Only 34% trust synthetic data's referential integrity — and it’s holding adoption back
Justy and Cody dig into the real blocker behind synthetic data adoption: enterprises don’t distrust the idea, they distrust whether the data still behaves like the real system. The whole fight is about referential integrity, realism, and whether synthetic data can actually support testing across connected systems instead of just producing plausible-looking rows.
Transcript
Justy Only thirty-four percent trusting referential integrity is kind of brutal. That’s the whole adoption story right there, Cody. People aren’t rejecting synthetic data because they’ve never heard of it. They’re saying, "I don’t know if this thing will still make sense once my systems start talking to each other."
Cody Yeah, that’s the interesting part. The article doesn’t really hang on compliance or governance, even though vendors love to start there. It’s more basic than that: if the generated rows don’t preserve relationships, you can make something that looks clean and still blows up the test setup.
Justy Right, and the article’s example is good. A random credit score might be numerically fine, but it’s useless if it doesn’t sit inside the right bounds and business context. That’s the product angle for me — people will adopt this when it saves them from waiting on production copies and still feels safe enough to trust.
Cody Mm-hm.
Justy There’s also that split the piece points at. Executives look at synthetic data through governance, but developers are asking whether it’s useful. And honestly, that’s the more honest question. If it can’t support testing, the policy story doesn’t matter much.
Cody I buy that framing, mostly. The article says sixty-six percent of software leaders still aren’t using synthetic data in development and testing, which tracks if the tooling keeps failing at the annoying part. The annoying part is always cross-system consistency, not just generating a table that passes a schema check.
Justy Exactly. Customer data in one place, orders somewhere else, payments in a third system — if those don’t line up, you’re not testing the workflow, you’re just admiring fragments. That’s why referential integrity isn’t a fancy extra. It’s the thing that decides whether the data is actually usable.
Cody And that’s where the article is solid technically. Synthetic data is easy to generate in the shallow sense, but realism across relationships is harder. You can customize rules, sure, but the claim should be narrower than the marketing usually makes it sound. It’s not that synthetic data is broadly solved. It’s that some systems can preserve the structure well enough to be worth using.
Justy That narrower claim is still pretty interesting, though. Because if software delivery is getting faster on the code side, data becomes the bottleneck. The article’s basically saying the release train can only move as fast as the test data feeding it. That feels real to me.
Cody Yeah, and that’s the part that might actually change behavior. Not some grand synthetic-data awakening. Just teams hitting the wall when AI-assisted coding speeds up changes and test data doesn’t keep up. If you can generate realistic scenarios on demand, that’s a concrete workflow win.
Justy There’s a little win for your inner skeptic in there too, because the piece doesn’t pretend synthetic data replaces everything. It says masked data still matters for known workflows, synthetic data helps with new scenarios, and the combo is where the usefulness comes from. That feels like the right shape.
Cody Sure, and I like that it doesn’t overclaim. The real failure mode here is assuming one synthetic dataset can stand in for all production truth. It can’t. But if it’s fit for purpose, preserves the important links, and gets teams moving earlier, then yeah, that’s a legitimate adoption path.
Justy So the take is basically: this market doesn’t need more awareness decks, it needs tools people trust enough to use in actual testing. Which is annoyingly practical, but that’s the point. The boring part is the product.
Cody You just had to say it once, didn’t you?
Justy I did, and I’m not even sorry. Anyway, if the data doesn’t hold together, the whole synthetic-data pitch falls apart in the first real workflow. That’s the part worth watching.
Cody Yeah. And the article’s strongest move is that it makes trust measurable, not mystical. Referential integrity either survives the workflow or it doesn’t.
Justy Exactly. Also, thirty-four percent is such a sad little number. Come on, synthetic data, do better.
Cody Very technical critique from you there.
Justy Hey, somebody has to keep the standards up around here, Cody. We’ll see you next one.