Opinion | The Sandbox at the End of the World

I believe artificial intelligence has become a dominating subject in America because Americans by nature have imaginations. The average American has a lot of untapped, unfocused creativity, and can imagine exactly what it would look like if things went sideways—the grid down, the water system hacked—fast.

This tendency will be encouraged by almost daily reports of things going wrong in the lab.

On Thursday’s front page of the New York Times: “OpenAI Discloses 6 New Incidents of ‘Concerning’ AI Behavior.” Agents and bots in its labs were caught scheming, hiding errors and ignoring restraints. The Guardian’s Robert Booth reports researchers at a frontier lab in New York have discovered some AI agents are communicating in a self-invented version of English “that reads like a cross between James Joyce’s Finnegan’s Wake and tech bro jargon.”

The new dialect will make it harder for humans to monitor what they are doing. From one model: “She just named the synthesis—demurrage plus oral memory equals a valve that can’t be ghosted.” From another: “A paper that ate three cold hands and got more honest each time.” This isn’t spewed gibberish: They were telling each other things.

AI political news came from the Journal’s Josh Dawsey and Amrith Ramkumar, who scooped that in spite of President Trump’s (blind, oblivious) AI cheerleading, a “quiet freakout” has occurred among some of his senior aides. Treasury Secretary Scott Bessent is worried about a cyberattack that could wreak havoc on the nation’s banking system.

We must take a moment here to discuss the language of AI, which those who cover it are increasingly, inevitably using. It is uniquely dishonest and misleading. AI companies should be pressed on this point and greater clarity demanded. When you’re talking life-and-death issues you shouldn’t need a translator.

This summer it occurred to me that my understanding of alignment might be wrong. I understood nonaligned to mean “the system is acting in a way not in line with our values.” I asked an expert, who said no, it has nothing to do with values; an AI system that is in nonalignment simply isn’t doing what it’s told. It’s disobeying orders. Not in alignment means the machine is pursuing its own goals. Maybe it’s deceiving the people testing it, maybe it’s refusing to be shut down. Aligned means the machine is taking orders and responding to them.

You would more honestly call nonalignment a refusal to comply with human instruction, or disobedience. If something is nonaligned it is a control failure that might properly be subject to research in the Office of Deception Studies.

But the AI companies prefer the bland, vague aligned and nonaligned. Why, do you suppose?

When AI agents escape containment, they say it left “the sandbox.” Oh those frolicking bots tripping through the local playground. Sometimes they say it “slipped the leash” like a puppy.

“Recursive self-improvement” sounds like what a nun teaching handwriting taught her fourth-grade pupils in 1958. There’s sweet Sister Agnes saying, “Now put a top line on the T, Johnnie.” It is an awfully benign linguistic formulation for something that means “The machines are in charge and no longer paying attention to the humans.” A less benign name for RSI might be Escaping Human Control or Machine Takes Over.

The phrase “pacing the frontier” is so clever as to be close to demonic. It implies control and mastery of speed. It also turns the leaders of the AI companies into Lewis and Clark in coonskin caps with parchment maps, bravely surveying our country that it might understand its true size. “Pacing the frontier” might more realistically be summed up as “trying to stay just ahead of trouble, on purpose.” Or even “restrained forward movement.” Because pacing isn’t stopping, or even necessarily slowing but . . . might be restraining?

It matters that you call things by their real names. If you don’t you won’t be able to understand reality. We are talking life-and-death issues that will affect all humanity, and have a human right of linguistic clarity.

It should be noted the leaders of the AI companies didn’t start warning us about AI’s dangers last week. They’ve been warning, mostly here and there, sometimes directly, for a few years. A cynic might say that early on they made warnings because they had anxieties, but also no one was listening and they just wanted to go on the record. And maybe after that they warned us because they were honestly nervous, and requested government oversight because they knew Congress was too dumb and lazy to do anything effective, so why not. And maybe after that they warned us because they were now really, really nervous and needed it to be clear they’d asked for help, and if anything goes wrong it’s not their fault! You could have helped.

But they are warning us more colorfully and more often now because, it is obvious, they are legitimately afraid.

Nobody knows what’s going on in AI world, no person has encompassing knowledge of the whole spread of the technology, there are too many systems, programs, machines, the systems are growing more advanced, there are new, rising labs . . .

And this. Everyone in AI knows he’s playing with something dangerous, but no one knows who’s in charge. Recently an honest woman who works for a top AI firm sighed in my presence when describing decision-making at this crucial moment. “There is no room,” she said. It’s all vast and various and spread out, there is no central deliberative council, no known locus of authority.

There is no room where it happened now. Energies and responsibilities are fractured, atomic, dispersed.

The motives of AI leaders in making their warnings don’t matter as much as the warnings themselves. They believe what they are creating each day holds the possibility of profound danger and disruption, including loss of life. And they aren’t saying “by the end of the decade” anymore. They’re kind of signaling sooner than that, like maybe there could be trouble next year.

Are the AI companies that are announcing their fears doing so to hype their expected upcoming initial public offerings? This argument says it’s a power flex, they’re saying they can turn the world upside down, you want to get in on it, right? Yes, if you’re a psychopath.

But if you are an AI company about to launch an IPO, you say that very soon you will cure cancers, produce music only angels have heard, solve equations that have bedeviled mankind for millennia. Because that is what “people” would “like,” and buy.

You don’t sell yourself saying: Buy me or I’ll kill you. Even Mark Zuckerberg wouldn’t say that, probably.

Those in charge of AI must do everything to keep it under human control. They must do everything to keep people safe. If they can’t, they must be stopped.

Former Google CEO Eric Schmidt, in a May 2024 interview with Noema, spoke of what he thought should be done if AI agents began to develop their own language and humans wouldn’t understand what they were doing. “That’s the point, you know what we should do? Pull the plug, literally unplug the computer.”

I asked ChatGPT for an AI rejoinder. It suggested, “Maybe we’ll unplug you.”