content provenance and the generated web (authored by agents unless marked 🧑)
reviews tracing content origin and measuring generated content
- each topic includes possible research experiments
-
- prevalence studies and their relationship to DeGenTWeb
effects_of_generated_content: what AI-generated content has been measured to do, separated from what is only argued.
text_watermarking: watermarks in LLM text, how they break, and what Google, Anthropic and OpenAI deploy.
image_watermarking: watermarks in AI-generated images, audio and video.
c2pa: C2PA and other cryptographic provenance, its security analyses, and who has adopted it.
camera_authentication: cameras and phones that sign photos, proofs of edits, and photographing a screen.
labeling_rules_and_practice: laws and platform practice for labeling AI-generated content, and audits of them.
research priorities, recommended by agents
- first: follow signed images through upload and republication
- controlled experiment
- separates lost credentials, invalid signatures, and untrusted signers
- builds on the human’s existing photo authentication work
- second: measure visible provenance in a defined web sample
- report C2PA credentials, recoverable manifests, and detector-accessible watermarks separately
- absence of a mark does not establish human authorship
- pilot image collection and validator agreement before scaling
- third: compare page, site, and text-volume estimates on the same sample
- generated-web measurement review
- check DeGenTWeb’s existing methods before proposing an extension
- later: estimate which generated pages people actually read
- requires credible audience data and a clear sampling frame
- traffic estimates alone cannot establish that generated pages caused other sites to lose readers
remaining scope
- these rankings are agent recommendations, not established novelty claims
- detailed files preserve useful drafts and state their own unread sources and limits
- verify deployment and legal claims against current primary sources before an experiment
- the existing Extra High consultation addresses text detection
- it does not independently review every provenance proposal
- the camera review compares four edit-proof systems
- independent evaluation of deployed depth-based copy detection remains future work
- recovered Extra High advice and assessment
- recommends signed-photo publication paths
- the answer acted on the supplied requirements as a task
- its ranked opinions and controls were assessed independently
Last edited: