Mmastodon TechnologyAI first seen 8 h ago, last 7 h ago, peak #11

OpenAI text watermark detection drops sharply when words are altered

Original: OpenAI text watermark detection drops to 17% after 25% of words change: About 80% of 200-token passages are caught at a

OpenAI's text watermarking system detects about 80% of 200-token passages at a 1% false-positive target, but accuracy falls to 17% once 25% of the words are changed. The tool is available to API users on an opt-in basis, and text from ChatGPT and Codex will be marked in the EU within weeks. Observers are highlighting the system's vulnerability to simple paraphrasing and edits.

Why now: Fresh details on the effectiveness and limits of OpenAI's watermarking raise questions about AI-content detection reliability.

OpenAIChatGPTCodexEuropean Union

Open on mastodon →

Rank over time, top of the chart is #1. 2 snapshots from 8 h ago to 7 h ago.

Evidence

• OpenAI text watermark detection drops to 17% after 25% of words change: About 80% of 200-token passages are caught at a 1% false-positive target. API users can opt in now; ChatGPT and Codex text in the EU is marked within weeks. https:// ppc.land/openai-text-watermark… · ppcland · 1

API: https://socialmediatrends-api.osmike.com/v1/trends/1207915