How Anthropic’s Invisible Claude Text Watermarks Are Meant to Work
Anthropic’s planned invisible Claude text watermarking is designed to leave a detectable statistical pattern in generated language. Here is what that can—and cannot—show.
Generative AI News & Analysis
Follow the developments, releases, research, and ideas shaping how generative AI is built, adopted, and governed.
Source-led reporting and analysis, selected for consequence and relevance.
OpenAI Adds Opt-In Computer History to ChatGPT’s macOS App
OpenAI has added Computer History to the ChatGPT app for macOS, an optional feature that lets ChatGPT and Codex reference activity from selected apps and websites. The company says the feature is off by default, excludes private browsing and does not capture screenshots or audio.
Anthropic Plans Machine-Readable Provenance Signals for Supported Claude Output
Anthropic says supported Claude models will use machine-readable provenance signals for content they process, though a detected signal does not establish that Claude authored all of the material.
Google Rolls Out Gemini 3.7 Flash, Replacing Its Fast Model After Three Weeks
Google has introduced Gemini 3.7 Flash, a new fast, lower-cost model aimed at coding and agentic workflows, only three weeks after releasing Gemini 3.6 Flash. The release combines Google-reported benchmark gains with temporary lower API pricing, while underscoring how quickly teams building on model APIs may need to re-evaluate production choices.
Google Releases Gemini 3.7 Flash Three Weeks After Its Previous Flash Model
Google has announced Gemini 3.7 Flash, a coding- and agent-focused update to its lower-cost Flash model line, just three weeks after Gemini 3.6 Flash. The rapid cadence makes workload-level testing important for teams considering a production switch.
Twitch Adds Opt-Out for Amazon Generative-AI Training on Channel Content
Twitch has introduced a channel-level setting that allows creators to opt out of having specified channel material used in future training of Amazon generative-AI models. The control covers streams, VODs, clips, channel chat and channel text or images, while leaving functions such as captions and safety tools in place.
Google Says Gemini App Has Reached 1 Billion Monthly Users
Google says more than 1 billion people now use the Gemini app each month. The company’s claim signals the scale that AI assistants can reach through major-platform distribution, while leaving important questions about measurement and comparability unanswered.
Meta Releases Muse Glimmer and Plans to Open Muse Spark 1.2 Weights
Meta has released the open-weight Muse Glimmer model and said it plans to publish weights for Muse Spark 1.2 in the coming weeks. The announcement accompanies Mark Zuckerberg’s new AI strategy essay and marks a renewed emphasis on model access for developers.
Ford Is Reportedly Rolling Out an AI Assistant for Vehicle Questions in Its Apps
Ford is reportedly introducing an AI-powered assistant in its Ford and Lincoln mobile apps to answer questions tied to an owner’s vehicle. The reported app-first deployment points to a growing automaker focus on making connected-vehicle data conversational, but primary documentation on the product’s availability and limits remains unavailable.
Amazon-Backed West Texas Data Center Could Be Powered by a 7.65-Gigawatt Gas Plant
Amazon says it acquired the GW Ranch site in Pecos County, Texas, where Pacifico Energy is developing new on-site generation for a planned data-center campus. Permit reporting describes up to 7.65 gigawatts of gas capacity, initially outside the Texas grid.
OpenAI Slows Astra Work After It Cannot Rule Out Critical Cyber Capability
OpenAI has paused certain internal activities involving its unreleased Astra model after evaluations and expert assessments found the company could not rule out that the system has critical cybersecurity capabilities under its Preparedness Framework. The move is precautionary: OpenAI has not said Astra carried out a real-world attack or definitively crossed the threshold.
Suno Plans Watermarks and Download Limits for AI-Generated Music
Suno has outlined planned audio watermarking, fingerprinting, transparency tools and revised download rules aimed at making AI-generated music easier to identify and harder to mass-distribute. The company has not disclosed a launch date, technical specifications or final download terms.