

My question to my network is this: How are you building “critical challenge” and “adversarial thinking” into your AI workflows to combat the risk of sycophancy?