合成用户与数字孪生的定义、风险及适用场景：为何无法替代真实人类研究

If you ask an AI research tool what your users think, it'll give you an answer with total confidence. But what exactly is that answer based on?

"Synthetic users" and "digital twins" are often used interchangeably in our space, but it’s important to recognize the differences between these new tools.

To quickly distinguish them: Synthetic users are AI personas built from general training data. They’re useful for fast, directional feedback. 

Meanwhile, digital twins are AI models trained on one real person's data. They are meant to predict what that particular person would say. 

Most importantly, neither replaces real human research for consequential decisions. But, they can be used in very specific use cases. Let’s break it down further. 

What is a synthetic user?

A synthetic user is an AI-generated persona built from general training data, not from any specific real person. It's an LLM playing a fictional user type, shaped by broad patterns from its training. It's a composite, not a prediction about anyone in particular.

Synthetic users can be useful for early, directional feedback. For example, a quick reaction to a button label or a pressure test of a feature idea. Low-stakes situations where an extra test can help move things along. 

What is a digital twin?

A digital twin is a virtual representation of an actual person, built from that person's own data and often enhanced with AI models.

While a synthetic user represents a type of person, a digital twin claims to represent one actual individual.

What are the risks of using these tools?

The risk is right there in the names: "synthetic" and "twin." Neither one is the real thing, so neither behaves exactly like a real human does.

That gap shows up most in the context-dependent, sometimes irrational behavior that real human research is built to catch. People answer questions based on their mood, how tired they are, what's in their budget that week—all sorts of factors that shift from moment to moment. 

Research 101 is knowing that people don't always do what they say they will (New Year's resolutions are proof enough of that!). So how could a synthetic persona or a digital twin, trained on a snapshot of past data, ever fully capture that?

But the biggest risk isn't the tools themselves—it's when we treat their output as definitive instead of directional.

“Relying solely on simulations risks overlooking critical user concerns, such as issues of fairness, personalization, or trust.

Keeping humans in the evaluation loop is therefore both a methodological and an ethical imperative, ensuring that design decisions are accountable to real users and that agent-based insights are treated as directional rather than definitive.” — Sun et al., 2025

The issues surrounding digital twins

A September 2026 Columbia University study identified five systematic distortions in how digital twins behave:

1. Insufficient individuation: Twins compress people into patterns and miss what makes someone genuinely different.
2. Stereotyping: Twins lean on group-level traits over individual signal.
3. Representation bias: Twins perform better for educated, higher-income, ideologically moderate people, who are already overrepresented in research panels. The accuracy gap widens for underrepresented audiences.
4. Ideological bias: Twins reflect the political and social leanings baked into the underlying model.
5. Hyper-rationality: Twins reason too cleanly, and failed to reproduce two well-established findings in decision research (the attraction effect and the compromise effect), both of which depend on humans behaving a little irrationally.

The research even said: "Digital twins of today may best be described as funhouse mirrors that systematically distort human behavior."

The problem with synthetic users

NN/G found roughly 80% accuracy for synthetic users on attitudinal questions about familiar topics. But that figure doesn't extend to behavioral questions, unfamiliar contexts, or situations involving divided attention or unusual constraints. 

The more consequential the decision, the wider that accuracy gap gets.

So when are these tools useful?

Used cautiously, both tools can have a place in research. The challenge is knowing which tasks are actually a good fit for them.

When to consider using a synthetic user

It makes sense to use a synthetic user for early ideation and low-stakes, directional checks. Things like catching an obviously confusing button label, pressure-testing a feature idea before a real session is scheduled, or getting a quick gut check on wording before it goes anywhere near a real participant. 

It's a speed tool, not a robust research tool. It can help you narrow down options fast, but not confirm you've landed on the right one.

When to consider using digital twins

Use a digital twin narrowly: to fill a gap in existing data. 

Columbia's researchers found twins are better at capturing relative differences between participants than at predicting any one person's exact answer. So be sure to treat any twin output as directional, never definitive.

What should teams use instead?

Synthetic users and digital twins simply can't stand in for real behavioral research with real people, in real context. 

Even Outset and Dovetail (two platforms building digital twins) say publicly that twins are meant to extend human research, not replace it. Columbia's researchers frame their own goal the same way: building a transparent, responsible science of digital twin development, not disqualifying the technology.

That's the throughline worth carrying out of all this: AI can get you to a better question faster, but it still can't answer it for you. 

The job for any research team is figuring out where fast, directional AI feedback earns its keep—and where nothing but the real thing will do.