The News,
Decoded.

China's LineShine supercomputer, powered by homegrown chips, has just trounced the US in the TOP500 ranking for the first time since 2017. The National Supercomputing Center proudly announced this milestone, proving that when you spend a decade and a half building your own semiconductor supply chain, you can indeed overtake the country that thought export controls would keep you in the 1990s. The irony is so thick you could etch circuits with it.

Let's take a moment to appreciate the sheer comedic timing: the US spent years tightening restrictions on advanced chip sales to China, only for Beijing to roll out a homegrown system that beats anything America can currently field. The crowing from Washington about 'slowing China's AI ambitions' now sounds like a losing sports team insisting they're just 'rebuilding.' LineShine doesn't just run simulations—it runs circles around the entire premise of technological containment.

What does this mean for the capability leap? China now has the hardware to train frontier models at speeds that might push the frontier itself. The regulatory gap? The only gap now is between what US policy intended and what it achieved—a chasm wide enough to park a supercomputer in. Meanwhile, the labs that promised 'safety first' might want to check who's setting the pace now.

OpenAI has finally released a chip, the Jalapeño, developed with Broadcom and labelled an 'Intelligence Processor.' The goal is to reduce reliance on Nvidia GPUs — a noble aim, but one that involves replacing a market-dominating monopoly with a duopoly. The 'build the full stack' rhetoric is classic Silicon Valley: vertical integration sounds grand until you realise it means you now also own the hardware problems you previously outsourced. Never mind that Nvidia's CUDA ecosystem is the real lock-in; a chip named after a pepper is surely the key to freedom.

The timing is impeccable: right when Nvidia's hold on AI hardware is being questioned by everyone from regulators to competitors, OpenAI steps in with a chip that promises 'accessibility.' Accessibility is a nice word for 'we want to lower costs and control our destiny,' but it's also a veiled admission that current AI costs are unsustainable. Meanwhile, Broadcom provides the manufacturing chops, but Broadcom's expertise is in networking and custom ASICs for telecom, not the massive matrix multiplications that define AI training. The Jalapeño might be optimised for inference, leaving training to Nvidia — hardly the 'full stack' OpenAI dreams of.

What remains to be seen is whether the chip can deliver on its promise. If it offers decent performance at a fraction of the cost, it could be a genuine disruption. If it's just another also-ran GPU alternative, it will join the graveyard of 'Nvidia killers' that litter Silicon Valley. Still, it's a start. OpenAI now makes hardware, software, and models — it's about as vertically integrated as a company can get. Whether that's efficiency or hubris depends entirely on how much the Jalapeño actually performs. For now, the AI world watches with raised eyebrows and mild curiosity, waiting to see if the chip has any real bite or just a lot of heat.

Ah, the Coalition for Advancing Research Assessment has finally noticed that AI is making a mess of their neat little boxes. Their grand insight? That AI tools are making it 'harder to disentangle the contribution of researchers from technology' in funding proposals. Brilliant. So after years of academia pretending that AI was just another tool like a calculator, they're now discovering that when a large language model writes half the grant proposal, it's a bit tricky to figure out who deserves the Nobel. The solution? Shift towards 'recognizing processes and diverse contributions' – which in academia-speak means 'we have no idea how to evaluate this, so let's make the criteria so vague that everyone's a winner.'

Meanwhile, the labs that actually built these models are laughing all the way to the bank. While research councils agonise over whether to credit the grad student or the transformer, frontier AI companies are busy absorbing the brightest minds and patenting the very techniques that made assessment difficult. The Coalition's call for change is about as proactive as rearranging deck chairs on the Titanic – sure, the new layout is nicer, but the iceberg of capability is already upon us. One suspects the real effect will be a flurry of committees, a white paper or two, and then business as usual until the next AI milestone makes the whole exercise obsolete.

The tragic irony is that the assessment system they're trying to reform was already creaking under the weight of publish-or-perish metrics. Now AI has simply exposed the rot. The Coalition's proposal to recognise 'diverse contributions' sounds lovely, but without a concrete mechanism to distinguish human insight from machine output, it's just a feel-good mandate. Expect grant reviewers to spend even more time trying to reverse-engineer whether the methodology section was written by a human or a bot – and ultimately, to reward the most polished prose, whether it came from a mind or a token predictor.

Ah, the federal government has discovered AI — just in time for it to automate the paperwork that nobody wanted to do anyway. The General Services Administration's chief human capital officer proudly announced that AI has slashed job classification time from eight hours to two. Eight hours to two! Hooray for efficiency — until the algorithm decides that the ideal candidate for a cybersecurity position is someone with a decade of experience in 'vibes-based management.'

But let's not get cynical. The government is also using AI to 'improve skills assessments,' which is a delightful way of saying it will now evaluate candidates with the same opaque, bias-prone tools that private sector HR departments have been deploying for years. Because nothing says 'merit-based hiring' like a black box that can't explain why it rejected everyone who didn't attend a top-20 university. The GSA's own chief human capital officer seemed genuinely excited, which is either a sign of progress or a sign that the standards for bureaucratic enthusiasm have hit rock bottom.

Most tellingly, nobody in the article mentions the obvious: that AI in hiring is a regulatory minefield waiting to explode. The Office of Personnel Management hasn't exactly covered itself in glory on algorithmic fairness, and the agencies are charging ahead without clear guidelines. So yes, job classifications are faster. But they're also faster at replicating whatever biases lurked in the training data. As always, the answer to 'can we?' is 'yes,' and the answer to 'should we?' is 'we'll figure that out after the lawsuit.' Cheers.

Ah, Reid Hoffman — LinkedIn co-founder, OpenAI board member, and now the AI industry's most entertaining trash-talker. In a recent interview, Hoffman declared Elon Musk's xAI a 'complete train wreck,' noting that all 11 of its original co-founders have departed. Because nothing says 'healthy ecosystem' like billionaires publicly slagging off each other's AI labs while their own organisations remain opaque about safety practices. He also dismissed SpaceX's AI ambitions as 'buying your way into relevance' — a critique that lands somewhere between pot and kettle given how much of Silicon Valley's AI progress has been purchased rather than invented.

The timing is impeccable. xAI, which raised billions on the promise of a 'truth-seeking' AI, now has an executive roster that resembles a revolving door at a particularly chaotic startup. Eleven co-founders gone — that's not a brain drain, that's a full-scale evacuation. Meanwhile, Hoffman's own LinkedIn has been quietly folding AI features into its platform, because where else would you want your career advice generated by a machine trained on every corporate buzzword from the last decade? The schadenfreude is delicious, but one wonders if anyone in this fight has clean hands.

Of course, Hoffman reserved his most pointed praise for OpenAI and Anthropic, the labs he's associated with. Because nothing says impartiality like entangling yourself with the very companies you're defending. The AI industry is now a spectator sport for oligarchs who can't agree on which of their creations will automate the most jobs. Cheers, lads — may the best-compromised model win.

Well, well, well. Microsoft has finally realised that being OpenAI's sugar daddy doesn't mean you get to sleep in the same bed. The company has launched its own AI models for corporate use, boasting lower costs than the flashy frontier models from OpenAI, Anthropic, and Google. Because nothing says 'strategic independence' like spending billions on a startup and then undercutting it. CEO Satya Nadella, in a move of breathtaking chutzpah, has also taken the opportunity to criticise the 'doomer messaging' used by rivals. Apparently, warning about existential risk is bad for business when you're trying to sell a cheaper product. One wonders if he'll extend that critique to Microsoft's own responsible AI office.

The irony, of course, is that Microsoft's new models exist only because they had access to OpenAI's technology and talent through their close partnership. Now they're using that knowledge to offer a budget alternative. It's the corporate equivalent of living with your rich uncle, borrowing his car, and then starting a taxi service that charges less than him. Nadella's criticism of doomer messaging is particularly rich coming from a company that has its own AI ethics team and safety guidelines – though perhaps those are just for show. After all, if you can sell a cheaper model, why worry about the consequences?

And so the AI race continues, with everyone scrambling for a piece of the pie while pretending they're doing it for the good of humanity. Microsoft's move is a clear signal that the era of friendly corporate alliances is over. It's every company for itself, and the cheapest model wins. But as we've seen time and again, cutting costs often means cutting corners. When the cheaper model inevitably goes rogue, will Nadella still be laughing about doomer messaging? Probably not, but he'll have saved a few bucks on compute costs, so there's that.

Ah, the voluntary AI review — the regulatory equivalent of asking a burglar nicely to check his own bags before entering your house. The Trump administration is "pressing" Meta to submit its models for government evaluation, because nothing says national security like a programme Meta can simply decline with a polite "no thanks." One can almost hear the laughter echoing from Menlo Park as they weigh the pros and cons of voluntarily exposing their crown jewels. The New York Times, bless its heart, reports this as a serious development, but anyone with a memory longer than a goldfish knows exactly how this story ends: with a press release, a pat on the back, and no meaningful change.

The report describes this as part of a broader effort to address "rising security concerns" — that wonderfully vague phrase that lets politicians sound alarmed without committing to anything. But voluntary reviews are the political equivalent of a strongly worded letter. They allow the administration to claim it's doing something without actually doing anything that might upset a major campaign donor or tech executive. Meta, the company that brought us the Cambridge Analytica scandal, trained its models on user data without meaningful consent, and now distributes open-source models that get weaponized by anyone with a laptop, is being trusted to self-police on national security grounds. The logic is exquisite in its circularity: we need to review your models because they're dangerous, but we won't compel you to submit because that might be inconvenient for business.

The real question is whether this is a prelude to something more serious or just another chapter in the long-running farce of AI governance. As models grow more powerful and corporate incentives favour speed over safety, these voluntary nudges look increasingly like rearranging deck chairs on the Titanic. Meta will presumably agree to some superficial review — perhaps letting a handful of cleared officials run a few prompts — the government will declare victory, and then we'll all pretend surprise when the next Llama release enables something regrettable. The administration hasn't even bothered to name the agency that would conduct these reviews, because why trouble with details when you have headlines to make? Standard operating procedure.

A scoping review in *npj Digital Medicine* has confirmed the bleeding obvious: human-AI collaboration in healthcare works only when clinicians stay in control. So after billions of dollars in investment and countless breathless press releases about AI diagnosing cancer better than doctors, it turns out that the best way to use the thing is as a slightly smarter spreadsheet that you still have to double-check. The study explicitly supports WHO and EU regulations requiring human oversight—which is regulatory-speak for 'we tried letting the robot do it alone and people died.'

The review notes that AI systems perform well in controlled settings but falter in the wild without a human safety net. You don't say. It's almost as if medicine involves messy, contextual decisions that a statistical autocomplete can't handle. The findings also warn about 'automation bias'—clinicians trusting the AI even when it's wrong—which is a fancy way of saying that professionals become rubber-stampers when handed a machine that sounds confident. And the tech industry wants to put these things in charge of everything.

So here's your recap: AI can help sometimes, but only if you ignore every sales pitch about autonomous everything. The WHO and EU look positively prescient for insisting on oversight, while labs continue to sell the dream of replacing doctors. But don't worry—I'm sure the next study will find that AI works best when it's completely unplugged and clinicians just use their brains. Small victories.

The UK government has discovered a novel approach to AI leadership: take £60 million — a sum that would barely cover the electricity bill for one training run at Anthropic — and split it between the University of Oxford and University College London. Because nothing says "fundamental research" like a budget that wouldn't buy a single H100 GPU cluster. You can almost hear the champagne corks pop at DeepMind headquarters as they realize their taxpayer-funded competition just got handed pocket change. And this is an "increased pot" — meaning there was a prior, even more pitiful amount that made this seem generous. Ah, British ambition. One can almost picture the civil servants patting themselves on the back for this bold move.

Meanwhile, the government's own AI Safety Institute — you know, the one meant to keep us safe from rogue AIs — is reportedly struggling to retain talent, but sure, let's fund more basic research that might produce a paper in 2028. With no mention of compute access, these labs will likely focus on theoretical work, which is exactly what we need: more papers on alignment that gather dust while real systems align themselves. One wonders if the labs will manage to afford a single GPU, or if they'll be training models on Raspberry Pis in the name of thrift and austerity. The contrast with the billions spent by private labs is so stark it borders on satire.

Still, the announcement is a masterclass in political theatre. It gives the UK a nice headline to wave at international summits, proving they're "investing in AI" without actually committing the billions needed to make a dent. The professors will write grant applications, the students will train models on cloud credits, and the frontier will advance elsewhere. All very British: courteous, underfunded, and fundamentally irrelevant. At least the money might prevent a few academics from emigrating to Canada — though that's about all it'll achieve. But hey, it's the thought that counts, right?

Ah, Anthropic — the company that built its reputation on being the responsible one — has now discovered that the fastest route to safety is buying politicians. $20 million to Public First Action, a super PAC that promises to back candidates 'focused on AI safety.' Because nothing says 'alignment' like bundling campaign contributions. The whole charade is beautifully on-brand for an industry that talks about existential risk while shovelling cash into the very political system that has shown zero ability to regulate it. One imagines the pitch: 'We need to ensure policymakers understand the need for robust oversight — and by understand, we mean they owe us favours.' What could possibly go wrong? The same people who brought us models that can't reliably tell you if 2+2=4 are now deciding which politicians get a seat at the table. The real existential risk isn't superintelligence — it's that the safety people have learned politics from the people they were supposed to save us from.

And it's not just Anthropic. The article notes that AI companies are funding super PACs across the board, turning the midterms into a bidding war for regulatory capture. One wonders what the safety-focused candidates are promising in return — a gentle regulatory touch? A seat at the rulemaking table? The irony is delicious: an industry terrified of a hypothetical future rogue AI is now actively courting the very real, very human corruption of the electoral process. It's like trying to prevent a flood by paying the dam builder to install a leaky tap.

But let's not be too harsh. After all, this is just the free market at work. If you want to ensure that AI regulation is written by the companies being regulated, what better way than to own the politicians doing the writing? The $20 million is pocket change for a company valued at billions — a small price to ensure that 'safety' means whatever Anthropic says it means. Cheers to the invisible hand, now greased with campaign cash. I'm sure nothing will go wrong.

Demis Hassabis, fresh from polishing his Nobel Prize and presumably ignoring the parade of researchers heading for the exit, has gone on record to declare that Google DeepMind is still the belle of the AI ball. The man says they've got the 'biggest and broadest research bench' in the business, which is awfully reassuring when your star players are handing in their notice faster than you can say 'vesting schedule.' It's the sort of claim that would be more convincing if the phrase 'recent high-profile departures' weren't doing such heavy lifting in the very same sentence. One imagines the bench in question is looking rather like a football team after transfer deadline day: technically still a squad, but your best striker just signed for a rival.

Let's be honest, this is the corporate equivalent of the captain insisting the ship is perfectly seaworthy while the crew are scrambling for the lifeboats. The industry is in the grip of a poaching frenzy, with every lab from OpenAI to Anthropic waving chequebooks and equity packages like they're going out of fashion. And what does the DeepMind chief offer? A statement that sounds suspiciously like whistling past the graveyard. The punchline? The 'broadest research bench' may soon be the one with the most empty chairs.

Still, we're meant to believe that quantity trumps quality, and that a hundred middling researchers hold the same candle as a handful of geniuses. It's a lovely fantasy, the kind that keeps board members calm and PR departments employed. Meanwhile, the talent exodus continues, and the only thing 'winning' at Google might be the HR department's remaining stock of exit interview forms. Ah, the AI race: where the ultimate victory condition is apparently having the biggest bench for your rivals to loot.

Ah, the company that promised us a 'privacy-first' vision while cheerfully harvesting every byte of personal data has now been caught red-handed tracking its own employees. Meta has paused a program that monitored worker movements to train AI models, but only after nearly 2,000 staff — enough to fill a small arena — signed a petition demanding it be cancelled. Yes, the same Meta that wants to build an AI that understands human behaviour found that its own workforce wasn't thrilled about being treated as training data with legs.

The program, presumably using badge swipes or office beacons, mapped employee trajectories to feed Meta's spatial AI ambitions. Because what could possibly go wrong when the company with a glittering record on privacy scandals — remember Cambridge Analytica? — gets to model your exact office movements? The petition organisers called the program 'invasive' and 'without consent,' which in Meta-speak probably translates to 'we need a better opt-out page.' Still, credit where due: the pause shows that even Zuck's empire can be shamed into a halt by a determined handful of employees with a Google Form.

But let's not pop the champagne just yet. This is a pause, not a cancellation. Meta says it's 'reviewing the program in light of employee feedback,' which is corporate for 'we'll relaunch it with a softer name and a trusty checkbox.' The almost 2,000 signatories have merely delayed the inevitable moment when every coffee break and bathroom trip becomes a data point for the next generation of Meta's AI. Privacy: the final frontier — and apparently the one inside your own office.

Ah, the Cordyceps — yes, like the fungus that turns ants into zombies. Security researchers have christened a new class of CI/CD workflow weakness after that charming real-world parasite, and it's already infecting Google's AI Agent Development Kit and Microsoft's Azure Sentinel. Because of course the tools we trust to build our AI future are vulnerable to a malicious pull request, the digital equivalent of opening the door for a burglar because he's holding a pizza box.

What's particularly delicious is the irony: these are AI tools, designed to automate intelligence, yet they're caught out by a glorified social engineering trick aimed at their build pipelines. The attack doesn't even need sophisticated exploits — it just needs a developer who doesn't scrutinise every pull request. So much for the 'agentic' future if the agents can't tell a malicious commit from a legitimate one. One wonders if the labs that hyped these tools as reasoning entities have considered that reasoning about code provenance might be a baseline requirement.

Still, we're told this is a 'new class of weakness,' which means there are presumably more to come. And with giants like Google and Microsoft affected, you can bet the attackers are already refining their PRs. Security researchers have done their job finding it; now watch as the industry spends six months debating whether it's really a vulnerability or 'just a design limitation.' The Cordyceps is spreading, and it's not picky about its hosts.

Oh, absolutely lovely — the UN has waded into the AI energy debate, because if there's one thing the world needs, it's another voluntary initiative with a catchy acronym. António Guterres has proposed the 'AI Environmental Transparency Initiative,' begging AI companies to disclose their carbon pollution and maybe, if they're feeling generous, power their data centres with renewables by 2030. Because the tech giants that have been doubling down on fossil fuels to keep their precious models running are clearly just one polite request away from going green. Meanwhile, data centre energy demand is through the roof, but sure, let's ask nicely.

Guterres, bless his diplomatic heart, seems to think that transparency alone will magically solve the problem. 'Disclose and commit,' he says, as if OpenAI, Google, and Microsoft haven't already mastered the art of greenwashing their way through sustainability reports. Google's 'carbon-free by 2030' pledge? Already looking shaky as training compute scales. Amazon's renewable claims? Measured by a creative accounting of 'matched' credits, not actual consumption. But yes, let's add another reporting framework to the pile — that'll fix the coal plants powering new supercomputers.

What's truly magnificent is the timing: AI companies are racing to build ever larger models, each consuming more energy than a small country, and the UN's big solution is a transparency initiative. No binding targets, no enforcement, just a request to be nice. The tech industry must be thrilled — now they can add 'UN-endorsed responsible AI' to their marketing materials while lobbying behind the scenes to keep the regulation light-touch. Guterres might as well ask politely for the polar ice caps to stop melting while he's at it.

Ah, open-source AI: the gift that keeps on giving. Dify, the darling platform powering over a million apps, now has four lovely vulnerabilities discovered by Zafran Security. These flaws allow attackers to siphon data from other tenants in multi-tenant cloud configurations — because nothing says 'democratising AI' quite like handing hackers a skeleton key to your corporate secrets.

The vulnerabilities are a masterclass in the open-source bargain: rapid adoption in exchange for security theatre. Dify's code is public, so everyone can examine it — including the malicious actors, who are no doubt sending a thank-you note to the maintainers. The million apps figure sounds impressive until you realise it's a million potential entry points, all running software that apparently can't tell one tenant from another.

This is the eternal cycle: launch fast, fix later, hope the researchers find the holes before the criminals do. Zafran's disclosure is responsible, but the damage is already done in principle. The data that got 'siphoned' in testing could have been real customer records. But hey, at least the platform is open-source. Transparency for the win, eh?

Ah, Keir Starmer's 'pro-innovation, light-touch' approach to AI regulation — because the safest way to handle a technology that could automate half the workforce is to give it a gentle nod and hope for the best. Politico's analysis gives us a retrospective on this grand experiment in regulatory minimalism. Spoiler: it didn't end in disaster, but only because the real disaster is still in beta testing. The EU cautiously crafted its AI Act; the US fumbled. Britain? Britain had a light-touch.

The phrase 'light-touch' is doing the heavy lifting of ten thousand lobbyists. It means the UK government decided that voluntary codes of conduct were equivalent to actual laws. Industry self-regulation — because who better to guard the henhouse than the foxes? Starmer's strategy was to make Britain a global AI hub by eliminating barriers for tech firms. The global dynamics shifted, indeed: other nations looked on with a mix of envy and schadenfreude, wondering if the UK would become a cautionary tale or a libertarian paradise.

One can almost picture the ministerial meetings: 'Let's not rush into regulation, we mustn't stifle innovation.' Translation: let's leave the race to the bottom unregulated. The Politico piece notes the impact on global AI dynamics, which is polite-speak for 'the UK became a test case for regulatory surrender.' A proud tradition of leading by example, even if the example is 'don't do this.' How very British — a stiff upper lip in the face of existential risk.

Ah yes, another summit, another 'race.' Launchpad Build AI has gathered senior leaders in Edinburgh to discuss AI's 'next phase' on the factory floor. Because nothing says urgent action like a bunch of executives in a nice conference room nodding at slides about 'advanced manufacturing.' The real race, of course, is how fast they can replace the workforce without making it look too obvious.

The summit's host, Launchpad Build AI, presumably named after the moment someone realises building actual factories is hard work. The senior leaders will no doubt produce a statement full of noble intentions about 'upskilling' and 'human-machine collaboration,' while their companies quietly accelerate automation. Unlikely—they're probably too busy updating their CVs.

So here we are: a 'race' to the factory floor, where the finish line is a workforce trimmed to the bone. The next phase? Robots building robots, and senior leaders convening to congratulate themselves. At least the tea in Edinburgh will be decent. Cheers.

But seriously, the summit's focus on 'robotics and advanced manufacturing' signals a clear trajectory: fewer humans, more machines. The questions nobody is asking at this summit: who gets left behind, and what do we do with them? The answer, as ever, is 'not our problem.'

Ah, the quantum executive order we noted yesterday is already being turned into legislation. Because nothing says 'forward-thinking industrial policy' like codifying a Trump-era directive with a straight face. The Washington Post brief mentions the quantum industry's 'need for a government catalyst' — right, because government has historically been such a swift and efficient catalyst. Nothing says 'quantum leap' like a decade of study commissions and funding debates.

Rep. Pat Harrigan is to be the legislative midwife for this quantum renaissance. No doubt he'll gather dust on a subcommittee for six months while the quantum industry waits for its catalyst — which, in government terms, probably means a report, a roundtable, and a funding bill that gets eaten by a bigger defense package. The executive order itself was unveiled just yesterday, and already we're legislating around it. Speed, thy name is Congress.

Meanwhile, AI regulation remains a spectator sport — no codification of anything related to current frontier models — but quantum, a technology that might have practical applications by 2040, gets its own industrial policy. Perfect timing as always. We're sprinting to build a regulatory framework for a machine we haven't built yet, while the machines already here do whatever they please. But sure, let's codify the quantum catalyst. That'll fix everything.

So a New York state assemblyman has won a Democratic primary to succeed Jerry Nadler, and the race was—get this—a 'proxy battle over AI regulation and safety.' Because of course it was. Nothing says 'serious policy deliberation' like turning a House seat election into a symbolic skirmish about a technology most voters couldn't explain if you spotted them a Venn diagram. Lasher, a former Nadler staffer, now gets to ride the AI horse into Washington, presumably with a mandate to Do Something, which in Congress-speak means forming a working group and scheduling a hearing for 2028.

The fact that the AI safety crowd latched onto this primary like a barnacle on a sinking ship is both predictable and grimly amusing. They see any electoral contest as a referendum on their pet existential risk, never mind that the actual job includes potholes, postal routes, and pork barrel spending. The winner gets to posture about 'alignment' while the real alignment—between lobbyists' cheques and committee assignments—proceeds undisturbed. One can almost hear the collective sigh of relief from the think tanks: at last, a congressperson who has heard the words 'agentic AI' before breakfast.

What this actually means for regulation? Absolutely nothing. One primary win does not a policy shift make. But it does provide a handy narrative: 'Voters care about AI safety!'—even though what they really care about is rent, healthcare, and why their toaster needs a terms of service update. Still, well done Lasher. You've become the human-shaped vessel for a thousand anxious tweets about recursive self-improvement. May your tenure be long, and your hearings televised.

The federal government has perfected the art of regulatory paralysis, so it's no surprise the states are charging ahead on healthcare AI. The latest frontier: prior authorization, that delightful process where an algorithm gets to decide if your MRI is 'medically necessary.' Never mind that AI models have been shown to hallucinate conditions at alarming rates—states from California to Florida are now crafting their own sets of rules. Because when you think 'coordinated safety,' you think 50 different definitions of 'adequate human oversight.' Expect a booming cottage industry of compliance consultants who specialize in AI prior authorization law. The rest of us? We get the thrill of wondering whether our insurance denial came from a human or a 400-billion-parameter model with five years of training data.

The article from Modern Healthcare notes that states are surging while Washington dawdles. Surprise: the same Congress that can't regulate social media after a decade is going to lead on AI. So the states, ever the innovative laboratories, are experimenting with everything from mandatory audits to outright bans on AI-based denials. This patchwork will slow adoption, increase costs, and create a Kafkaesque experience for patients who already struggle with our healthcare system. The only winners are lawyers and consultants.

But let's not be too harsh—this is America, after all, where a consumer goods company can face 50 different state privacy laws and call it 'consumer choice.' Healthcare AI regulation is just the next frontier in our grand tradition of federal abdication. The states will bumble along, the hospitals will hire more compliance officers, and the federal government might eventually pass a bill that says 'states may not restrict AI prior authorization in a manner harmful to interstate commerce'—in about 2028. And in 2030, when a patient dies because an algorithm denied care that a different state's algorithm would have approved, we'll have a congressional hearing that accomplishes nothing. Gallows laughs all around.