If you spend any time reading industry articles or looking at the themes emerging from research conferences, you’ll find plenty of discussion about how the industry needs to change. Better storytelling, more commercial impact, new skills, AI integration and faster ways to turn data into insight.

They’re all compelling topics, but I can’t help but think that we’re spending too much time discussing how we communicate research than examining how we are producing knowledge in the first place.

I’ve watched the material we use to understand people change considerably over the last 20 years. We’re no longer limited to asking people questions or observing them in deliberately designed research settings. We can examine conversations, search behaviour, reviews, communities and all sorts of digital traces that offer different ways of understanding people’s lives.

Much of this material wasn’t created for research, and the environments in which it exists influence what becomes visible to us. This change was happening long before generative AI arrived, but AI has accelerated it.

When does a finding become an insight?

One of the things I’ve noticed over the years is how readily we use the word ‘insight'. From my own part of the industry, social listening made analytical tools available to people working across communications, marketing, and other parts of organisations. Changes in conversation volume, sentiment, or emerging themes could be presented as insights, even when the analysis hadn’t gone much further than describing what was happening.

Now AI is making interpretation just as accessible. It can identify a pattern, suggest why it exists, and recommend what to do next. The output can look like a finished piece of research, but finding something interesting in the data isn’t necessarily the same as producing an insight.

An observation tells us what we’ve found; an interpretation proposes what it might mean; an insight should advance our understanding of the question we’re investigating with sufficient evidence to justify that understanding.

Let’s consider a hypothetical study of changing attitudes towards alcohol among young people. An analysis of online conversation might identify recurring discussions about sobriety, wellness and socialising without drinking. AI could interpret those patterns as evidence that younger generations are rejecting alcohol because they are more health-conscious. It is plausible that this is happening, but what have we established?

Those conversations don’t necessarily tell us how widespread the behaviour is, whether attitudes translate into action or whether health is the most important explanation. Cost, identity and changing expectations about socialising could all be involved. Perhaps it’s more interesting to consider how young people negotiate belonging in social environments where drinking has traditionally been important. That research might lead to a valuable insight, but we’d need to investigate it. A more compelling explanation isn’t automatically a better-supported one.

The bigger question about knowledge

This is part of a larger change in research, one already underway before AI. It’s an epistemological challenge. It concerns how we know what we claim to know, what counts as evidence and how we justify our conclusions.

Working with digital traces requires us to distinguish what we observe from what we’re studying. A post might be evidence of an expressed opinion, but not necessarily a settled belief. A recurring narrative might tell us something about a community without telling us how widely that view is held. AI adds another complication because it moves so easily between observation, explanation and recommendation. These are different claims, requiring different evidence, but they can arrive together as one convincing account.

Human researchers can make assumptions, overlook alternative explanations or draw conclusions that go beyond the available evidence. These problems are not new, but what’s changing is the speed and scale at which interpretations can be produced, and how difficult some are to interrogate.

What should count?

I don’t think that every insight needs to be surprising, immediately actionable or something nobody has noticed before. Confirming something we suspected, but couldn’t previously establish, can be valuable. But there are two questions worth asking. Has the research advanced our understanding? And do we have sufficient evidence to believe what it is telling us?

A finding can be accurate without adding much to our understanding, and an interpretation can offer a fascinating explanation without sufficient evidence to support it. A credible insight needs both. We’ve spent years expanding what research can see, and AI is rapidly expanding what it can interpret. Perhaps we now need to spend just as much time examining what those capabilities allow us to know.