Anthropic Explains Claude’s Text Watermark Through Statistical Word-Choice Patterns

01Anthropic Details Claude Text Watermark: No Hidden Characters, Detectable Pattern Comes From Random Word Choice

Anthropic says future Claude models will watermark generated text without inserting hidden characters or attaching extra information. The company is making the change to comply with the EU AI Act: since August 2, AI providers serving the European market have been required to mark AI-generated content, and other major model developers have signed the same Code of Practice.

The company has now explained how its watermark works and what a detection result can—and cannot—show. Claude will use a version of SynthID-Text, a technique published by Google DeepMind in a 2024 Nature paper, to create a statistical pattern through the ordinary process of choosing words.

Large language models generate text one word at a time. At each step, several candidates may be similarly sensible: after “The weather today was cold and,” for example, either “overcast” or “grey” might fit. When such choices have little effect on meaning, a random number normally helps determine which candidate is selected.

Watermarking changes the source of that randomness. Instead of drawing an arbitrary random number, Claude uses a secret key together with several preceding words to settle the choice. It does not force the model to use words it would not normally consider. Across many selections, however, the resulting sequence can form a pattern detectable by someone who possesses the key.

Anthropic compares the process to replacing dice rolls in a board game with successive digits of pi. The moves would still appear random to players, but someone who knew the underlying sequence could later assess whether it probably determined the moves.

Likewise, a matching text pattern supports only a probability estimate that Claude was involved in writing at least part of the text. It does not confirm authorship or prove that an entire document was generated by Claude. Anthropic also says the watermark contains no information identifying a person, organization, or individual conversation.

According to Anthropic, the technique has no practical effect on output quality or content and is indistinguishable to readers. Its internal tests found no impact on content, creativity, or readability. Google DeepMind’s SynthID-Text research likewise found no statistically significant difference in user ratings, while human evaluators in a controlled comparison saw no quality difference. The process uses no extra tokens, so Anthropic says it will not increase costs.

Anthropic has not disclosed in the provided explanation who will receive the detection key or how access to detection will work.

02Apple Reportedly Trained Its Own China AI Model With Alibaba as Apple Intelligence Clears a Regulatory Hurdle

Apple has reportedly worked with Chinese technology giant Alibaba to train a custom large language model for China, changing its approach as it prepares to introduce Apple Intelligence in the country. Apple Intelligence is the company’s suite of on-device generative AI features, but the American models it uses elsewhere, including OpenAI’s ChatGPT, are unavailable in China.

The China-focused model was developed with Alibaba and trained with the company’s support, Reuters reported, citing three unnamed people familiar with the matter. Neither company has formally confirmed the partnership details in the supplied report, and the model’s name and capabilities were not disclosed.

Building its own customized model would represent a departure from Apple’s previous strategy in China. Historically, the company has relied on domestic Chinese models to bring generative AI to devices sold there. According to the Reuters report, developing a proprietary model with Alibaba’s help would give Apple more control over its products in China’s competitive smartphone market.

Apple has also made progress on the regulatory path for Apple Intelligence. The company formally registered its on-device generative AI service with China’s cyberspace regulator last month, clearing what the report described as a major hurdle for deployment.

Registration does not mean the model has received final permission for public release. AI models must be registered with and cleared by the Chinese government before they can be offered publicly, and the report did not identify the remaining approval steps. Reuters said Apple would become the first US company approved to offer its own proprietary AI model in China if it obtains clearance.

People familiar with the matter told Reuters that Apple Intelligence is expected to arrive in China “in the coming months” following an update to the iOS operating system. That remains an estimate rather than a confirmed launch date, and no specific version of iOS was named.

Apple did not immediately respond to The Verge’s request for comment. The precise launch timing, the model’s capabilities and whether regulators will grant final approval remain unknown.

03Cerebras and OpenAI Preview GPT-5.6 Sol Ultrafast Service, Claiming Up to 750 Output Tokens per Second

Cerebras and OpenAI have previewed Ultrafast Mode, a new OpenAI API service tier that runs GPT-5.6 Sol on Cerebras hardware. The limited preview is initially available only to selected customers, with access set to expand as capacity grows. Cerebras says the service can generate as many as 750 output tokens—the units of text produced by a model—per second without reducing quality.

The companies are targeting workflows in which waiting for a capable model can carry an immediate cost. Cerebras identifies diagnosing production outages and responding to cyberattacks as possible uses, though it does not report customer results from either scenario. The company’s role is to provide the computing infrastructure intended to make GPT-5.6 Sol respond faster.

Cerebras attributes that speed to its Wafer-Scale Engine architecture. It says each wafer-sized chip contains 44GB of SRAM, allowing model weights to remain on the chip instead of being repeatedly moved between on-chip memory and external storage as successive tokens are generated. Tokens then flow through model layers arranged as a pipeline across multiple wafers, reducing data-movement overhead.

All published performance and quality findings come from Cerebras. In one test, GPT-5.6 Sol Ultrafast answered all 2,500 questions in Humanity’s Last Exam in 11 hours and 11 minutes. Cerebras said Claude Fable 5 required 78 hours and 27 minutes to reach comparable accuracy, making Ultrafast nearly seven times faster on total benchmark completion. The company tested Sol through Codex at “xhigh” reasoning on July 10 and Fable 5 through Claude Code at the same reasoning setting from July 13 to July 15.

A separate July 31 test measured complete workflow time rather than raw output speed. Using Codex at medium reasoning on GDP-Val, a benchmark for economically valuable knowledge work, Cerebras reported a 5.6-fold end-to-end speedup for Ultrafast over standard GPT-5.6 Sol with no quality degradation.

Those figures have not been independently verified in the supplied source. Cerebras and OpenAI also have not disclosed pricing, preview customer numbers, available capacity, or a date for broader access.