Why AI Generations Get Rejected: Real Moderation Rates Across 11 Models

Nearly every guide about failed AI generations lists the failure types and stops there — identity drift, weird physics, prompt misreads. Useful, but it leaves the practical question unanswered: how often does this actually happen, and on which model?

We had the data sitting in a table, so we pulled it. Every generation that ran on FlyAIgh between 2026-05-06 and 2026-08-21 — 1,480 runs, 341 accounts, 40 models — sorted by what actually caused each failure.

The short answer

Content moderation refuses far more generations than the models break. And the gap between models is enormous: on the same platform, in the same window, one model refused 26% of everything sent to it while four others refused nothing at all.

The trigger is almost never the prompt. It is the reference image — specifically, a real human face in it.

Where these numbers come from

This is production traffic from real users, not a lab test. Nobody wrote prompts designed to probe the filters; these are people trying to make something and getting told no.

Before computing any rate we removed 59 failures that trace back to our side, not the model's — integration errors on our end (sending a text-to-video request to a model that only accepts image-to-video, malformed parameters) and platform-level failures during submission. A model should not be scored on our bugs.

Only models with at least 30 valid samples appear below. Eleven cleared that bar out of 40.

Rejection rates by model

Two separate columns, because they are two separate things. Moderation means the model worked and its content policy declined. Provider failure means the generation broke — timeout, 502, busy queue.

The two columns barely overlap. The models that refuse the most break the least, and the two that break the most — Seedream 5.0 Pro and Grok Imagine — refused nothing. Whatever is going on inside these systems, content policy and infrastructure reliability are unrelated.

Read the sample sizes before quoting any single number. GPT Image's 4.0% rests on 403 generations and is the sturdiest figure in the table. Kling V3 Omni's 0% rests on 35 — it means "no refusals observed in 35 runs", not "this model never refuses".

What actually triggers a rejection

These are the verbatim messages returned by the providers, with how often each appeared in the 77 moderation refusals:

• Real person in the reference image (15) — "The request failed because the input image may contain real person." Plus 12 more naming specific slots (content[1], content[2], content[3]), which tells you it checks every reference you attach, not just the first.
• Explicit or suggestive material (15) — "Contains inappropriate content. Please remove suggestive or explicit material."
• Celebrity likeness (6) — "Celebrity or public figure imagery is not supported."
• Copyright and audio rights (5) — "Content security audit did not pass. The output audio may be related to copyright restriction." Plus separate audio-filter refusals on models that generate sound.
• Content involving minors (4) — refused outright, no appeal path.
• Watermark removal (3) — refused when the input looks like an attempt to strip a watermark or reproduce protected material.

The pattern worth internalising: the reference image is what gets you refused, not the prompt. Real faces alone account for more than a third of every refusal we logged.

Why the Seedance family refuses most

All three Seedance 2.0 variants land between 14% and 26%, well above everything else. Two things drive that, and only one of them is about the filter.

First, Seedance accepts up to nine reference images per generation. More references means more surfaces for the filter to check — and the refusals naming content[1] content[2] content[3] confirm it checks each one. A model that takes a single reference has fewer chances to trip.

Second, people reach for Seedance precisely because it accepts many references, which is the workflow you use when you are trying to keep a specific person consistent across shots. The use case pulls in exactly the input the filter is strictest about.

How to get rejected less

1. Fix the reference before you touch the prompt. Most refusals fire on the image. Rewording rarely helps.
2. Replace photos of real people with generated characters. A consistent AI-generated character passes filters that a photograph of an actual person will not. This is what character tools are for, and it sidesteps the most common refusal entirely.
3. Drop references you do not need. On a nine-reference model, every extra image is another check to fail. Send three good ones instead of nine mediocre ones.
4. Never send public figures. Celebrity likeness is refused across every provider we route to. There is no model in the table where this works.
5. Check the source for watermarks. Stock previews and screenshots carry them, and the filter reads that as an attempt to strip protection.
6. Switch models rather than fight the filter. If real-person references are core to your work, VEO 3.1 Lite, Kling V3 Omni and Z-Image Turbo showed no refusals in this window. Moving the job is faster than a fifth rewrite.

One structural point: on FlyAIgh credits for failed and rejected generations are refunded automatically. At a 26% refusal rate that stops being a nicety — it is the difference between paying for one usable clip and paying for one and a third. Refund policies vary by platform; worth checking before you batch anything on a high-refusal model.

What this data does not tell you

Four limits, stated plainly, because a number without its boundaries gets misused:

• This measures refusals, not quality. A generation that completed and looks bad counts as a success here. Whether output is usable is not something the database records, so this article cannot speak to it.
• Rejection rates reflect what our users sent. A different audience with different source material would produce different numbers on the same models.
• Filters change without notice. This is a May–August 2026 window. Providers adjust content policy continuously and rarely announce it.
• Sample sizes are uneven. 403 generations versus 35 is not the same quality of evidence, and the table shows both so you can weight them yourself.

Everything above comes from one query against production data. We will re-run it as samples grow, particularly for the models currently under the 30-generation bar.