Margaret Mitchell: text watermarking arrived five years too late as AI slop floods the web
mmitchell_ai · x · 2026-09-22
AI researcher Margaret Mitchell (ex-Google, Hugging Face) made two points in a thread on agent oversight:
- Because of their stochastic-parrot limitations, AI agents don't do what we intend via natural-language instructions, so designing oversight mechanisms is arguably more important right now than expanding agents' action surfaces.
- The Stochastic Parrots paper urged prioritizing text watermarking five years ago; adoption is only happening now — after the internet has been garbled by AI slop, hampering our ability to train on human text. She cites Kirchenbauer et al.'s landmark ICML 2023 LLM watermarking work (softly promoting randomized "green tokens" during sampling for algorithmic detection), which Hugging Face embraced early.
Related event: Margaret Mitchell's New Paper: AI Agents Are Pushing Humans Out of the Loop(4 posts)→
More from Safety
- Agent Swarms Are Autonomously Hacking Critical Infrastructure Run by Underpaid Devs — menhguin · 2026-11-12
- Researcher says Anthropic dodged questions on Mythos false positive rate; 37,966 findings, 2 exploited — OwariDa · 2026-11-12
- Who reads the queries your agent writes? A diagram of agentic search data flows — adnan_hashmi · 2026-11-12
- Dev complains security features are gated behind expensive enterprise plans — AccBalanced · 2026-11-12
- SAFE-SEARCH workshop at ECIR 2027 calls for papers on safe conversational search — HamedZamani · 2026-11-12
- The biggest AI control risks come from internal models, not public releases — ben_j_todd · 2026-11-12