The digital twin problem: more data isn’t enough

You’d think richer customer interviews would build a better digital twin. A new study suggests otherwise.

For a while, digital twins felt like the next big thing in research.

The idea is simple: build an AI copy of a customer from what you know about them, and test ideas on the AI instead of the person.

Many startups are selling exactly this promise.

The assumption behind it: the more you know about someone, the better their twin should be.

I recently came across a study from researchers at Columbia Business School and Wharton that tested this directly.

What they did

They interviewed 317 homeowners about replacing windows and doors using three approaches:

• AI interviews that asked follow-up questions
• Human interviews
• Fixed questions with no follow-ups

The researchers then built digital twins for people in the AI and fixed-question groups.

The twins were asked to predict how those people had actually responded to real marketing materials — including mailers, commercials, and marketing claims.

What happened?

The AI interviewer did well at collecting insight.

People talked more than they did with fixed questions. The AI also covered more of the planned topics, with similar depth to human interviews in most sections.

And because AI interviews cost much less, the same research budget surfaced more than twice as many customer needs as human interviews.

That’s pretty impressive.

There was one clear difference, though: people sounded more positive and emotionally engaged when talking to a human.

But the most interesting finding came next.

Twins that had a person’s interview did a little better than twins that only knew their age, income, and job.

But the richer AI interviews did not produce better digital twins.

Twins built from the detailed AI interviews were no better at predicting people’s responses than twins built from the simpler fixed-question interviews.

So more information helped researchers find more insights.

It didn’t necessarily help AI predict the person better.

There was another clue, in how people and their twins explained their reactions to the ads.

When explaining their reactions to ads, people often relied on immediate reactions and gut feelings.

The AI twins tended to explain things more analytically — especially when responding to images and video.

What I took away

This study made me think about three things that we often treat as the same:

Collecting human insight.
Understanding a human.
Predicting a human.

They aren’t the same problem.

AI may already be very good at helping us collect and organize much richer human insight at scale.

But having more information about someone doesn’t automatically mean AI understands them — or can accurately predict how they’ll respond to something new.

And I think that’s an important question to keep asking as digital twins become more common:

Does knowing more about a person actually make the prediction better?

This study suggests we shouldn’t assume that more is always better.

One caveat: this is one study in one category, a big, considered purchase. Results may differ for quick, emotional, everyday choices.

Study: “AI-Moderated Interviews for Market Research and Digital Twins Calibration” by Yuting Deng, Jingxuan Liu, Olivier Toubia, and Naman Jain (2026).

The digital twin problem: more data isn’t enough was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.