How to spot convincing nonsense in synthetic research

Synthetic research output looks like data even when it isn't. The believability is the trap. This is how you read a finding and tell signal from filler.

The hardest synthetic research to deal with isn't the obviously bad kind. It's the kind that looks great. Clean structure, confident quotes, tidy themes, a persona who answers every question in full sentences. It reads like data. Sometimes it is. Sometimes it's the model telling you a plausible story with nothing real underneath, and you can't tell which at a glance.

The researchers who reviewed 182 studies on synthetic participants gave this a name: misleading believability. The output is convincing on the surface and shallow underneath, and the believability is the problem, because it lends credibility to conclusions that haven't earned it.

So you need a way to read a synthetic finding and tell signal from filler. Six tells worth checking, and what to do about each. They work on any synthetic output, including ours.

1. Everyone agrees, and nobody contradicts themselves

Real audiences are messy. They disagree with each other and with themselves, they hedge, they change their mind mid-sentence, they hold two incompatible views at once. If every persona converges on the same answer and no one is internally inconsistent, you're probably looking at the model's average opinion dressed up as a population. Check for: disagreement across personas and contradiction within them. The absence of mess is a warning, not a mark of quality.

2. It tells you exactly what you expected

Synthetic output skews agreeable. Feed it your brief and it tends to hand the brief back to you, more articulate. Check for: surprise. If nothing in the findings caught you off guard, that isn't validation. It's an echo of your own framing.

3. It's too articulate

Real people ramble. They go off-topic, give half-answers, bring up things you didn't ask about, trail off. Output that's all clean, complete, well-organized sentences with no texture is suspiciously smooth. The review found synthetic responses tend to be verbose and over-structured while staying shallow beneath the polish. Check for: rough edges. Their absence is a tell.

4. The numbers are more precise than they could possibly be

"62% of buyers would switch," stated like it came off a survey. Synthetic research can tell you direction. It can't produce a real percentage with a margin of error, because there's no sample doing the math. Check for: false precision. If a number is presented like panel data but couldn't carry a confidence interval, treat it as a vibe with a decimal point.

5. You can't trace a single claim back to anything

Pick the most striking finding in the report and ask where it came from. If the only answer is "the model generated it," every finding is unfalsifiable, and you have no way to separate the grounded ones from the invented ones. Check for: provenance. A finding you can trace to a source is evidence. One you can't is a guess. This is the whole point of how evidence grounding works.

6. Invented findings that trace to nothing

This is the one to handle carefully, because it's easy to get backwards. LLMs do hallucinate. They invent needs, topics, and usability issues that were never in any real signal, and those sound exactly as authoritative as the genuine ones.

But the tell is not that a finding is new to you. Surfacing something you didn't think to ask about is one of the best things synthetic research does. You set out to test two hypotheses and the study hands back a third you never considered, and it's the most interesting thing in the report. That's not a bug to filter out. That's often the whole reason you ran it.

The difference between a discovery and a hallucination isn't novelty. It's whether it traces to anything. A surprising finding that links back to real evidence is a lead worth chasing. A surprising finding that links back to nothing is a guess in the same outfit. Check for: provenance on the unexpected findings specifically, because those are the ones you're most tempted to act on and most need to verify. Don't discount a finding for being new. Discount it for being unsupported. The ones that are both new and grounded are frequently the point.

Why grounded synthetic gives you something to check

Run that list against ungrounded persona output and most of it fails. There's no provenance to trace, no real variance so everyone agrees, no guard against invented findings. That's not an accident. It's what you get when the model is improvising from its training average.

Grounded synthetic is built to pass the same checklist. Provenance means you can trace any claim to its source. Calibrated personality and bias mean disagreement and contradiction show up by design rather than getting flattened out, which is the job of archetype clustering. Directional framing means the output doesn't dress itself up as a survey. And a critic pass exists specifically to catch drift and invented findings before they reach you. The fidelity comes from the grounding, not the model, which is the case I made in grounding beats the model.

Use it on us too

The reason to care about provenance is that it means you don't have to take the vendor's word for anything, including ours. You can run this checklist on a Candor report and trace the findings yourself. A tool that's confident enough to be inspected is the one worth trusting. A tool that asks you to trust the output because it looks good is asking for exactly the thing the 182-study review warned against.

The goal here isn't to distrust synthetic research. It's to read it like a professional instead of being impressed by it. Believability is the trap. The checklist is how you don't fall in. See how Candor works for how we try to pass it, and the evaluator's framework for synthetic research tools for sizing up a platform before you ever get to an output. What's the most convincing piece of research you later found out was hollow?

Common questions

Inspect the output against a few tells rather than trusting how polished it looks. Trustworthy synthetic research shows variance (personas disagree and contradict themselves like real people), surprises you rather than just confirming your brief, has texture instead of suspiciously smooth prose, stays directional instead of quoting false-precision percentages, and lets you trace any finding back to a source. The single most useful check is provenance: pick a striking finding and try to trace where it came from. If you can, it's evidence. If you can't, treat it as hypothesis-grade. The believability of the output tells you almost nothing on its own, which is exactly why surface polish is the trap.

Misleading believability is the tendency of synthetic output to look credible (clean, structured, confident, full of complete sentences) while being shallow, stereotyped, or fabricated underneath. The term comes from the largest systematic review of synthetic participants, which found that experts often couldn't distinguish synthetic personas from human-generated ones on a first read. It's dangerous because surface polish lends false credibility to conclusions that haven't earned it. A confident fiction is worse than an obvious one, because the obvious one gets caught and the confident one gets put in a deck and shipped. The defense is provenance and inspection, not impression.

Not statistically-bounded ones. Synthetic research is strong at direction and ranking (which option is stronger, where a value prop wobbles, which way a segment leans), but it can't produce a percentage with a margin of error, because there's no real-respondent sample doing the math. A synthetic figure presented like survey data is a warning sign, not a feature. A practical test: if a number is shown like panel data but couldn't carry a confidence interval, treat it as directional rather than precise. Use synthetic for the comparative questions and a real-respondent panel for the absolute-magnitude numbers that need statistical confidence.

Not for being unexpected. Surfacing a thread you didn't think to test is one of the best things synthetic research does. You can set out to test two hypotheses and have the study hand back a third that turns out more interesting. The discriminator isn't novelty, it's provenance. A surprising finding that traces back to real evidence is a lead worth chasing; a surprising finding that traces back to nothing is a hallucination in the same outfit. So check the provenance of the unexpected findings specifically, because they're the ones you're most tempted to act on. Discount a finding for being unsupported, never for being new.

Yes, deliberately. The point of provenance is that you don't have to take any vendor's word for the output, including ours. You can run the same checklist on a Candor report: look for variance across personas, trace striking findings to their sources, and verify that the unexpected results are grounded rather than invented. A tool confident enough to be inspected is the one worth trusting. A tool that asks you to trust the output because it looks good is asking for exactly the thing the research warns against. Reading synthetic output like a professional means inspecting it, not being impressed by it.

Candor is in development.

Be the first to know when it launches.

No spam. Just a note when Candor is ready. Powered by Highline Beta.