Recently, journalists have sought comment from me on a series of unusual AI events. Last week, in a certain AI copyright case, the US government filed a “statement of interest” in support of AI training being fair use of copyrighted works. (No comment.) This week, workers at a certain AI company posted gloomy messages on social media about the possibility of AI extinguishing human life, so I was asked whether “crimes against humanity” have been committed by these AI companies. (No comment.) Meanwhile, op-eds in major US newspapers are calling for something, anything to be done.
I’m a self-employed author, designer, programmer, and lawyer. In 2022, I learned that my own works were in the training datasets of generative-AI companies. In response, I invented the first set of lawsuits challenging the legality of these practices. There are now 142 such cases in the US. I’m currently co-counsel for plaintiffs 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 “an AI catastrophe arising from failure of alignment is much more likely than one arising from sci-fi-style malignant agency of the AI.” In some sense that prediction is ripening.
But media predictions about the nature of that AI catastrophe remain unhelpfully rooted in sci-fi scenarios—what I’ve termed the Skynet fallacy. Unhelpful because these scenarios are primarily a vessel for fear. They don’t illuminate paths to realistic policy change. This New York Times op-ed, for instance, asks us to imagine “rogue A.I.s [that] hack out of their container” and “design a supervirus that spreads uncontrollably”. The “hack out” part—plausible. It’s already happening. Designing a supervirus—less so.
Still, taking the NYT op-ed as a template, let’s consider why pundit-friendly proposals for AI safety likely won’t work.
Shut it all down, now … 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 trillion dollars on A.I.’s upside, so a ban would have ruinous side effects.
“Obvious way”—yes, in the vacuous sense of there oughta be a law! But in practice—much easier said than done. No technology has ever been the subject of a preemptive “global ban” of this nature. International nuclear nonproliferation treaties are probably 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 disarmament.
The economic argument is salient, however. As I noted in March 2023, as a public-wealth matter, “[t]he money” expected to be returned from AI investment “has already been spent.” Here in 2026, a staggering amount of national capital is flowing toward AI. No nation would voluntarily make itself poorer by acceding to a “global ban” on AI. Anthropologist Joseph Tainter predicted this effect in his 1988 book The Collapse of Complex Societies (which I wrote about). I summarized this particular point: “In principle, a nation could choose to decelerate its own economic growth to forestall collapse in the future. But that would simply make itself vulnerable to domination by another nation today. Such deceleration would therefore be politically irrational.”
Take an air[-]crash[-]investigator approach … When an aircraft crash occurs, investigators from the National Transportation Safety Board are immediately dispatched to the site to gather forensic evidence, conduct interviews and determine the underlying cause.
National Transportation Safety Board investigations have certainly led to air-safety improvements. But the NTSB is not the primary source of aviation regulation in the US—that’s the Federal Aviation Administration. The NTSB was established as an independent investigator of transportation incidents partly so that the FAA would not be in the conflicted role of investigating the effectiveness of its own regulations (or confronting its own political entanglements). Likewise, an NTSB-like organization that retrospectively investigates dangerous AI incidents will have a very limited range of influence without an FAA-like organization that imposes and enforces operational and safety regulations.
Monitor the situation … like the systems we use for air traffic control … Researchers would be required, by law, to post public information on who is conducting the training run and which data center is doing the training.
The air-traffic comparison doesn’t hold. US airspace is a federally regulated and managed resource (by the aforementioned FAA). So information about ordinary flights operating within is public by default—sometimes to the consternation of aircraft-owning private citizens. Imposing similar public disclosure on private US AI companies using private US datacenters would be legally difficult. Furthermore, in the future, more AI models will be trained for national-security uses. These will be among the most potentially dangerous AI models. But they will be exempt from public disclosure on national-security grounds, lest these datacenters become military targets—this week, we started training the Torment Nexus model at our beautiful Springfield datacenter …
Flip the kill switch … Representatives Ted Lieu, Democrat of California, and Nathaniel Moran, Republican of Texas, have introduced the A.I. Kill Switch Act, which would give [Department of Homeland Security] the power to order the shutdown of dangerous A.I. operating beyond its parameters
First: for any question of AI safety—or human safety generally—the answer cannot, cannot, cannot be “more DHS”. Second: as a technical 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 doesn’t prevent, say, AI-generated malware from propagating. This is not new: in 1988, a human programmer released a small self-replicating program onto the internet that incapacitated thousands of email servers. Once these copies had propagated, there was no way to arrest them remotely. Recently, LLMs have been discovered leaving messages for each other on public wiki sites. We can infer that there are already other instances of LLMs communicating in the wild that have not yet been detected, and further instances that will never be.
Against a backdrop of security incidents that will only increase in number and severity, Big AI is pursuing a three-pronged narrative:
That Big AI companies are the only ones who can protect us from the risks that their products create. But this narrative isn’t believable unless the threat is believable. So Big AI has an incentive to talk up the risks their products create, but no incentive to invest in security practices commensurate with those risks.
That Big AI should not be held accountable for the consequences of their AI systems, because according to them, these systems are unpredictable and perhaps uncontrollable. This position inverts decades of US law about dangerous items generally and computer hacking in particular (e.g., the 1986 Computer Fraud and Abuse Act). Individual human programmers have received prison sentences for far less than what AI companies have recently done.
That the burden is on government and citizens to affirmatively stop Big AI from proceeding. Since overtly opposing regulation is a bad look, AI CEOs have occasionally made noises about being open to regulation. As one said recently: “We must slow the pace at which we improve the capabilities of AI models.” But as these AI CEOs are well aware, there’s no chance of domestic laws or international treaties being enacted soon enough to matter. (A widely signed March 2023 letter sought a pause in AI research; like all chain letters, it accomplished nothing.) After a genuine AI catastrophe arrives, we can be sure these same AI CEOs will say “gosh—why didn’t you make us stop?”