AI Companies Have Turned to Invisible Watermarks to Identify AI-Generated Text

OpenAI, Google and Anthropic are introducing technologies designed to identify artificial intelligence-generated text through invisible statistical watermarks, a shift that could give publishers, educators and regulators new ways to trace AI-assisted writing as synthetic content becomes harder to distinguish from human work.

The systems embed machine-readable signals in the words or tokens selected by AI models without adding visible labels, unusual characters or formatting. The resulting sentences can look entirely ordinary to readers while retaining statistical patterns that specialised detection systems can identify.

OpenAI disclosed details of its textGrain watermarking technology on Oct. 5, saying it was developing the system in response to European Union requirements for identifying AI-generated content. Google DeepMind has developed SynthID for text generated through its Gemini products, while Anthropic says supported versions of its Claude models are being equipped with embedded watermarks under its commitments to AI transparency rules. 

The technologies reflect a growing effort by AI developers to establish the provenance of machine-generated content as governments, news organisations and educational institutions grapple with the spread of synthetic material and the limitations of conventional AI detectors.

Unlike systems that attempt to determine whether a passage was written by AI by analysing its style, vocabulary or sentence structure, watermarking introduces a signal during text generation that can subsequently be tested by a detector.

How invisible text watermarks work

Large language models generate text by repeatedly estimating which token, or unit of text, should come next. A token can represent a word, part of a word or a punctuation mark.

Under a watermarking system, the model’s selection process is subtly adjusted to create a statistical pattern across the generated passage. The pattern is designed to be difficult for a human reader to notice while remaining detectable by software with access to the appropriate detection mechanism.

OpenAI says textGrain changes the randomness used in a model’s word-selection process rather than inserting hidden characters or special punctuation. Its detector checks whether a passage’s word choices contain the expected statistical signal.

The company says the system uses a secret key to select among adjusted probability distributions for possible next tokens. Those distributions are balanced so that, when averaged, they match the model’s original probabilities. A detector with the corresponding key and compatible settings can then test whether the observed choices are consistent with the watermark. 

Google’s SynthID for text follows a related principle, modifying the probability distribution of tokens as the model generates a response. The resulting pattern can be assessed against the expected watermark to estimate whether the text came from a system that uses the technology.

Google introduced SynthID text watermarking in the Gemini app and web experience in May 2024, describing it as an imperceptible signal intended to preserve the quality, accuracy, creativity and speed of generated text.

The company has cautioned that SynthID is not a definitive solution for identifying every piece of AI-generated writing. 

Anthropic is adopting embedded text watermarks for supported Claude models, alongside Content Credentials, a standard for attaching provenance information to supported files. Its published guidance says models launched in the European Union on or after Aug. 2, 2026, support machine-readable marking at launch, with marking also applying to supported models and products worldwide.

The company is working to extend watermarking to earlier models. Detection remains in private preview for eligible organisations, including regulators, media organisations, fact-checkers and researchers. 

OpenAI begins with Europe

OpenAI’s textGrain announcement comes as the European Union prepares to enforce transparency obligations for AI-generated content under the AI Act.

The company says it will introduce text watermarking for eligible ChatGPT and Codex users across all subscription plans in the European Union, initially limiting the rollout to that region rather than making it a global default. OpenAI also said on Oct. 5 that API customers worldwide could opt in to watermarking for selected models.

Approved researchers and expert organisations can apply for access to OpenAI’s detector, which is designed to report whether it finds the company’s watermark without identifying a user or revealing their prompts or conversations. The detector is not being made publicly available at launch. 

The European rules are helping push AI companies toward systems that can identify generated material even when ordinary file metadata is missing. The EU’s transparency framework covers machine-readable marking as well as other forms of disclosure, with requirements and implementation details depending on the type of content and the applicable provisions.

Watermarks differ from Content Credentials, which use the Coalition for Content Provenance and Authenticity’s C2PA standard to attach signed information about a file’s origin and history. Metadata can provide richer information but may be stripped when a file is copied, edited or converted. Embedded watermarks can survive some transformations because their signals are incorporated into the generated content itself.

Neither method is guaranteed to survive every alteration, and neither automatically establishes who wrote a passage or whether its claims are true. 

Detection remains imperfect

The central challenge is making a watermark sufficiently strong to detect without compromising the quality or diversity of the text.

Watermarking works best when a model has multiple plausible ways to express an idea. It becomes more difficult when the output must be exact, such as a mathematical answer, a short factual response, computer code or text that closely reproduces material supplied by a user.

Google says SynthID performs best on longer passages that allow greater variation in wording. Minor edits and some paraphrasing may leave enough of the signal intact, but extensive rewriting or translation can substantially reduce detection confidence. 

OpenAI has reported similar limitations for textGrain. Its published evaluation says shorter or more constrained passages are harder to detect and that strong performance under controlled testing does not guarantee reliable results in everyday use.

The company also reported that textGrain matched or exceeded SynthID in its own evaluations. That comparison is OpenAI’s assessment of the approaches it tested, rather than independent evidence that one system is universally more reliable. 

Anthropic similarly warns that a detected watermark indicates that content may have been generated or processed by Claude, not that Claude necessarily originated every idea or sentence. A model may have edited, translated or summarised material written by a person, and the resulting text may carry a watermark.

The reverse is also true: failure to detect a watermark does not prove that a passage was written by a human. The text may come from an unsupported model, predate the watermarking rollout, be too short to analyse reliably or have been extensively rewritten. 

These limitations matter for schools, newsrooms and employers considering whether to use automated detection to investigate plagiarism, verify authorship or assess compliance with disclosure rules. A statistical signal can provide evidence of a model’s involvement, but treating it as conclusive proof could produce false accusations.

A new layer in the AI accountability debate

The expansion of watermarking also raises questions about who can inspect the signals, how widely detection tools will be available and whether independent researchers can test their reliability.

OpenAI says it plans to release textGrain as open-source technology, potentially allowing other developers to study the approach and build compatible tools. Google has also announced plans to open-source SynthID for text through its responsible AI toolkit. The availability of source code, however, does not necessarily mean that every detector will have access to the secret keys or proprietary settings needed to verify a particular watermark. 

The systems are also not interchangeable by default. A detector designed for one company’s watermark may not recognise another company’s signal, while text generated by models that do not use watermarking may carry no detectable mark at all.

For news organisations, the technology could eventually provide an additional verification tool when assessing submissions, public statements or other material suspected of being generated by AI. For educators and regulators, it could help establish whether supported models were involved in producing a passage.

But the technologies do not determine whether a text is accurate, original, deceptive or ethically produced. Nor do they establish how much a person contributed to an AI-assisted document.

Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our 
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Contact Us to reach a global audience with Impact Newswire.