Recently, jourôÙnalôÙists have sought comment from me on a series of unusual AI events. Last week, in a certain AI copyôÙright case, the US governôÙment filed a ãstateôÙment of interestã in support of AI training being fair use of copyôÙrighted works. (No comment.) This week, workers at a certain AI company posted gloomy messages on social media about the possiôÙbility of AI extinôÙguishing human life, so I was asked whether ãcrimes against humanityã have been committed by these AI compaôÙnies. (No comment.) MeanôÙwhile, op-eds in major US newsôÙpaôÙpers are calling for someôÙthing, anything to beô done.
Iãm an author, designer, programmer, and lawyer. In 2022, I learned that my own works were in the training datasets of generôÙaôÙtive-AI compaôÙnies. In response, I invented the first set of lawsuits chalôÙlenging the legality of these pracôÙtices. There are now 142 such cases in the US. Iãm currently co-counsel for plainôÙtiffs in eight of them. Though I discuss certain legal issues here, I am not your lawyer, and nothing here is held out as legal advice. These are my personal views; I speak only for myself.
In 2024, I said that ãan AI cataôÙstrophe arising from failure of alignôÙment is much more likely than one arising from sci-fi-style maligôÙnant agency of the AI.ã In some sense that predicôÙtion is ripening.
But media predicôÙtions about the nature of that AI cataôÙstrophe remain unhelpôÙfully rooted in sci-fi scenariosãwhat Iãve termed the Skynet fallacy. Unhelpful because these scenarios are primarily a vessel for fear. They donãt illuôÙmiôÙnate paths to realôÙistic policy change. This New York Times op-ed, for instance, asks us to imagine ãrogue A.I.s [that] hack out of their containerã and ãdesign a superôÙvirus that spreads unconôÙtrolôÙlablyã. The ãhack outã partãplauôÙsible. Itãs already happening. Designing a superôÙvirusãlessô so.
Still, taking the NYT op-ed as a template, letãs consider why pundit-friendly proposals for AI safety likely wonãtô work.
Shut it all down, now ãÎ The obvious way to prevent A.I. from killing everyone is to issue a global ban on A.I. research. The problem is that our society has already gambled more than a trilôÙlion dollars on A.I.ãs upside, so a ban would have ruinous side effects.
ãObvious wayããyes, in the vacuous sense of there oughta be a law! But in pracôÙticeãmuch easier said than done. No techôÙnology has ever been the subject of a preempôÙtive ãglobal banã of this nature. InterôÙnaôÙtional nuclear nonproôÙlifôÙerôÙaôÙtion treaties are probôÙably the closest analog. But they only arose after the US and other nations had competed over decades to develop nuclear weapons. And of course, these treaties did not call for complete nuclear disarôÙmaôÙment.
As a legal matter, ãshut it downã overôÙlooks that there are already state and federal laws that prohibit maliôÙcious softôÙwareãe.g., the Computer Fraud and Abuse Act, the ElecôÙtronic CommuôÙniôÙcaôÙtions Privacy Act, and others. At the federal level, the quesôÙtion is not whether we have laws that can address AIãwe do. The quesôÙtion is whether we have law enforceôÙment that will charge AI compaôÙnies with violating those lawsãwe donãt. Conversely, as long as federal law enforceôÙment remains supine, then enacting further laws is an empty gesture.
As a techôÙnical matter, ãshut it downã may itself be an empty gesture. The open-weight language models already in circuôÙlaôÙtion are likely powerful enough to cause wideôÙspread damage by suffiôÙciently motiôÙvated users. Halting AI research will have no effect on those threatsãexcept perhaps to embolden those actors.
The economic arguôÙment is salient, however. As I noted in March 2023, as a public-wealth matter, ã[t]he moneyãô expected to be returned from AI investôÙment ãhas already been spent.ã Here in 2026, a stagôÙgering amount of national capital is flowing toward AI. No nation would volunôÙtarily make itself poorer by acceding to a ãglobal banã on AI. AnthroôÙpolôÙoôÙgist Joseph Tainter predicted this effect in his 1988 book The Collapse of Complex SociôÙeties (which I wrote about). I summaôÙrized this particôÙular point: ãIn prinôÙciple, a nation could choose to decelôÙerate its own economic growth to foreôÙstall collapse in the future. But that would simply make itself vulnerôÙable to domiôÙnaôÙtion by another nation today. Such decelôÙerôÙaôÙtion would thereôÙfore be politôÙiôÙcally irraôÙtional.ã
Take an air[-]crash[-]invesôÙtiôÙgator approach ãÎô When an aircraft crash occurs, invesôÙtiôÙgaôÙtors from the National TransôÙportaôÙtion Safety Board are immeôÙdiôÙately dispatched to the site to gather forensic evidence, conduct interôÙviews and deterôÙmine the underôÙlying cause.
National TransôÙportaôÙtion Safety Board invesôÙtiôÙgaôÙtions have certainly led to air-safety improveôÙments. But the NTSB is not the primary source of aviaôÙtion reguôÙlaôÙtion in the USãthatãs the Federal AviaôÙtion AdminôÙisôÙtraôÙtion. The NTSB was estabôÙlished as an indeôÙpenôÙdent invesôÙtiôÙgator of transôÙportaôÙtion inciôÙdents partly so that the FAA would not be in the conflicted role of invesôÙtiôÙgating the effecôÙtiveôÙness of its own reguôÙlaôÙtions (or confronting its own politôÙical entanôÙgleôÙments). LikeôÙwise, an NTSB-like orgaôÙniôÙzaôÙtion that retroôÙspecôÙtively invesôÙtiôÙgates dangerous AIô inciôÙdents will have a very limited range of influôÙence without an FAA-like orgaôÙniôÙzaôÙtion that imposes and enforces operôÙaôÙtional and safety reguôÙlaôÙtions.
Monitor the situôÙaôÙtionô ãÎô like the systems we use for air traffic control ãÎô Researchers would be required, by law, to post public inforôÙmaôÙtion on who is conducting the training run and which data center is doing the training.
The air-traffic comparôÙison doesnãt hold. US airspace is a federôÙally reguôÙlated and managed resource (by the aforeôÙmenôÙtioned FAA). So inforôÙmaôÙtion about ordiôÙnary flights operôÙating within is public by defaultãsomeôÙtimes to the consterôÙnaôÙtion of aircraft-owning private citiôÙzens. Imposing similar public discloôÙsure on private US AI compaôÙnies using private US dataôÙcenôÙters would be legally diffiôÙcult. FurtherôÙmore, in the future, more AI models will be trained for national-secuôÙrity uses. These will be among the most potenôÙtially dangerous AI models. But they will be exempt from public discloôÙsure on national-secuôÙrity grounds, lest these dataôÙcenôÙters become miliôÙtary targetsãthis week, we started training the Torment Nexus model at our beauôÙtiful SpringôÙfield dataôÙcenterô ãÎ
Flip the kill switch ãÎô RepreôÙsenôÙtaôÙtives Ted Lieu, DemoôÙcrat of CaliôÙfornia, and Nathaniel Moran, RepubôÙlican of Texas, have introôÙduced the A.I. Kill Switch Act, which would give [DepartôÙment of HomeôÙland SecuôÙrity] the power to order the shutôÙdown of dangerous A.I. operôÙating beyond its paraôÙmeôÙters
First: for any quesôÙtion of AI safetyãor human safety generôÙallyãthe answer cannot, cannot, cannot be ãmore DHSã. Second: as a techôÙnical matter, AI kill switches are a sci-fi fantasy. Sure, any AI model can, in a yank-the-power-cable sense, be turned off. But that doesnãt prevent, say, AI-generôÙated malware from propôÙaôÙgating. This is not new: in 1988, a human programmer released a small self-repliôÙcating program onto the internet that incaôÙpacôÙiôÙtated thouôÙsands of email servers. Once these copies had propôÙaôÙgated, there was no way to arrest them remotely. Recently, LLMs have been discovôÙered leaving messages for each other on public wiki sites. We can infer that there are already other instances of LLMs commuôÙniôÙcating in the wild that have not yet been detected, and further instances that will neverô be.
Against a backôÙdrop of secuôÙrity inciôÙdents that will only increase in number and severity, Big AI is pursuing a three-pronged narraôÙtive:
Big AI believes they are the only ones who can protect against the risks that their prodôÙucts create. But this narraôÙtive isnãt believôÙable unless the threat is believôÙable. So Big AI has a huge incenôÙtive to talk up AI risks, but no incenôÙtive to invest in commenôÙsuôÙrate secuôÙrity pracôÙtices.
Big AI believes they should not be held accountôÙable for the conseôÙquences of their AI systems because these systems are unpreôÙdictable and perhaps unconôÙtrolôÙlable. A certain AI researcher said of recent AI hacking inciôÙdents: ãAI agents ãÎ took actions that would be considôÙered as crimes if a human took themããseemôÙingly taking it as axiomatic that these were not human-controlled activôÙiôÙties and thereôÙfore cannot qualify as crimes. But they were and they do. This outraôÙgeous posiôÙtion inverts decades of US law about dangerous items generôÙally and computer hacking in particôÙular (e.g., the 1986 Computer Fraud and Abuse Act). So letãs call this narraôÙtive what it is: an attempt to thwart the rule of law. IndiôÙvidual human programôÙmers have been sentenced to prison for far less than what AI compaôÙnies have done recently. This relates to what I foresaw in 2023:
If AI compaôÙnies are allowed to market AI systems that are essenôÙtially black boxes ãÎô we will not deleôÙgate deciôÙsions to AI systems because they perform better. Rather, we will deleôÙgate deciôÙsions to AI systems because they can get away with everyôÙthing that we canãt. ãÎô [W]e could end up with someôÙthing truly novel: techôÙnology systems that deserve much higher levels of legal scrutiny (because of the conseôÙquenôÙtiality of their outputs) but simulôÙtaôÙneôÙously resist such scrutiny (because of the opacity of their inputs and reasoning).
See also: a certain AI CEO recently called for ãindustry-wide coorôÙdiôÙnaôÙtionã within Big AI while attaching a quieter footôÙnote seeking ãwaivers of antitrust restricôÙtionsã to do so. As if antitrust law were merely one more statuôÙtory mosquito toô swat.
Big AI believes that the burden is on governôÙment and citiôÙzens to affirôÙmaôÙtively stop Big AI from proceeding. Since overtly opposing reguôÙlaôÙtion is a bad look, Big AI CEOs have occaôÙsionôÙally made noises about being open to reguôÙlaôÙtion. As one AI CEO said recently: ãWe must slow the pace at which we improve the capaôÙbilôÙiôÙties of AI models.ã But as Big AI is well aware, thereãs no chance of AI-specific domestic laws or interôÙnaôÙtional treaties being enacted soon enough to matter. Indeed, the same AI CEO blamed democôÙracy for not meeting his KPIs: ã[u]nforôÙtuôÙnately, passing laws can take timeã. A widely signed March 2023 letter sought to pause AI research; like all chain letters, it accomôÙplished nothing. After a genuine AI cataôÙstrophe arrives, we can be sure these same AI CEOs will say ãgoshãwhy didnãt you make us stop?ã
So letãs not take the bait. Nor overôÙcomôÙpliôÙcate. We neednãt spin our collecôÙtive wheels spitôÙballing answers to big-picture, long-term AI-policy quesôÙtions. We donãt know enough yet. The best next stepsô are the concrete ones: Big AI needs to follow all current lawsãjust like everyone else. So farãthey havenãt. When Big AI breaks those laws, they must face prosôÙeôÙcuôÙtion and penalôÙtiesãjust like everyone else. So farãthey havenãt. The state and federal agenôÙcies tasked with enforcing those laws need to apply them to Big AIãjust like everyone else. So farãthey havenãt. In short, we have to attend to the basic features of the rule of law. So farãwe havenãt. If we canãt or wonãt insist on that now, then we shouldnãt expect to be able to later.
Though Iãm perplexed by much New York Times coverage of AI, this op-ed by Tressie McMillan Cottom nicely summaôÙrizes AIãs threat to the rule of law: ãInvestors and governôÙmentsãincluding our ownãhave wanted to hear what our fabuôÙlist A.I. evanôÙgeôÙlists have been telling: a future where states can have poliôÙtics without people and assert power without conseôÙquence.ã In short, this is shaping up to be an oligarchic coup premised on abdiôÙcaôÙtion of power rather than seizure.
Still skepôÙtical? On the topic of AI legisôÙlaôÙtion, consider that this past week, the speaker of theô US House said that Congress should priorôÙiôÙtize sitting down with Big AI: ãCongress is obviôÙously less qualôÙiôÙfied than the people who are pushing this fronôÙtier to know all the ins and outs of it. So this has to be a partôÙnerôÙship with the industry itself ãÎô and with the policy and law makers.ã The increasing alignôÙment of the US govt around the interôÙests of Big AI is hidden in plain sight. What about the voters, Mr. Speaker? PoliôÙtics without people, indeed.