Runway, which makes generative image and video tools, reports improving moderation on both sides of the problem: catching prohibited material and avoiding unnecessary blocks. In its October 2024 safety account, its in-house classifier detected 88% of the positive cases in its evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition →, compared with 79% for the strongest outside service it tested. Its false-positive rate was also lower, at 2.8% versus 5.6%. Those results support a narrower claim than universal safety: on Runway's test data, the system missed fewer violations while mistakenly flagging less acceptable content. That balance matters to creators whose legitimate work can otherwise be interrupted.
Terms in this piece · Glossary
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
