The Synthetic Data Trap
The pitch is still seductive.
Why pay a credentialed reviewer when a frontier model can generate a million preference pairs for the cost of the GPU time? Why wait weeks for human feedback when a judge model returns a verdict in a second? The economics look so good that an enterprise team can make a compelling internal case for running every post-training iteration on synthetic data alone.
The economics are real. The trap is real too. And the trap is the exact part of the work that matters most.
Where synthetic data works
Be honest about this first. Synthetic data is not a gimmick.
On the easy cases — the cases where the task is well-defined, the distribution is covered, and the answer is knowable by any reasonable evaluator — a frontier judge is competitive with, and sometimes better than, a median human reviewer. Cheaper. Faster. More consistent across batches. More available at 3 a.m. on a Sunday when the next training run is scheduled.
Most of the volume in a post-training pipeline lives on the easy cases. Some of it is boilerplate: well-formed answers to well-formed questions, style matching, tone adjustment. Some of it is augmentation: generating variations of a human-written example to expand the training set. Some of it is cheap labeling: rank-ordering responses where the top choice is obvious.
If a team uses synthetic data for this layer and stops there, the team is not being foolish. The team is using the right tool for the right part of the job.
Where synthetic data breaks
The breakage is structural. It is worth saying slowly.
A judge cannot grade what a judge cannot do.
The fringe cases — the ones where safety matters, where bias hides, where the right answer is not obvious, where context changes the correct call — are by definition the cases where the judge model is also uncertain. A judge that produces confident output on a case it does not actually understand is the same judge that would produce confident output if it were the model being trained. The two are correlated, not independent. The "evaluation" is the same model grading itself, with extra steps.
On the easy cases this does not matter. On the hard cases it matters enormously. The hard cases are where the release decisions live. The hard cases are where the safety incidents come from. The hard cases are where the training signal has to be most reliable.
A team that uses a frontier judge on the hard cases is producing training data whose quality degrades exactly where the stakes are highest.
What teams learn the expensive way
Three patterns keep appearing.
Drift the judge cannot see. The production distribution shifts. The judge model was trained on a distribution that no longer matches. The judge scores the new release favorably. The new release ships. Users find the failure in a week. The team goes back to human review on the drifted slices. The slices are usually the most valuable part of the product.
Hidden bias. A judge model shares the biases of its training data. A preference pair ranked by the judge inherits those biases. A reward model trained on the preferences inherits them in aggregate. The customer-facing model reproduces them. A credentialed human reviewer would have flagged the bias. The judge did not, because the bias is invisible to a judge trained on the same data.
Capability ceiling. On novel capabilities — new kinds of reasoning, new kinds of tool use, new domains the model is being extended into — the judge cannot grade work beyond its own ceiling. Training data above the ceiling is unavailable to the pipeline. The team paying for synthetic feedback is capping its own upside at the judge's best day.
The pattern that works
Hybrid. Routed. Tiered.
Synthetic data for the volume. Credentialed humans for the cases where synthetic data breaks.
The routing is the product. An incoming training example hits a classifier. The classifier decides whether the case is inside the judge's confidence region or outside it. Inside-region cases get the judge. Outside-region cases route to a reviewer whose credentials match the domain. The tiering is explicit. The ratio is tunable. The measurement runs live.
This is hybrid routing. It sits inside the AI Labs product — AuraQC measures the judge's calibration, Workforce holds the reviewer roster, and Cleo's structured interviews keep the roster calibrated as the work shifts. Control Center shows the ratio going into every training batch so a release lead can see, before the model trains, what percentage of the training signal came from which source.
The result is not "cheaper than all-human." It is not "as reliable as all-human." It is better than either, for the reason that neither alone is sufficient for the work the team is actually doing.
What to measure
Three numbers will tell a post-training team whether the hybrid is working.
Judge-reviewer agreement on sampled cases. Every batch, a sample of judge-scored cases goes to a human reviewer for adjudication. The delta is tracked over time. Drift in the delta is the earliest signal that the judge is failing at a task it used to handle.
Override rate on routed cases. The cases the classifier routed to a human reviewer get scored. How often did the human disagree with the judge's pre-routing score? If the override rate on routed cases is high, the routing is working. If it is low, the classifier is being too generous with the judge's confidence.
Capability ceiling drift. As the model's capabilities extend, the judge's utility shrinks at the frontier. Track the ratio of training examples that require human judgment by month. A healthy ratio for a team pushing capability frontiers is going up, not down.
What to do this quarter
If your post-training pipeline runs on synthetic data alone, three moves.
One. Identify the fringe cases. Walk the last quarter's incidents. Classify each one by whether a credentialed human reviewer would have caught the failure before it shipped. Count.
Two. Wire a sampled human review into the next training run. Not a full re-label. A sample, on the cases the classifier thinks are hardest. Measure the delta between the judge's score and the reviewer's score. That is your ceiling data.
Three. Make the routing explicit. No more "we use synthetic data for everything." Start classifying. Start tiering. Start measuring.
Synthetic data is a real help on the easy cases. It is not a substitute for human judgment on the cases that matter.
A judge cannot grade what a judge cannot do.
Build for that.
---
Ready to see what hybrid routing looks like in one system?
→ Workforce → AuraQC → Cleo → Talk to us