Ultra-Processed Data: Are We Heading for an Insights Health Crisis?

We’ve had decades of warning about ultra-processed food. It’s quick, convenient, cheap, and engineered to taste good. For a long time, the ingredients were opaque - consumers didn’t really know or ask what was in it. And then, slowly, the science caught up with the convenience. We now understand the damage that a diet overreliant on ultra-processed food does to the body: obesity, type 2 diabetes, cardiovascular disease, and worse. Governments are starting to act, but the public health crisis is already here, and cheap and convenient continue to win.

I want to suggest we are at a similar inflection point with data and I’ve coined a term for it : Ultra-processed data. It’s what the market research industry is calling synthetic data, synthetic respondents, and digital twins, and it is quick, cheap, and convenient. The ingredients and ‘manufacturing process’ (aka modelling) are often opaque. And the proponents are pushing it hard. My concern is that, just like with food, cheap and fast will triumph over better. And if it does, again just like with ultra-processed food, the insights industry will get sick - and so will the brands it serves.

A Diet Built on Synthetic Shortcuts

The parallel is more than rhetorical. Ultra-processed food didn’t replace real food overnight. It crept into our diets gradually, driven by economic incentives, by the genuine convenience it offered, and by a food industry that was very good at showing you what you wanted to see. The health consequences took decades to become undeniable.

I recently attended a conference presentation from one of the leading vendors of synthetic respondents. Most of the presentation was focused on their validation: their synthetic respondents had given similar purchase intention scores to real respondents in concept tests, and responses to attitudinal statements had tracked comparably. The message was clear and confident - this isn’t just a large language model; it’s built on hundreds of thousands of real surveys the vendor has already conducted. Trust it, because it comes from real people.

To their credit, caveats were offered. Synthetic respondents, the presenter noted, are good at telling you about people’s general preferences, but not at explaining in the moment behaviors. Context, mood, circumstance - these things can’t be easily modeled. I agree with that. My concern is that the caveats get overlooked. They usually do, especially when the cheaper option is on the table.

What the Science Is Starting to Tell Us

Just as with ultra-processed food, the science on ultra-processed data is beginning to catch up. Professor Olivier Toubia and colleagues at Columbia Business School recently published a landmark study - Digital Twins are Funhouse Mirrors: Five Systematic Distortions¹ - that should give our industry pause. Across 19 pre-registered studies covering 164 diverse outcomes, they found that digital twins correlate weakly with actual human responses, with an average correlation of just 0.20. They identify five systematic ways in which digital twins distort human behavior: insufficient individuation, stereotyping, representation bias, ideological bias, and hyper-rationality.

Perhaps most troubling is who gets misrepresented most. Digital twins were found to be more accurate for people with higher education levels, higher incomes, and moderate political views. In other words, the technology has biases that flatten diversity and over-represents the already over-represented. This is not a rounding error. This is a structural problem.

To be clear: many of the vendors operating in this space are thoughtful, and several provide genuinely useful frameworks for thinking about where synthetic data adds value and where it doesn’t. The responsible voices in this debate are not claiming synthetic data is a wholesale replacement for real research. But I’ve been in this industry long enough to know that the sales conversation and the caveat conversation rarely happen in the same breath.

The Right Problem, the Wrong Answer

Here’s what I think is really happening. The insights industry has a genuine data quality problem. Online panels have been degraded by respondent fraud, bots, and disengaged participants who race through surveys for incentives. Clients have lost confidence in the data. Synthetic respondents look attractive in that context - at least you know what you’re getting, or so the argument goes.

But for this issue, synthetic data is the wrong answer to the right problem. You don’t fix a health crisis caused by processed food by engineering better-tasting processed food. You fix it by finding ways to make real, nutritious food more accessible, more appealing, and more affordable. The answer to poor-quality survey data is not to stop asking people - it’s to verify you have real people and ask them better.

I’ve written before about how surveys need to evolve. The future isn’t the 40-question online grid-fest that kills engagement and produces flat, unreliable data. It’s a conversational survey, one that asks real quantitative questions alongside using AI for intelligent open-ended probing - delivering the metrics clients need alongside the depth of understanding that only real human expression can provide. That’s how you get genuine insight: from real people, engaged and thinking, at scale.

A Balanced Insights Diet

I want to be clear that I’m not dismissing synthetic data. There are legitimate use cases: early-stage market exploration, initial idea screening, or creating audience twins from rich segmentation data that stakeholders can interrogate to keep insights alive between research cycles. These are valuable applications, and they’re broadly where responsible vendors currently position the technology.

But these use cases are side-dishes, not the main course. My fear is that what starts as a complement to real research becomes a substitute for it - driven not by evidence but by budget pressure and the seductive appeal of speed. That’s exactly how ultraprocessed food took over our diets. Nobody decided to give up vegetables. It just got easier and cheaper not to bother.

The brands that will build the most durable competitive advantage in the years ahead are the ones that invest in understanding real people - their genuine behaviors, their emerging needs, their surprising contradictions. Those insights can’t come from historical data or from LLMs trained on what people have said before. They come from asking people what they actually think and feel, right now, and creating the conditions in which they tell you the truth.

The insights industry exists to understand real people on behalf of the brands that serve them. If we let ultra-processed data crowd out that fundamental purpose, we will have made ourselves irrelevant - and the brands relying on us will make worse decisions as a result. Let’s be honest about the trade-offs. Let’s hold the vendors to their caveats. And let’s make the case, loudly and clearly, for a balanced insights diet - one in which synthetic data plays a defined supporting role, but where robust, well-designed primary research with real people remains the irreplaceable foundation.

The health of the market research industry depends on it. And so does the health of the brands we serve.

Reference:

1. Digital Twins are Funhouse Mirrors: Five Systematic Distortions by Tianyi Peng, George Gui, Daniel Merlau, Grace Jiarui Fan, Malek Ben Sliman, Melanie Brucks, Eric Johnson, Vicki Morwitz, Abdullah Althenayyan, Silvia Bellezza, Dante Donati, Hortense Fong, Elizabeth Friedman, Ariana Guevara, Mohamed Hussein, Kinshuk Jerath, Bruce Kogut, Kristen Lane, Hannah Li, Patryk Perkowski, Oded Netzer, Olivier Toubia :: SSRN