Don’t forget that manual review is not inherently required for all 70,000 prompts. GenAI also classifies images, and it’s probably a reasonable workflow within this studio’s capabilties to, say, mindlessly create a bunch of minor variations of the same thing, feed a sample of a few frames from the first N frames of each video into a classifier model, and tell it to mark them as probably bad and put them to the side if they look anomalous.

Obviously a GPT classifier has zero understanding of what that actually means, but you can probably filter out a sizable chunk of prompts this way – heuristic reason being that an AI-generated video tends to decohere over time as it feeds on itself, so the initial frames are probably the best you’re going to get.

You’ll have tons of false positives and won’t filter out all of the bad ones, but who cares about false positives when you can generate another one with a change of the prompt? This method I made up just now in five seconds is probably well-surpassed by a rapidly developing knowledge base about how to microoptimize the shit out of GenAI.