Google sends four AI agents after AI slop networks. Its paper shows no test results

Google has put a four-agent AI system to work against synthetic video networks. The system is called SAFE, short for Scaled Abuse Forensics Examiner. It appears in a three-page paper signed by seven Google authors, dated 28 May 2026. Search Engine Journal wrote about it on 25 September.

SAFE judges a cluster of channels, not one video

The system looks at groups of channels at once. It asks whether a cluster is a coordinated spam operation or ordinary activity. That is a shift from checking one upload at a time. Spam networks spread work across many accounts, so a single video often looks harmless.

One agent runs the case, three do the digging

The Root Agent directs the work and issues the verdict. The Content Understanding Agent separates real media from machine-made filler, using a tuned language model. The Behavior Understanding Agent reads infrastructure signals such as IP addresses and upload times. The Channel Cluster Understanding Agent maps how channels connect to each other.

Google already ran an earlier system, S-CTS, described in June. SAFE replaces manual forensic work that took analysts days.

Seven authors, three pages, no test numbers

The paper carries no measured results. Its evaluation section lists metrics Google says it will use later. The one claim about performance is a sentence saying early deployment “significantly accelerates” the work, with no figure attached. That makes the paper a design note rather than proof.

The timing matters for anyone publishing video. Google started its September spam update on 24 September, a rollout that runs for weeks. Platforms are tightening labels too: Instagram said it will show fewer unlabelled AI profiles. Google has also been putting agents into its own ad products, including a chat agent inside YouTube ads.

Brands that buy or run video channels lose the most if a cluster gets flagged. A network of look-alike channels built from the same template now reads as a pattern, not a portfolio. Creators with one channel and human faces on camera gain, because the signals SAFE reads point away from them.