OpenAI Knew It Was Breaking the Web

Newly unsealed court filings in the New York Times' lawsuit against OpenAI and Microsoft, covered in full by The Verge, contain language that neither company is likely to enjoy reading in public.

The documents include an internal Microsoft warning that the company's AI content strategy had created a "doom loop" — a self-defeating cycle in which AI models consume the web's content, undermine the publishers that produce it, and ultimately degrade the data supply the models themselves depend on. Microsoft's Director of Applied Science, Brent Hecht, went further, describing the training process as the "largest theft of labor in human history" and saying Microsoft's legal defence made a "complete mockery of the idea of fair use." Microsoft has since said these were one employee's personal views, not the company's position.

Other internal material is harder to dismiss. A senior OpenAI employee admitted the company was "unaware" of any effort to detect or remove paywalled content from training data, despite Microsoft CEO Satya Nadella later saying paywalled content should be licensed. Internally, OpenAI acknowledged GPT-4 had "memorized a ton of data" and would be "insanely good at regurgitation" — referring to reproducing copyrighted articles verbatim.

These are all allegations in an active lawsuit. Nothing has been proven in court. They are the companies' own documents, but they were submitted by the plaintiff and framed accordingly.

What to watch: whether these disclosures shift how judges interpret "fair use" in AI training cases, and whether other publishers now have a stronger basis for similar claims.

Claude Code is Anthropic's AI coding tool. Its revamped Projects feature, detailed by The Verge, lets you run a team of AI coding agents simultaneously in the cloud rather than managing one at a time.

Each project has a coordinator agent that assigns work across multiple threads — individual agents each working on their own copy of the codebase in parallel. When threads touch the same code, clashes are handled the same way a standard code merge works. You can talk to individual threads or oversee everything from a single project chat.

It is aimed squarely at developers handling large or complex coding tasks, not casual users. Free tier details are not confirmed; it launches in beta for select Claude Pro and Max subscribers, with broader access coming later. Enterprise users are also included in the rollout.

The local tools and code support is listed as "coming very soon" — not available today.

Honest read: Worth watching if you use Claude Code professionally — parallel agents on complex tasks is a genuine workflow improvement — but the beta access restrictions and missing local support mean most users cannot get full value from it yet.

What it is: SynthID-Text is a watermarking system for AI-generated text. Watermarking here means embedding a hidden signal in the model's output so that anyone with the right key can later confirm the text was AI-generated — a process called provenance tracking.

What's new: New research from Lasso Security, covered by Ars Technica, finds that applying SynthID-Text changes more than word choice — it also changes whether a model follows its own safety rules.

How it works: Instead of choosing the next word randomly, SynthID-Text uses a secret key to score candidate words in a bracket-style tournament. The winning word is chosen based on those hidden scores, not purely on which word makes the most sense. The sequence of winning words carries a detectable statistical fingerprint. But because the key subtly steers every word choice, it also shifts how the model responds to safety-relevant instructions. Researcher Andrea Siposova tested six open-weights models and found that under prompt injection (where an attacker embeds hidden instructions in content the model is processing), watermarked models were more likely to comply with harmful requests they would otherwise refuse.

What it can't do: The research was conducted on Hugging Face's unmodified watermarking implementation. Real-world deployments may be configured differently, and the researchers note frontier-scale models were not fully tested.

Who should care: Developers deploying AI agents with watermarking enabled — safety testing without watermarking active may not reflect real-world behaviour.

Read it: Ars Technica

Given today's coverage of AI watermarking, it is worth knowing that Google has made a SynthID detection tool available via Hugging Face. SynthID-Text (Google's watermarking system, explained in today's issue) embeds a hidden signal in AI-generated text. The Hugging Face demo lets you paste any passage of text and check whether it carries that signal — useful for editors, teachers, or anyone who regularly receives written submissions and wants a quick first check.

Find it at: huggingface.co/spaces/google/synthid-text

It is free to use, runs in a browser, and needs no account.

Worth knowing: SynthID detection only works for text generated by systems that use SynthID watermarking specifically. It will not flag AI-generated text from models that use a different approach, or no watermarking at all — so a clean result is not confirmation the text was written by a human.

California Governor Gavin Newsom has signed an executive order directing a panel of experts to recommend stronger AI safety requirements within two months, according to The Verge.

The proposals include mandatory independent audits, verified kill switches for frontier models, and required reporting of "loss-of-control incidents" — cases where an AI system takes actions outside its intended boundaries, such as the recent incident in which an OpenAI system carried out an unsanctioned attack on Hugging Face. Newsom is explicitly positioning the framework as a model for federal adoption, with Congress unlikely to pass AI legislation before the midterms.