Labelling requirements for generated content are taking effect. Unlike the early single "AI-generated" tag, several platforms have adopted graded labelling.
Three tiers are now common: produced entirely by a model, assisted by a model on top of human work, and modification of existing material by a model.
Where does "assisted" begin
The trouble sits in the middle tier. Does polishing text with a model count as assistance? Does removing a passer-by from a photograph count as modification? Platforms define these differently, and the burden of understanding falls on creators.
The rules themselves are not hard to follow. What is hard is that every platform’s rules are slightly different.— A content creator
- Tiers: generated / assisted / edited
- Methods: visible labels plus invisible watermarks
- Difficulty: defining the line between assisted and edited
- Effect: higher compliance cost for cross-platform publishing
Technically, visible labels and invisible watermarks are advancing in parallel — the first for readers, the second for provenance — though watermark survival through repeated transcoding remains an engineering problem.
I had not thought about it from this angle. It did shift my view a little.
Hope there is a follow-up. These things tend to drop out of view after a few weeks.
Hope there is a follow-up. These things tend to drop out of view after a few weeks.
I have seen something similar locally, and this matches what I noticed.