The industry is having a pretty unhelpful debate right now.
Synthetic vs. real. AI respondents vs. human panels. Everyone’s picking a side like it’s a zero-sum game and acting like they’ve cracked something. After sitting through a year of panels and pitches, it’s pretty clear this is mostly noise. The binary framing is lazy—and it’s pushing people toward bad decisions in both directions.
Here’s where I land:
Synthetic data is a real tool. A useful one. But it’s not magic, and it’s not a threat. It’s about knowing where it belongs—and where it doesn’t. That’s it.
Before even getting there, though, one thing that gets lost: “synthetic data” means completely different things depending on who’s talking.
- Synthetic personas
- Simulated conversations
- Models trained on large-scale human response data
Those are not the same thing. Not even close. They come with very different levels of rigor, but people bundle them together anyway. That’s how buyers end up with noise when they needed signal. Step one should be agreeing on what we’re actually talking about.
So where does it make sense?
Early-stage work, mostly.
If you’ve got ten concept variations and need to narrow to two, running full panels on all ten is slow and expensive. Using synthetic to pressure-test before committing real budget is just smart. It can also make sense for hard-to-reach audiences—rare disease patients, highly niche B2B buyers, ultra-high-net-worth consumers—where panel economics break down. And in privacy-constrained environments, which will only get more important as regulation tightens.
Where it doesn’t belong—and I’ll be direct here—is anywhere near decisions that really matter.
Brand repositioning. Product launches. Anything where you need to understand how real people actually feel right now.
An LLM trained on internet text can’t tell you that. It tells you what the internet has said historically, smoothed into something that sounds like insight but isn’t. That distinction matters.
This is especially true for qualitative work—the hesitation, the contradiction, the moment in a conversation that changes how you interpret everything. You can’t synthesize that. Anyone saying otherwise is selling something.
There’s also a dependency most people gloss over: synthetic data is only as good as the real data underneath it.
If your human data foundation is weak, the synthetic output will be wrong in ways that are hard to spot—because it still sounds plausible. That’s actually more dangerous than data that’s obviously off.
The teams doing this well have moved past the “versus” debate entirely.
They use synthetic data at the front end to move faster and cheaper. Then they use real human data when the decision actually matters.
Both. In sequence.
It’s not a conservative take. It’s just being disciplined about methodology instead of getting swept up in a compelling vendor story.
The next few years will shake this out.
The organizations that will come out ahead are the ones using AI to make human data work harder—not the ones trying to quietly replace it and hoping no one notices.
They won’t hold up.
