Runway details its generative video safety guardrails
Runway has published the safety, fairness and provenance measures behind its generative video tools, including an in-house moderation model it says beats the best third-party API it tested.
Runway published an outline of the guardrails it uses to limit misuse of its generative media models. The post is organized around safety, fairness and integrity, and it includes some performance figures. All of them are self-reported.
Moderation. Runway says it built its own visual moderation system that works on both AI-generated and real-world images and video, and that it automatically blocks users who repeatedly try to generate inappropriate content. On held-out data, the company reports an F1 score of 83% and recall of 88%. The best third-party API it tested scored 70% and 79%. Runway also reports a false-positive rate of 2.8% against 5.6% for that API. It does not name the API or describe the evaluation dataset. The company says it reports child sexual abuse material, and attempts to generate it, to the National Center for Missing and Exploited Children.
Fairness. Runway says a "diversity fine-tuned" model reduces gender and racial stereotyping in prompts about professions such as doctor, CEO or nurse. It reports that the model improved a group fairness metric by 150% for perceived skin tone and 97.7% for perceived gender. It does not say which metric it used. The company also acknowledges that its tools work best with English prompts and says it wants to improve multilingual support. It gives no timeline.
Integrity. Runway says it has safeguards against generating content that depicts known personalities. It has also adopted the Coalition for Content Provenance and Authenticity (C2PA) standard, adding invisible watermarks and AI-generation metadata to every output.
The post contains no independent verification of these results and no technical detail beyond the headline numbers.