TextGrain
A text watermarking technology developed by OpenAI that encodes machine-readable invisible markers by subtly adjusting word choices in model outputs. TextGrain operates without altering the user experience and is designed to satisfy the transparency and provenance disclosure requirements of the EU AI Act.
Core Facts
Timeline (last 90 days)
Italy's Corriere della Sera reported on TextGrain, describing it as an invisible watermark (filigrana invisibile) used to track chatbot text.
TextGrain works by shaping the AI's word selection preferences to leave statistical traces, causing the model to favor specific vocabulary or grammatical structures during generation.
Because the watermark bias is distributed throughout the entire article, simply replacing a few words or paraphrasing sentences cannot erase the underlying statistical fingerprint, making the watermark robust.
OpenAI has stated that the use of its detection tools will be restricted, but has not yet clarified who has the right to view watermarks (employers, educators, or governments).
TextGrain forms patterns recognizable by detection tools through fine-tuning word choices, but editing, translation, or the use of very short excerpts may weaken the signal.
OpenAI launched a text watermarking system called TextScreen, which embeds statistical signals in the model's word choices rather than inserting visible markers for subsequent detection, in response to the EU AI Act, and plans to open-source it.
The longer the text, the higher the confidence of watermark detection; short texts have high statistical noise and a greater risk of false positives, while detection accuracy is more reliable for texts of several hundred tokens or more.
There is currently no reliable method that can accurately detect whether a text was generated by AI.
All Facts (8)
The longer the text, the higher the confidence of watermark detection; short texts have high statistical noise and a greater risk of false positives, while detection accuracy is more reliable for texts of several hundred tokens or more.
65%UnverifiedBecause the watermark bias is distributed throughout the entire article, simply replacing a few words or paraphrasing sentences cannot erase the underlying statistical fingerprint, making the watermark robust.
50%UnverifiedOpenAI has stated that the use of its detection tools will be restricted, but has not yet clarified who has the right to view watermarks (employers, educators, or governments).
50%UnverifiedTextGrain works by shaping the AI's word selection preferences to leave statistical traces, causing the model to favor specific vocabulary or grammatical structures during generation.
50%UnverifiedTextGrain forms patterns recognizable by detection tools through fine-tuning word choices, but editing, translation, or the use of very short excerpts may weaken the signal.
50%UnverifiedThere is currently no reliable method that can accurately detect whether a text was generated by AI.
50%UnverifiedItaly's Corriere della Sera reported on TextGrain, describing it as an invisible watermark (filigrana invisibile) used to track chatbot text.
50%UnverifiedOpenAI launched a text watermarking system called TextScreen, which embeds statistical signals in the model's word choices rather than inserting visible markers for subsequent detection, in response to the EU AI Act, and plans to open-source it.
50%