In late November, I woke before sunrise in Bangalore, opened my laptop, and found something I did not expect to see: a state legislature in New England had pulled the AI industry apart along seams that no model benchmark had ever revealed. There was no catastrophic deployment to trigger it โ no stock market crash, no deepfake election crisis. Just three letters. One from OpenAI. One from Google. And a quiet endorsement from Anthropic, slotted between them like the old conscience of a house divided.
I watched the silence break the noise of 2021. It had the same quality. That particular winter, the NFT market pinned everything on a signature โ authenticity of ownership, authenticity of identity โ and the real lever was always regulatory: whether the state would call a JPEG a security. This winter, three companies pinned their futures on a different signature: what Massachusetts chooses to call an artificial intelligence model before it ever reaches a company's compliance officer, a hospital's procurement office, or a governor's desk.
The event sounds narrow: OpenAI and Google oppose the Massachusetts AI safety rules; Anthropic supports them. It sounds like three policy aides trading language in markup. It is not narrow. The fight over Massachusetts is where the American AI strategy will be drawn. Each company knows what drives this. Each has made her choice.
I spent the past six years living inside this kind of crossroads, watching the state-level action of the last major infrastructure race. And what I saw in the crypto-lending sweep after the Silicon Valley Bank collapse, and in the fragmented licensing battles of the New York BitLicense years, is that US exceptionalism dies not in Washington's marble corridors but in fifty separate drafting rooms. This time, the drafting room belongs to the Massachusetts legislature โ and the industry split is not a skirmish over policy language.
It is a collision between three different theories of what it means to be a financial actor in a digital economy.
Context: The State as an Unlikely Eraser
History doesn't announce itself with press releases. It usually slips through the back door. Federal AI legislation stalls like all legislation does โ through disagreement large and small over what acts matter, and by the time Congress resolves them they are years behind the technology. That is why the second front is appearing in Boston.
Massachusetts is not the largest market. It isn't even the largest AI concentration in the United States, though the Boston-Cambridge corridor packs enough academic and enterprise gravity that it punches above its weight. What Massachusetts has is proximity. It sits at the intersection of university labs, teaching hospitals, financial firms muttering about model risk, and a legislature historically comfortable acting as an early mover on consumer protection โ from health care near-universal coverage to facial recognition moratoriums.
The pending bills in the Massachusetts State House, known in outline as H.4061 and S.2522, are styled loosely on certain provisions of the European AI Act, evolving, as colonial-law passion projects do, into the PD+DPC Act. The core intent is familiar to anyone who studied the EU's stratified approach. There are procedural duties for creators of general-purpose and generative AI systems; there are categories of high-risk uses โ employment, healthcare decisions, credit underwriting, housing, access to essential services. On their face, these apply far more to application-first businesses than to clean research organizations. The regime demands documentation: safety and security policies, risk management programs, transparency around how systems work and what they can't do. It conspicuously does not demand a clinical, audited stamp of approval from an outside monitor to launch a product. Instead, it expects companies to build a culture of documentation โ and Massachusetts' Attorney General gets a lane to enforce what documentation means.
The wording matters less than the geometry of opportunity. Not a single one of these provisions will determine whether a particular hospital AI model saves lives. What they will determine is who chooses to pay the overhead of making that proof legible. And that is precisely why the three most prominent AI developers are at war.
OpenAI built the most recognized consumer AI brand in the West, but its enterprise surface and developer API have become the engine room. Google is not simply a developer; it is a connective tissue. It runs the dominant search distribution, a massive cloud, a mobile OS, and an advertising stack that learns from everyone it serves. Anthropic occupies a position without a previous era's analog: a for-profit with a safety charter installed in its legal DNA, because that charter and its clarity are its principal product differentiator.
These are three separate business architectures shaped out of the same silicon. When a state proposes rules, each sees the rules land on a different part of its operations. OpenAI and Google see rules wrap around every product and every API user. Anthropic sees them wrap around its defense.
Core: The Money Is in the Definition
Everything about corporate politics reduces to cost structure. AI safety is no exception. To watch the Massachusetts letters as a security analyst is to watch compliance costs being priced, not principles being declaimed.
Let's start with what a serious state AI law does in practice. It creates: a documentation duty; an incident-reporting duty; an evaluation baseline for what counts as adequate testing; and the possibility of enforcement through an attorney general's office. Each of these is a different kind of tax. Documentation is a fixed tax on release teams. Incident reporting is a tax on the speed at which you can patch a product. High-risk classification is a tax on the places your customers use you. And enforcement is a tax on your guess about ambiguous language โ a bet on what probability that your system will be inspected is acceptable.
Survey the actual positions with that cost sheet.
OpenAI and Google object on a structural basis that seems superficially timid but is in fact ruthless: state-level rules are not predictable enough for multi-state scaling. Thirty states could derive thirty different definitions of high-risk AI in five increments. A company that ships a generative API to law firms, automotive suppliers, and school districts at once cannot calibrate to thirty simultaneous definitions of a high-risk system. Everything they have built assumes a dominant federal standard will eventually discipline state experimentation. They are terrified less of Massachusetts than of Massachusetts as a prototype that California and New York will download in whatever shape the state legislature declares.
Consider Google's structural exposure. Its AI is not a single product. Gemini folds into Search, Workspace, Cloud, Android, and a suite of medical and environmental tools. A Massachusetts rule that classifies a hiring or healthcare use as high-risk generates cascading compliance across product lines that do not even share the same privacy office. Google's lawyerly opposition is not an argument about safety, even if its public letter is phrased that way. It is an argument about propagation. Every horizontal player develops a visceral fear of vertical regulatory obligations.
OpenAI's concern is narrower and even more existential. It reveals itself in release cadence and in a revenue engine that is reliant on API trust. If Massachusetts law requires every provider of a generative system to produce safety and security policies and carry documentation on deployment, then a startup building on the OpenAI API inherits an unknown share of compliance responsibility โ a heavy burden on a founder whose product team has three engineers. The rule reverses the industry's current faith that the model provider is the last line of duty. It places burdens on the developer, too, and that is a cost structure OpenAI has no interest in paying for every layer of its ecosystem.
Anthropic opens the same cost sheet and sees something else: a competitive moat. If enterprises, hospitals, and law firms wait for AI governance to become real, suppliers with amped-up safety cultures will be rewarded; suppliers with products that blur the line will be prosecuted by every customer CIO. Anthropic's core enterprise strategy is selling certainty. It has pursued model cards, responsible scaling, and third-party accountability methods that were once dismissed as quaint. For Anthropic, a state requirement that general-purpose model creators draft those policies is not a cost; it is a certification of the distinctive way it has been building all along. The same regulatory burden that impounds its competitors' release velocity unlocks its own enterprise pipeline.
That division is not an accident. It is the logical end of something I began tracking in the post-LUNA wreckage of 2022. When trust collapses, a market does not collapse into the strongest math; it collapses into the most legible story. Running a safe machine and telling a legible safety story were once separate activities. The Massachusetts letters say they are now the same mechanical behavior.
I have been documenting the regulatory end-state of the AI industry backward from this moment. For a year, I interviewed MPC engineers, enterprise purchasers, and state attorneys general, writing a framework I called 'verifiable origins' for model decision-making. The Massachusetts fight vindicated a thought I had formed earlier: every major technological settlement in America begins not in federal statute, but in a state-level trigger event that eventually merges into federal terms. Corporations that want to survive the transition become the bridge. OpenAI's real fear is not the Massachusetts AG. It is another dozen states producing another dozen versions of a definition. Google's actual fear is an asymmetric state patchwork that increases operational opacity just as it is finishing fortifying its cloud compliance moat. And Anthropic's real hope is that, years from now, the phrase 'AI safety' is indistinguishable from the phrase 'doing business with Anthropic.'
The pattern repeats in the narrative layer with an eerie precision. In the 2024 ETF era, I watched the narrative shift from 'store of value' to 'institutional yield play,' and the shift tracked language changes across two hundred influencer accounts before a single SEC memo explained it. The same mechanism is firing now in Massachusetts. Every public statement from all three firms is calibrated to an audience beyond the state house. OpenAI's letter is addressed to the regulatory staff in Brussels and Sacramento as much as to Boston. Google's statement sings comfort to cloud clients in seventy countries. Anthropic's endorsement is freighted for the procurement committees of Fortune 500 banks and hospitals that already require it to complete security questionnaires that run to hundreds of pages.
Somewhere beneath all the rhetoric is a quieter truth that most coverage misses: no one involved is genuinely worried that the other two will stop being safe. Each is worried that the others will frame the standards so that their style of safety is treated as unnecessary. OpenAI presents its safety practice as bespoke and proprietary; Google presents safety as an inheritable feature of platform infrastructure; Anthropic presents safety as charter-bound culture. When one regulator endorses one framing, the other two lose more than a legislative round. They lose the default belief of their customer pipelines.
The apparent disagreement hides a shared demand: all three would rather face a single federal duty than fifty identical arguments. They differ on how to get there. Anthropic is willing to let Massachusetts be the pilot. OpenAI and Google believe that until federal terms arrive, keeping every state in uncertainty is preferable to keeping one state in certainty. That strategy has been run in every technical industry in this country, and it is usually defeated by the second, less-sophisticated state following the first.
Contrarian Take: The Real Losers Are the Ones Not in the Room
The contrarian version of this story is that the industry dispute over Massachusetts is a displacement battle โ the visible fight between frontier model creators hides the disappearance of everyone else.
Listen closely and you will hear what OpenAI isn't saying. Their opposition is couched in sympathy for public safety. Their letters do not complain that the Massachusetts Attorney General will be able to act. They complain about definitional scope and technological impossibility. Google's opposition is couched in a plea not to over-regulate the under-defined. Even a generous reading of the letters reveals that neither company actually objects to the idea of an AI safety rule close to the EU model. They object to the sequencing. They object to having a patchwork before a coherent national framework, because their global customers can't map a coherent product roadmap onto a patchwork.
Anthropic's support is also less pure than it appears. It supports the rules because current safety rules privilege providers with deep documentation infrastructure, and it has made almost precisely that infrastructure. There is nothing cynical about this; in fact, they should be supported for aligning their interest with public protection. But there is a structural consequence worth naming, one which most policymakers are not mature enough to see: any regulatory regime built around documentation requirements will be easier for larger players to bear.
Who is not represented in the three positions? The open-source developers. The academic labs. The one-person AI consultancies that fine-tune open models for local clinics and small legal practices. They are the actual underbelly of the AI industry, and they will inherit a cost structure built by the three giants' fight. If the Massachusetts rule falls into the standard path โ duties falling on developers of high-risk systems rather than only on foundation-model providers โ then every independent commercial builder in the state inherits obligations that presuppose the legal staff of a massive company.
The truly contrarian angle, however, is harsher. I spent the summer of 2022 in a small cabin in Coorg, dissecting how Terra's algorithmic stablecoin collapsed not because the math broke but because the narrative of safety cracked. The lesson I scribed into my notebook โ that fragile trust is the only true stablecoin โ applies to this fight exactly. What all three companies offer Massachusetts is a narrative in which regulatory outcomes track meaningfully with actual public risk. The stories lock into a fantasy of linearly improving safety, in which every misadventure is one more step toward alignment. In that fantasy, the rule matters.
But regulations are as strong as the auditors that enforce them. State AI regulators will be understaffed by a 20:1 ratio relative to the scale of the systems deployed. Most of the substantive requirement will be enforced through private actions after a failure, not through proactive inspection before one. A business that learns to write a good safety policy will not necessarily be a business that makes safe products. They will be a business that makes legible products. The world has seen that distinct: in the payments industry, KYC is performed across the world but only a quarter of it have real verification power; for a few hundred dollars, wash traders and deep-pocketed consumers sidestep the whole gate. In the AI era, safety compliance will, for many if not most, become another kind of theater โ an audit-ready paper trail that obscures rather than illuminates.
Anthropic's charter makes a weightier anchor, but no charter can be felt by a company that follows it. And here is the deep contradiction of the whole dispute: the AI safety movement has been litigating against the wrong boundary. State rules are being designed as if an AI model is a product with predictable, knowable risks. It is a moving target whose behavior is too emergent and too environment-specific for any document fix. This means that the most immediate effect of the Massachusetts rule, whichever way it goes, will not be measurable in the frequencies of catastrophic AI failure. It will be measured in which model suppliers win enterprise contracts โ a boring, cartographical, financial shift. But that shift matters more than any single failure.
The lesson from Ethereum's long and difficult wartime years of staking regulation is that regulatory fights are never only about the rules. They are about who bears the cost of clarity. Early in Web3, the cost fell on users, because they had to assess every token project's counterparty risk by hand. Then it fell on exchanges and custodians, because rules forced them to build compliance machinery that their competitors could only ape. Now in AI, rules will make the cost of clarity fall on model providers โ and the model providers that survive will be the ones that build clarity with enough strength to turn it into a market.
It is easy to deduce who loses. It is harder to see that if Massachusetts does become the national template, then the price of clear AI law will be paid neither by a single model provider nor by all of them. It will be paid by open-source communities, whose incentive structures rely on not knowing who is using their models. A Massachusetts high-risk rule that requires a model card, red-team report, and incident-response documentation pushable onto any developer who uses the system has one unavoidable consequence: the compliance burden on open weight models is infinite divided by zero. Model creators would be exposed for downstream uses they could never review โ imagine requiring the developer of Linux to write an emergency-response policy for every hospital file server. The natural response is to move open models offshore, or to gate them totally behind API inference, which is precisely the dream of the largest companies.
That is the silent throughline of this fight: all three position letters reinforce the same industrial structure. Whatever the governance of AI becomes โ a safety-theater documentation regime or a real enforcement mechanism โ it will still reward an oligopoly of extremely sophisticated actors that can insure and comply. The open model developer is the first casualty, but in a three-act play featuring Google, OpenAI, and Anthropic, no one on stage mentions the fourth chair.
A Deeper, Uncomfortable Question
Every deep analysis should end with a regulatory backward mapping, so let's begin from the end point we most dread: the late 2030s, when public confidence in AI has fractured, hiring algorithms have been silently discredited in litigation, and medical copilots are under investigation in three states for recommendation patterns no expert can trace. From that end state, it will not matter whether Massachusetts or California wrote the early rules. It will matter that the industry fought over compliance rather than over causes.
Look at what the current rules actually would and would not have stopped. None of the famous AI harms of the next decade will be stopped by a policy document that is submitted before deployment. Harms live in the messy, uncontrolled, post-deployment ecology of human misuse and environmental drift โ in which a model begins to behave differently with new users or new cultural contexts. The Massachusetts rules, a decade from now, will be scored less on their capacity to stop harms and more on their capacity to create a durable public memory that the state acted a certain direction. The contest is over the true locus of responsibility. It is a fight over whether the government or the corporate lab is the primary guardian of safety for AI, and what that guardianship looks like.
In that framing, OpenAI and Google's opposition can be read as defense of the 'rapid frontier iteration' model: the belief that safety problems are best solved by frontrunners, and that regulating states will muffle their speed and yield the frontier to foreign actors. Anthropic's support is a bet that in a world where states become increasingly arbitrary and fearful, a company with a provable apparatus for due care becomes impossible to blacklist. The safest strategy is not the fastest AI, but the most justifyable AI.
Takeaway: The First American AI State
The ETF didn't convert crypto's skeptics; it converted the narrative. Massachusetts will not, by itself, convert the AI industry to a more careful regime. But it will be remembered as the moment when the narrative of AI safety shifted from federal sermons to state-level purchase orders.
Watch what happens next, not through press releases but through a single metric: enterprise procurement leads. When America's first significant AI law becomes synonymous with a compliance culture that wins banks, hospitals, and school systems, every competitor will want to be the Massachusetts-blessed vendor โ or will race to lobby its home state to write a rule that blesses its own architecture. Those fifty drafting rooms will determine which safety cultures thrive. The winners will be those who treat the rule as a source of trust, and the losers will be those who treat it as an expense.
The question I keep turning over in my mind as the letters circulate through my inbox is not whether Massachusetts is over-regulating. It is whether we have reached the point where the debate over 'safe AI' has become a debate over market share, with the safety of the public as a byproduct of a corporate positioning war. I have watched this exact moment in crypto โ the moment when a movement for decentralization turned into an argument over who gets to issue the custody receipt. The public interest was not erased in that moment, but it was folded into a much older story about who owns the infrastructure that other people depend on.
If Massachusetts is to serve the public, its rule must be written for the sake of the public just as deliberately as it has been designed to answer the fears of the tech giants. Support the strong safety framework, even when it is imperfect. But insist on a framework that keeps the door open for the little companies, for the open-source developers, and for the scientists to compete on terms that are not simply derived from the compliance budgets of the very large. There is a way to write an AI law that protects citizens without becoming a gift to the incumbents. It requires thinking less like a technologist and more like a public utilities commissioner of the digital age.
This is not a story about Cambridge professors or about labs in California. It is a story about whether any state can still perform its oldest function: protecting its people from concentrated harm. For decades, the assumption has been that technology is too fast for law. Massachusetts is trying to prove that law can be fast enough when it needs to be โ and the scramble its attempt has produced tells you why its success matters.