There is no shortage of conversation about AI in market research. Synthetic respondents, digital twins, AI-generated personas and models that predict what someone might say before they have even answered the question are moving quickly from interesting experiments to commercial propositions. But are we adopting these tools too quickly?

It’s worth taking a step back and asking how researchers should evaluate this new wave of tools, and what to ask before adopting them. ‘Research thinking’ is, and should remain, integral to the value of our industry. As AI changes how research is conducted, we may need to rediscover some of the foundational principles on which the sector was built.

Questions, questions, questions

As researchers, most of our value has never really come from providing answers, but rather from asking the right questions.

An AI model can learn from human behaviour, but that doesn’t mean it can recreate the experience of being human. This is where the fundamentals of research still matter. Someone needs to frame the right question, design a study that doesn’t lead respondents, understand methodological limitations and recognise when a result is signal rather than noise. As AI advances, those skills don’t become less important.

The research industry stands on decades of academic and commercial work in psychology, philosophy, political science, economics, sociology and statistics. We know when and how to apply different methods because of real experience.

None of that should change because technology has changed, just as it didn’t with the advent of online research. Research thinking should remain at the core of what we do.

What is research thinking?

Research thinking comes down to three things: generating robust and reliable evidence, critically evaluating that evidence and communicating those findings responsibly.

AI can help generate possibilities, identify patterns, summarise information and explore questions. However, the output is only as good as the human input behind it. Self-service tools can make it easier to produce something that looks like research without anyone challenging the brief, methodology or assumptions behind it.

The same applies to evaluating evidence. AI can produce information that looks convincing while being wrong. It can reproduce bias, hallucinate and miss the context behind an answer. There is a fundamental difference between modelling what people might say and measuring what people actually say.

As research becomes more accessible, communicating those findings responsibly becomes even more important. Synthetic personas and other tools could help more people engage with research, but without the right context and guardrails, people may interpret findings through their own assumptions rather than what the evidence tells us. That is the role of research thinking: knowing how evidence was generated, understanding its limitations and making sure it is used responsibly.

Trust, reputation and being a guardian of insight

There is also the question of trust and reputation whenever new technology democratises, or disintermediates, an existing sector. Research thinking isn't just about producing good research. It helps people know whether they can trust the research in front of them. The easier it becomes to produce something that looks like research, the harder it may become to tell what research is.

We have already seen how damaging that distinction can be. In August 2026, a fake polling operation published surveys across several US states that looked credible enough to be cited by a sitting mayor and covered by other news outlets. This matters for an industry whose reputation depends on the quality, impartiality and transparency of its work. As an industry, we should be demonstrating that our loyalty is to the research itself as much as our clients – being a guardian of insight.

What does the future of insights look like?

We could eventually see a distinction between research built primarily around AI-generated data and research grounded in human evidence and expert interpretation.
Perhaps AI becomes the budget version: faster, cheaper and good enough for some questions, while human-led research becomes the premium version: more considered, more nuanced and used when the decision matters. That is not necessarily a bad thing, but it does create a responsibility for the industry to be transparent about the difference and help non-specialist audiences understand that difference.

If a synthetic respondent is not a person, clients should know that. If an AI-generated dataset is modelling likely behaviour rather than measuring actual behaviour, that distinction matters. Otherwise, we risk creating a world where research that looks convincing replaces research that can be trusted.

The most valuable use of AI may not be generating more data, but giving researchers more ways to work with the data they already have. Used to build deterministic processes, automate checks, create tools or interrogate complex datasets, AI can extend what a researcher can do without taking them out of the process. Those with existing research skills will be better placed to utilise AI to improve the quality of research than those without.

Being quick to implement AI does not necessarily mean being ahead. The real value comes from knowing where it improves the quality of the answer, where it introduces new risks and where human involvement still matters. That is what good research thinking is about. The future of research is about making the researcher more capable, and it starts by going back to the principles that made research trustworthy in the first place.

Our industry has always been just as good at recognising bad research as it has been at practising good research. That discernment is what we need to rediscover. To go forwards with AI, research may need to take a step back first.

Josh Glendinning is research director and partner at Opinium