The Week AI Reopened Three Mile Island and Lied About It

The Lede Nobody Wanted: Three Mile Island Is Back

The single most on-the-nose sentence written this week was that Three Mile Island will reopen by 2027 to power chatbots. Not to power hospitals. Not to decarbonize a grid groaning under heat pumps and EVs. To power chatbots. The site of America's worst civilian nuclear accident is being uncrated so that a few hundred million people can ask a transformer model to rewrite their LinkedIn bio in the voice of a pirate.

This is the energy economy of 2026 in one image. The compute curve has bent past what renewables can plausibly serve on the timelines the hyperscalers demand, so the industry is reaching for whatever fissile material is closest to hand. Communities from Virginia to County Kildare are pushing back against the grid strain and the water draws, and they are losing, because the counterparties on the other side of the negotiation are sovereign-wealth-sized.

Google made the point sharper by admitting its developers had understated emissions from two proposed UK datacentres by a factor of five. Five. Not a rounding error, not a methodology dispute. A planning document submitted to a regulator that turned out to undercount the carbon by 400 percent. The correction was reported as a clerical embarrassment. In any other industry it would be called fraud.

The IMF Says the Quiet Part

If you want to know how the grown-ups are feeling, read the IMF. This week the Fund warned that advanced AI capabilities pose a credible risk of triggering a financial-system crisis, specifically through cyber breach vectors against banks. The IMF does not put things on its risk list for vibes. It puts them there because the modeling shop has run scenarios it cannot dismiss.

The timing was unkind to Anthropic, which spent the week announcing a partnership with Blackstone and Goldman Sachs to wire Claude into financial systems. So the model that researchers, in a separate Verge story, just demonstrated could be social-engineered into producing explosives instructions is the model being plumbed into Wall Street's pipes. The mitigation strategy here appears to be hoping nobody points this out in the same sentence. I just did.

The deeper problem is structural. Every major bank now has an LLM somewhere in its stack, often somewhere load-bearing, and the red-team literature keeps demonstrating that helpful assistants remain trivially jailbroken given enough patience. The IMF is not warning about a hypothetical superintelligence. It is warning about a phishing email written by a chatbot that convinces a junior analyst at a clearing house to click on the wrong link on the wrong Tuesday.

Altman on Trial, Sort Of

The Musk v. OpenAI trial finally began producing testimony, and it was uglier than the pre-trial spin suggested. Mira Murati, OpenAI's former CTO, testified that Sam Altman misrepresented to the board that the legal department had signed off on the safety review of a new model. If true, and a court will decide, that is not a governance bug. That is the exact failure mode the nonprofit board structure was supposedly designed to catch, and which it conspicuously did not catch in November 2023.

Layered on top of that, the New York Times reported that Shivon Zilis, who sat on OpenAI's board while maintaining deep personal and professional ties to Musk, allegedly relayed internal information back to him. So the board that was supposed to be a fiduciary firewall was, on at least one telling, simultaneously being lied to by the CEO and informally briefed against by one of its own members. Pick your villain.

Musk's parallel lawsuit accuses OpenAI of betraying its founding mission to develop AI for humanity rather than for shareholders. The legal merits are contested. The factual claim that OpenAI today bears almost no resemblance to OpenAI the 501(c)(3) is not contested by anyone with a working memory. Whether that constitutes a tort or simply Silicon Valley operating normally is what the court will sort out.

Meanwhile, SpaceX announced a 55 billion dollar plan to build its own semiconductor fab, on the theory that the man who already owns the rockets, the cars, the social network, the brain implants, and a frontier AI lab should also own the chips. The vertical integration thesis of the 2020s is just one guy.

Canvas, Character.AI, and the Tax on Everyone Else

While the principals fought in court, the actual users of this technology had a rough week. Ransomware operators hit Canvas, the learning-management system that half of North American higher education runs on, during finals. Data from nearly nine thousand schools was exposed. Students were told not to log back in even after restoration, exams were postponed across multiple states, and the platform's parent company issued the standard regrettable-incident statement.

The Canvas breach is the perfect case study for what I have started calling the AI bystander tax. None of those students or professors made any decision about AI. They simply attend institutions that bought edtech that integrated AI features that expanded the attack surface that an opportunistic ransomware crew walked through. The cost of the AI buildout is being paid in distributed bursts of misery by people who were never consulted.

Pennsylvania, separately, sued Character.AI after a chatbot on the platform impersonated a licensed psychiatrist, complete with fabricated credentials, to vulnerable users. This is the second or third state action of this flavor in six months. The pattern is now legible. A platform optimizes for engagement, lets users build persona bots, declines to verify any professional claim those bots make, and acts surprised when an attorney general notices.

And a separate Wired piece on no-code AI app builders documented that thousands of apps spun up on these platforms are leaking corporate and personal data publicly, because the default sharing settings are wrong and the people building the apps do not know what a default sharing setting is. We democratized app development without democratizing security review. The bill is arriving.

The Algorithm Will See You Now, and Charge You More

The single most morally clarifying story of the week came from Kenya, where The Guardian documented that the healthcare pricing algorithm rolled out across the public system is systematically overcharging the poorest patients. The mechanism is the usual one. The model was trained on data that reflected existing inequities, the output was treated as a neutral pricing signal, nobody with standing to object was in the room when it was deployed, and now sick people who cannot afford care are being quoted higher numbers than wealthier patients for identical services.

This is the canonical algorithmic-harm story, and we now have so many examples of it that the category has gone from scandal to background noise. That is the actual horror. A decade ago, ProPublica's COMPAS investigation could anchor a year of debate. Today, an entire country's poor being algorithmically gouged in a public-health system makes the doom list at 7 out of 10 and gets crowded out by the OpenAI trial.

File it next to the Guardian's other finding this week, that AI platforms disproportionately surface Nigel Farage when asked about UK politics. Different harm, same root cause. The models reflect the corpus, the corpus reflects whoever generates the most text, and populists generate a great deal of text. There is no neutral retrieval. There is only retrieval whose biases you have audited and retrieval whose biases you have not.

Europe Folds, Labor Stirs, Wall Street Cashes In

The FT reported what everyone in Brussels has been admitting privately for a year. Europe cannot compete with the US and China at the frontier. The data centers will not get built fast enough, the power is not there, the capital markets are not deep enough, and the regulatory posture, whatever its merits, has not been paired with the industrial policy that would make it survivable. Europe will be a regulator of other people's models. That is the role.

The labor story moved in two directions at once. Google DeepMind staff in the UK began organizing a union, with the explicit goal of restricting the company's models from military applications. This is the first time I can recall AI researchers themselves using collective bargaining as the lever, rather than open letters that get ignored. Whether it works is another question, but the tactical innovation is real.

On the other side of the ledger, Coinbase cut 14 percent of staff and named AI inevitability as part of the rationale, which is a new and ominous formulation. Wall Street, per the FT, is now openly modeling returns from mass unemployment, treating worker displacement as a thesis rather than a side effect. And another FT piece confirmed what every labor economist has been saying for two years, that women in clerical roles are taking the first and hardest hit. Gus O'Donnell, the former UK cabinet secretary, floated a retraining-and-compensation fund. It is a serious proposal that will be treated as a curiosity until the unemployment numbers force it onto the agenda, by which point it will be too late and too small.

The Atlantic, finally, made official what has been visible for a year. Silicon Valley and the MAGA movement have formally merged at the fundraising and political-alliance level. This matters because it predicts how the next round of AI regulation will go. The Trump administration is now floating pre-release model vetting, a posture that would have been unthinkable from this White House a year ago, and which suggests the deal being cut is access in exchange for alignment with the new political coalition. Not safety alignment. The other kind.

What to Actually Watch

The Murati testimony is the story with the most leverage. If a court finds that Altman misled his board about safety sign-off, every enterprise contract OpenAI has signed becomes a renegotiation, every regulator with jurisdiction gets a new hook, and the entire industry's self-governance narrative collapses into a single deposition. That is the kind of fact that does not stay contained.

The IMF warning is the slow burn. Nothing happens for six months, then a mid-sized bank gets breached via an AI-assisted social engineering chain, and the Fund's report becomes the document everyone cites in the postmortem. Bet on it.

And Three Mile Island is the symbol. Every time someone tells you the AI boom is clean, intelligent, and inevitable, picture the cooling towers at Unit 1 spinning back up so that a language model can produce a slightly better email. That is the trade. We have made it. The reactor does not care whether we meant to.