Anthropic and Google use hidden watermarks to identify AI text
New watermarking methods allow companies to tag AI-generated content for machine detection, though the public cannot yet verify these marks.

The short version
- Anthropic now watermarks text across all its models and surfaces worldwide.
- Google uses a method called SynthID-Text for its Gemini app.
- OpenAI is adding text watermarks to ChatGPT and Codex in the EU.
- Watermarks are designed to be read by machines rather than humans.
- Only the AI labs and approved organizations currently hold the keys to detect the marks.
Why it matters
These watermarks create a technical infrastructure to identify 'slop'—the mass production of low-quality AI content. For users, it means a future where automated filters can potentially keep human-made content visible on the internet.
The push for machine detection
Anthropic has updated its support documentation to show that its newest and some older models now watermark text on every surface where Claude is used worldwide. This move follows similar actions by other major players in the industry, such as Google and OpenAI. These companies are implementing these signals to comply with the EU AI Act, which requires marking AI-generated content in a way that machines can read.
While the official reason is regulatory compliance, some observers suggest these watermarks could serve as a tool to combat 'slop.' This term refers to the mass production of low-quality AI text that fills up various corners of the internet. By creating a hidden signal, companies hope to build filters that add friction to the automated stuffing of the web with machine-generated content.
Currently, the public does not have the ability to verify these watermarks independently. While Google provides a tool called synthid.com to check images, video, and audio, it does not allow users to check text from any AI lab. Only the big AI companies and a few authorized organizations currently hold the necessary keys to run these detectors.
How tokens form the basis
To understand how watermarking works, it is first necessary to understand how large language models process information. These models do not see letters; they work exclusively with numbers. Every model has a fixed vocabulary of tokens, where each word or piece of a word is assigned a specific number called a token ID.
When a user enters text, a tokenizer splits that text into these tokens and replaces them with their corresponding IDs. For example, the word 'strawberry' might be treated as a single token in one model, but split into three separate pieces in another. The model then takes these numbers into a context window, which is essentially a large array of numbers representing the current conversation.
The model's primary job is to predict the next token by generating a probability for every single word in its vocabulary. However, the model itself never actually picks a word. That task is handled by a separate program called the inference runtime, or runner. The runner takes the probabilities and either picks the most likely one or rolls the dice to choose one at random.
The mechanics of watermarking
Watermarking works by influencing the way the inference runner chooses the next word. One common method, known as the green list method, involves dividing the vocabulary into two groups. The runner then gives a probability boost to the tokens in the favored group, known as the green list, while the shares of the red list tokens decrease to keep the total probability at 100%.
Anthropic uses a different version of this called SynthID-Text, which was published by Google DeepMind in 2024. Instead of boosting probabilities, the runner rolls its dice a few times to find several candidates for the next word. It then flips a virtual coin for each candidate to decide which one to write. The result of that coin flip depends on a secret key, the candidate word, and the words that came before it.
Because the coin flip is tied to a secret key, anyone with the key can re-run the same calculation on a piece of text. If the text was watermarked, the results of those coin flips will show a statistically significant bias toward 'heads.' If the text was written by a human, the results will look like a random split. This allows a detector to confirm the origin of the text without the watermark being visible to a human reader.
Who is affected by these marks
The immediate impact of these watermarks is felt by the companies and regulators who are trying to manage the flow of information online. By creating a machine-readable signal, they are building the infrastructure needed to distinguish between human-authored content and high-volume machine output. This is particularly relevant as concerns grow over the sheer volume of AI-generated content appearing on social media and news sites.
For the average user, the effect is currently invisible because the detection tools are not public. However, as these technologies mature, they may eventually power browser extensions or search engine filters that automatically flag AI-generated content. This could change how users interact with information, potentially prioritizing human-written sources in search results or social feeds.
Developers and researchers are also affected as they look for ways to bypass or analyze these watermarks. Because the watermarks rely on secret keys and specific mathematical patterns, they represent a new layer of security in the AI ecosystem. The ongoing tension between the ability to generate content and the ability to identify its source will likely define the next era of AI development.
What comes next for detection
The next step for watermarking technology involves making these systems more robust and harder to strip away. Currently, if a user heavily edits an AI-generated paragraph, the statistical patterns of the watermarks may be broken. Companies are looking for ways to make the watermarks more resilient to human intervention while still remaining detectable by their proprietary tools.
We can also expect to see more diverse applications for these watermarks beyond just text. As AI becomes more integrated into software, watermarking may be used to track the provenance of code, data, and even complex interactive elements. This creates a framework where the origin of digital assets can be verified across different platforms and services.
The ultimate goal for many in the industry is to create a standard for AI transparency. While the current system is closed and controlled by a few large corporations, the underlying technology of SynthID-Text provides a blueprint for how these marks might eventually be standardized. This could lead to a future where AI-generated content is clearly labeled by default across the entire internet.
What to watch
- Whether detection tools will eventually be made public for independent verification.
- How resilient these watermarks are to heavy human editing and paraphrasing.
Sources and links
- Officialannounced it’s adding one openai.com
- Officialanthropic.com/news/claude-text-watermark anthropic.com
- Officialsupport page on marking AI-generated content support.claude.com
- PaperA Watermark for Large Language Models arxiv.org
- Reportingoreilly.com: the article we sourced
Read the original at oreilly.com
First reported by oreilly.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.