A confidence crisis is emerging in the insights industry. 

Research professionals are rapidly adopting AI across the research process, from sampling to analysis, but as it becomes more deeply embedded in research, many are starting to wonder if the outputs are grounded in real consumer understanding, or if they simply sound convincing.

Fraud has always existed within online research, targeting high-paying or easy-to-access surveys to maximise rewards. Large language models (LLMs) have accelerated that problem by making it dramatically easier to generate responses that appear human and evade traditional detection methods.

For the market research industry, it’s a fluency trap. It used to be that researchers could flag low-quality responses based on tell-tale signals: gibberish, clear nonsense or unrelated feedback. 

Those telltale signals are changing shape. LLMs are capable of producing articulate, relevant and rounded feedback – complete with being emotionally in tune and having perfect grammar. Because they’re trained on content from the internet, they are in tune with common narratives and cultural references. At a glance, they not only pass the smell test; they feel illustrative. 

But they’re almost too perfect – and that’s the reason they pose such a massive threat. They don’t represent consumers – they parrot them, trained on more content about audiences than from audiences themselves. LLM responses smooth over the messiness, contradiction and unexpected language that make open responses from real people so valuable.

As consumer data increasingly becomes an input for AI-based analysis, prediction and synthetic research, the consequences of poor-quality data multiply.

Bad inputs create bigger risks as synthetic data scales

We are now facing a data ecosystem increasingly susceptible to invalid data. When poor-quality sample trains synthetic tools, flawed inputs become flawed models and flawed models become flawed insights. 

That view is echoed in the IAB’s 2025 report, The State of Data, which argues that the real challenge isn’t tools, it’s trustworthy inputs. The stakes are too high to treat data quality as a technical detail treated as secondary to a tool’s function. Those inputs, more than the latest wrapper on a generic LLM, will move the meter. 

That is why the winners in AI-supported insights will be the ones building the most trustworthy systems beneath them: strong methodology, broad and consistently collected training data, continuously refreshed consumer inputs and clear human oversight.

Stronger foundations matter more than flashy features

Rather than focusing on the latest agentic product rollout, buyers should be focusing on the foundations needed to protect quality and earn trust.

Strong systems require continuously refreshed training data, monitoring for drift from human responses, consistent data collection methods and established safeguards such as Captcha and GeoIP fingerprinting. As AI-generated fraud evolves, so too must the systems designed to detect it. 

However, technical safeguards alone are not enough; human oversight is paramount. Experienced researchers develop an instinct for quality – they know when a response simply doesn't feel right, even if it passes traditional checks. The opportunity is to scale that intuition through systems designed to detect changes in response patterns, monitor them over time and proactively flag when something looks wrong. Combining multiple quality signals allows those systems to continuously learn and adapt as new threats emerge.

There is also value in looking beyond the insights industry. Sectors such as finance have spent years developing increasingly sophisticated fraud detection because the consequences of failure are so high. The same mindset is needed in market research: quality systems that continuously evolve, drawing on a broad range of positive and negative signals to identify emerging risks before they undermine the data.

The fluency trap is dangerous precisely because it provides a false sense of quality. AI is already deeply embedded in the generation of insights, so we need to place greater scrutiny on the data and processes that underpin them. AI has enormous potential to make insights faster, richer and more scalable, but only if the foundations are strong enough to support that ambition.

This is the moment for the industry to raise the bar: to look beyond surface-level innovation and focus instead on the data, systems and safeguards that make confidence possible. Because in the end, the value of AI-driven insight will be defined not by how intelligent it appears, but by how readily it reflects real people.

Kim Malcolm is vice-president of product solutions at Zappi