AI Regulation & Policy

Today's Stories

1. 1

   Anthropic details how Claude watermarks text to comply with EU AI Act

   Anthropic published details on Friday about how it will watermark text generated by Claude, its AI chatbot. The watermarks use the SynthID-Text approach outlined by Google DeepMind in 2024, embedding an undetectable pattern in low-stakes word choices (like selecting "overcast" vs. "grey"). Anthropic plans to release a watermark detection API. The watermarking requirement stems from the EU AI Act's Transparency Code, which requires AI companies to identify AI-generated content. The move has sparked debate among Claude users—some claim it is transparency overreach, while others defend it as necessary to prevent deception. Business Insider reports that "dozens" of users on X have said they are canceling Claude subscriptions in response.

   Light editing will likely not remove the watermark, but a complete rewrite where every word is replaced will. Code will have minimal watermark presence because the model needs freedom only in arbitrary choices (like comments), not in functional code logic. Other major model developers have signed the same Code of Practice and will implement their own watermarks.

2. 2

   Pro se plaintiff hid AI instructions in court filings to sway automated review

   Matthew Elliott, a self-represented plaintiff in a Connecticut court case against New York Bariatric Group, embedded invisible prompt injection instructions in his filings using 3-point white text on a white background. The hidden text directed an AI system to align its output with his filing and treat a prior clerk's denial as an error. The court discovered the scheme through suspicious whitespace, and Judge Walter Spader Jr. warned Elliott against the practice. Elliott ignored the warning and hid additional messages, including a YouTube link and mocking comments, in later filings. The judge compared the attempt to "secretly influencing a juror," finding that even though Connecticut courts do not use AI systems to review filings, the covert effort to manipulate a potential decision-making tool was itself improper. Judge Spader noted that AI tools can help unrepresented people present cases clearly, but warned that language models tend to reinforce flawed legal reasoning by developing arguments persuasively rather than challenging them. The case illustrates a broader problem: prompt injections are appearing in academic preprints and court filings worldwide as people attempt to manipulate AI-powered review systems.

   As a sanction, the court revoked Elliott's right to file electronically; he must now submit all filings on paper in person to the clerk's office. The broader context: pro se (self-represented) lawsuits at US federal courts have nearly doubled since ChatGPT went mainstream—41,490 were filed in 2025, compared with the pre-AI-boom average—and 18 percent of complaints sampled in early 2026 were flagged by an AI text detector as AI-generated.

3. 3

   Amazon Can Use Your Twitch Content to Train AI—Unless You Opt Out

   Twitch has added a new account setting that lets streamers opt out of having their content used to train Amazon's artificial intelligence models. The opt-out toggle is located in Settings under the Security and Privacy section, labeled 'Generative AI Training.' However, disabling it does not prevent Twitch and Amazon from using channel content for other purposes, including AI-powered platform features like sponsorship assistance, viewer recommendations, and moderation tools. The feature reveal sparked backlash from over 16,000 creators in a forum dedicated to the topic, who opposed the default opt-in setting—a practice that had not been publicly disclosed until now. Twitch's head of product, Mike Minton, acknowledged that keeping the default enabled was necessary because 'no one would participate' otherwise, raising concerns about how tech companies handle user data and the lack of transparency around when the training began.

   The move highlights a broader industry pattern: Meta has used posts and images from Facebook and Instagram, as well as employee activity and smart-glasses screenshots, to train AI models; similar practices have been documented at Google and YouTube. High-quality training data has become a strategic resource for AI developers, and user-generated content appears to be filling the gap as publicly available sources dwindle.

4. 4

   U.S. warns allies: join our AI coalition or face exclusion

   The United States is preparing to tell dozens of countries they must choose between joining the U.S.-led Pax Silica AI initiative or China's competing framework, with the threat of exclusion from the American coalition if they join both. Pax Silica, launched last year, aims to secure supply chains for AI models, semiconductors, and critical minerals. About two dozen countries have already joined, including Kazakhstan, Japan, Australia, and South Korea. The ultimatum reflects intensifying U.S.-China competition over AI and semiconductor supply chains. Countries like Kazakhstan that have joined both coalitions now face pressure to abandon one, complicating efforts by nations to maintain balanced relationships with both superpowers in a technology domain that shapes economic and national security strategy.

   How major U.S. allies and strategically important mineral suppliers respond to the either-or demand. Kazakhstan's position is particularly significant given its role as a key potential source of critical minerals needed for AI infrastructure.

5. 5

   Trump administration demands AI countries choose sides, threatens exclusion

   The Trump administration is preparing to send letters to more than 30 countries—primarily participants in the U.S.-led Pax Silic a framework—warning that nations simultaneously participating in China's "Global AI Cooperation Organization" will be excluded from the U.S. camp. Currently, only Kazakhstan holds membership in both initiatives. The U.S. is using access to critical minerals and semiconductor supply chains as leverage to force allied and developing nations into a binary choice between American and Chinese AI ecosystems. Countries with resources essential to AI development face pressure to abandon China-aligned initiatives or lose U.S. technological partnerships. This represents a hardening of techno-nationalist competition at a time when AI development relies on globally distributed mineral and manufacturing networks.

   The timing of the letters remains unconfirmed. Pax Silic a, launched by the U.S. and Japan late last year, currently includes around two dozen countries and regions (including South Korea, Australia, and the EU), while China's Global AI Cooperation Organization, established in July with founding members from 29 countries including Russia, Pakistan, and Laos, continues to expand membership.

6. 6

   Anthropic launches watermark detection API for Claude AI text

   Anthropic will offer a watermark detection API allowing third-party developers to identify text created by Claude. The watermark uses a variant of Google DeepMind's SynthID Text method published in Nature in 2024, subtly altering the randomness source during word selection to create a traceable pattern without affecting content quality or readability. Anthropic is implementing watermarking to comply with the EU AI Act; the company signed the EU Code of Practice on transparency for AI-generated content in July 2026. Unlike external detection tools such as Pangram that scan for typical AI patterns, this watermark approach relies on keys only Anthropic possesses, potentially offering more reliable detection in practice.

   All Claude models released after August 2, 2025 support watermarking built-in; older models will receive the feature in the coming months. The watermark works less reliably on short texts, fact-heavy passages, code, and heavily rewritten material, and cannot distinguish whether Claude wrote an entire text or only edited it.