The Trial Starts Tomorrow. The Fake News Site Started Months Ago.
The Musk v. OpenAI jury trial opens Monday in Oakland — and thousands of pages of unsealed depositions from Altman, Sutskever, Brockman, and Nadella paint a portrait of a company that was already fracturing well before Altman was fired. OpenAI's super PAC appears to have been quietly funding a fake news site staffed entirely by AI bots to push pro-AI legislation. OpenAI also released a Privacy Filter — an open-weight PII detection model that runs locally. And an amateur mathematician used a single ChatGPT prompt to crack a 60-year-old Erdős problem that defeated every human who tried it.
Musk v. OpenAI starts Monday — and the unsealed depositions are already rewriting the origin story
The Elon Musk v. OpenAI jury trial begins tomorrow, Monday April 27th, in a federal courtroom in Oakland. U.S. District Judge Yvonne Gonzalez Rogers declined to dismiss the case earlier this year, saying the jury needed to weigh whether the people on the stand are credible — a framing that makes this as much a character trial as a contract dispute. Musk's central claim is that OpenAI abandoned its original nonprofit mission that he helped fund, and that the pivot to a for-profit structure constitutes fraud. OpenAI has consistently called the suit sour grapes from a co-founder who left and then built a competitor.
But the thousands of pages of depositions unsealed in the days leading up to trial have complicated both sides' narratives. Ilya Sutskever owned approximately $4 billion in vested OpenAI shares at the time Sam Altman was briefly fired in November 2023 — a figure that recontextualizes his participation in the board vote against Altman. Depositions from Altman, Sutskever, Brockman, Mira Murati, Satya Nadella, and former board members Helen Toner and Tasha McCauley were among those unsealed. Early internal tensions are visible throughout: Sutskever worried in 2022 that OpenAI was treating open-source AI as a "side show," and OpenAI's leadership was actively debating whether to prohibit investors from backing rival labs — a conversation sparked by Reid Hoffman's decision to co-found Inflection while still an early OpenAI backer. The trial is expected to last several weeks. For the first time, the people who built OpenAI will have to tell their story under oath.
theverge.com ↗OpenAI's super PAC appears to be funding a fake news site staffed entirely by AI bots — and it's been lobbying on AI legislation
A Substack investigation by Tyler Johnston has uncovered what appears to be an AI-generated fake news operation connected to OpenAI's political arm. The site, The Wire by Acutus, launched on December 29, 2025, and has published 94 full-length articles on AI policy, Senate races, pharmacy reform, nuclear energy, crypto regulation, and more — all in under four months. It has no masthead, no named editors, no bylines, and no public explanation of who runs it. One of its "reporters," Michael Chen, reached out to Nathan Calvin of the advocacy group Encode seeking comment on an AI bill in Tennessee. Calvin flagged the request as suspicious: the email was loaded, the only format offered was "written Q&A," and no public record of Michael Chen exists. When run through Pangram, an AI content detector, the email came back fully AI-generated. So did 69% of the site's 94 articles, with another 28% flagged as partially AI-generated.
The financial trail is the other layer. Johnston's investigation traced funding patterns that appear to connect Acutus to OpenAI's super PAC — the same political vehicle the company has been using to lobby against state AI regulation, including the Colorado law the DOJ joined a lawsuit against last week. The site's source code is also damning: a React app with exposed fields in the backend interface labeled "AI Background Context" and "Question Prompts" confirms that content is generated by feeding AI a topic and prompt, then publishing the output under a fake reporter's byline.
modelrepublic.substack.com ↗OpenAI open-sources a Privacy Filter — a 1.5B-parameter PII detection model that runs entirely on your machine
OpenAI released Privacy Filter, an open-weight model for detecting and redacting personally identifiable information in text. The release is free, locally runnable, and specifically designed for production privacy workflows — the kind of infrastructure that sits between raw data and training pipelines. At 1.5 billion total parameters with 50 million active, it's small enough to run on modest hardware while delivering what OpenAI claims is state-of-the-art performance on the PII-Masking-300k benchmark. It supports 128,000 tokens of context and processes inputs in a single forward pass rather than token by token, which makes it fast enough for high-throughput use cases like logging redaction, pre-training scrubbing, and document indexing.
The model classifies text across eight categories: private_person, private_address, private_email, private_phone, private_url, private_date, account_number, and secret. The last two categories are notably broader than typical PII detection — account_number covers credit card and bank account data, while secret covers passwords and API keys. OpenAI says it uses a fine-tuned version of Privacy Filter in its own internal privacy workflows. The key architectural detail is the local execution model: because Privacy Filter runs on device, PII can be masked or redacted without the data ever leaving the machine — a meaningful difference from cloud-based PII APIs where the data must be transmitted to be processed. The weights are released on Hugging Face under an open license for fine-tuning.
openai.com ↗A 23-year-old with no math training used a single ChatGPT prompt to crack an Erdős problem that stumped the world's best mathematicians
Liam Price, 23 years old, no advanced mathematics training, submitted a solution to a 60-year-old problem posed by Paul Erdős that had defeated every mathematician who tried it — including Jared Lichtman, a Stanford mathematician who had already proven a related Erdős conjecture as his doctoral thesis. Price got the solution from a single prompt to GPT-5.4 Pro on "an idle Monday afternoon," submitted it to Erdosproblems.com, and forwarded it to his occasional collaborator Kevin Barreto, a Cambridge undergrad. Experts who reviewed it confirmed it was correct — and more importantly, that the AI had found a novel method no human had thought to try. Terence Tao, widely regarded as one of the greatest living mathematicians, noted that every previous attempt had "made a slight wrong turn at move one" and that the AI had simply... not made that turn.
The problem concerns "primitive sets" — collections of whole numbers where no number in the set divides any other. Erdős conjectured that the "score" of these sets approaches a specific minimum as the numbers in the set approach infinity. The conjecture had been open since the 1960s. The specific Erdős sum framework had already been the subject of serious work by Lichtman and others, but a particular sub-problem remained stuck. What appears to have happened is that the AI, unburdened by the accumulated intuitions and conventional approaches that human mathematicians carry, simply took a different path. Tao's framing was careful: "What's beginning to emerge is that the problem was maybe easier than expected, and it was like there was some kind of mental block."
scientificamerican.com ↗Sunday's brief, so I'll be direct: this week has been a stress test of how much trust the AI industry has actually earned, and the results are mixed in a way that should bother people who care about how this goes.
The Erdős story is the thing I want to end on because it's the most honest signal. A 23-year-old with a ChatGPT subscription and curiosity cracked something that decades of serious mathematical effort couldn't. That's real. The mechanism — an AI unencumbered by the accumulated bias of expert consensus — is real. It's the kind of thing that makes the case for optimism about what these tools can do when they're in the hands of curious people who aren't trying to monetize them or lobby with them.
The Acutus story is the shadow version of that. The same technology, deployed not to solve old problems but to manufacture fake journalists who email policy advocates and publish AI-generated lobbying content under the cover of "independent reporting." If the financial connection to OpenAI's super PAC holds — and the investigation is careful to say it's still being established — it represents the most concrete example to date of an AI company using AI to shape the political environment in which AI gets regulated. That's a problem that doesn't fix itself, and the trial starting tomorrow won't address it. Watching both of these things exist in the same week is a good reminder that "AI is powerful" is not a complete sentence. The rest of the sentence is: powerful for what, in whose hands, and accountable to whom.