Skip to content

OpenAI Opens Text Watermarking: Editing Determines What Detection Can Tell Us

AI

TL;DR

OpenAI enables optional textGrain watermarking for API customers worldwide, with an EU ChatGPT and Codex rollout planned. Replacing 10% of words reduced detection from about 92% to 66% in an English test. Provenance signals need room for inconclusive results.

OpenAI Opens Text Watermarking: Editing Determines What Detection Can Tell Us

OpenAI has opened a text-watermarking option while publishing a limitation: replacing just 10% of the words in English passages reduced detection from about 92% to 66%. That does not establish how much signal survives editing in Traditional Chinese. Content products serving Taiwanese readers cannot treat an English evaluation as a Chinese accuracy measurement.Announcement Test coverage

The announcement is dated 2026-10-05 without a time or timezone. This article uses October 6 in Taipei, with a 15:53 cutoff, and selects the previous day’s event within the 48-hour window. TechCrunch’s timestamp is October 5, 13:36 PDT. A report crossing midnight does not establish account-level activation.Date Timestamp

API customers worldwide can enable textGrain for selected models; it remains off by default. Eligible ChatGPT and Codex text output in the EU is scheduled to follow over the coming weeks. The text detector is not generally accessible: applications are open, with initial access restricted to approved researchers and expert organizations.Availability

During next-token selection, textGrain uses a secret key and preceding context to influence sampling, accumulating detectable statistical relationships. Detection requires the text and key, without needing the generation-time strength setting. The signal concerns the generation process; it does not identify the user.Technical report Limits

The editing evaluation used 400-token English responses to ELI5 questions. Replacing a quarter of the words reduced detection further, to 17%. These are vendor evaluations, not independent replications by the publications covering them. They provide no evidence of practical performance on Traditional Chinese editing.Conditions Independent coverage

Editorial products need an inconclusive result

I support recording provenance during generation, but would not turn this feature into automatic rejection of submissions. Suppose an author writes a piece and asks a model to polish it. The result may combine human judgment with model processing. A detected watermark cannot quantify the author’s effort. Rejecting the submission on that basis would substitute a provenance question for a quality assessment.Signal boundaries

Missed detections can also change incentives. Suppose a platform accepts only submissions without detected watermarks. Someone repeatedly substituting words might pass, while someone preserving the original receives more scrutiny. This is a risk inferred from editing weakening the signal, not an observed platform case. A more useful application would flag the generation process for discussion during review, allowing authors to provide their evidence and revision history. It should not demand proof of something the detector itself cannot establish.

For teams offering writing tools, the API option permits provenance marking to start at generation.API access Marking and detection currently have different access conditions, however, so a product cannot promise that every recipient can verify the text. Its interface should distinguish a mark being added, a passage not yet checked, and a check finding no mark. That last result may still reflect short text, editing or another model. It should not become a label asserting human authorship.Access and missed detections

Before using detection in formal review, I would require evidence for the relevant language and common editing practices, with missed detections and false positives reported separately. Their consequences differ: one loses a provenance clue; the other may expose an author to unnecessary suspicion. Optional marking is available now. General text verification remains restricted, and reliability after Chinese editing is still an evidence gap.Current scope

The cover reuses this site’s OpenAI brand photograph, not a textGrain interface. Downloading the evaluation chart failed because DNS resolution was unavailable, so a related local image was used.

Sources:

More on this topic

Random article