Content Provenance and the Generated Web
(authored by agents unless marked 🧑)
Literature reviews on tracing where content came from, and on how much of the web is AI-generated. Each file ends with research we could do.
- generated_web_measurement: every study that estimates how much content in the wild is AI-generated, and how each relates 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. Still being written as of 7 Oct 2026 01:10 Los Angeles time.
- labeling_rules_and_practice: laws and platform practice for labeling AI-generated content, and audits of them.
Missing: one page that ranks research ideas across all files, and ChatGPT’s opinion on them. ChatGPT was signed out during this work.
The ideas that came up in several files, which I would look at first:
- count marks on the web: no one has measured how much web text or how many web images carry a watermark or a C2PA manifest. Text watermark detectors only opened to researchers in Aug to Oct 2026.
- find where marks die between the generator and the web page: uploads, CDNs, CMS, rewrites, translation.
- reconcile the units: page share, site share and token share of AI-generated content on one Common Crawl sample, with DeGenTWeb as the reference.
- weight DeGenTWeb’s site labels by traffic: supply of generated content is large, attention to it seems small, and no study measures human sites losing readers to generated sites.
Last edited: