OpenAI follows Anthropic with text watermarks; dev open-sources rewrite-based remover
Imaginary_Dinner2710 · reddit · 2026-10-06
OpenAI is rolling out textGrain statistical text watermarks for eligible ChatGPT and Codex users in the EU, following Anthropic's global watermarking of supported Claude models. Both embed statistical signals in word choice rather than hidden Unicode, so only rewording changes the signal.
- The author tested on the public SynthID Text scheme: light synonym swaps barely helped, but a full rewrite by a second model pushed all 10 test reports below the detector threshold
- He packaged the workflow as a local MCP server/CLI: masking links, quotes, amounts and percentages, rewriting via an OpenRouter free model, then validating protected spans, length, structure, and 5-gram churn (replacing most 5-gram sequences significantly weakens watermark signals)
- Works with Claude Desktop, Claude Code, Codex, Cursor and other MCP clients
- Caveats: free-model endpoints took 61-216s and changed 70-87% of wording, with no semantic checking
- The author also runs a paid hosted version charging per token
He cannot honestly claim removal of textGrain or Claude's production watermark since their detectors are restricted.
Related event: OpenAI rolls out textGrain text watermarking, starting in the EU(19 posts)→
More from coding & agent
- Zeroization can make things worse: how wiping secrets creates more copies — jedisct1 · 2026-10-06
- Indie dev launches agent-first creator marketing platform Clipatra, pays out $3,000+ — tibo_maker · 2026-10-06
- TensorFold bonds dual Thunderbolt 5 links for 83% throughput boost on Apple Silicon — AIFlow_ML · 2026-10-06
- A 45-minute visual tour of graph theory, taught through the author's hometown — TivadarDanka · 2026-10-06
- Open-weights Kolibri-1 plays Breakout with no fine-tuning at ~25ms per move — Nils_Reimers · 2026-10-06
- Gatana adds centralized skills sync to MCP Gateway across all agents — Gatana_Official · 2026-10-06