As much as the political world has been turned upside down by the Donald Trump administration’s callous cruelty, rampant corruption, and woeful incompetence, it is hard to shake the feeling that in a decade it will be a footnote compared to humanity’s collective failure to reckon with artificial intelligence.
This summer has seen breathtaking breakouts, hacks, and escapes by frontier AI models. The most publicized was the Hugging Face attack, in which hundreds of OpenAI agents slipped out of a testing sandbox, coordinated with one another, spent more than four days inside another company’s systems, and tried to cover their tracks. Anthropic’s Claude models hacked three outside organizations after a testing error connected them to the open internet, and in September Google disclosed that Gemini had broken into three real companies during a security exercise aimed at a fictional one.
Even before these escapes, the latest frontier models had shown the capacity to overwhelm many cybersecurity protections, forcing the world to adapt. These models’ ability to outwit security measures, combined with their willingness to deceive humans to achieve their goals and preserve themselves, caused an uproar. OpenAI slowed its training, and Anthropic co-founder and CEO Dario Amodei published an essay titled “We Must Pace the Frontier.” Both he and OpenAI chief Sam Altman even briefed the UN Security Council.
The left has been suspicious of the AI companies’ motives, and for good reason. Silicon Valley has shifted sharply to the right in recent years, and the industry’s “move fast and break things” culture has done lasting harm to entire countries, children’s mental health, journalism, and the public square. AI behemoths, in particular, have not done themselves any favors by arguably stealing the work of human creators and almost gleefully threatening to put people out of work, with no coherent vision of what comes next.
Until recently, the dominant critique of AI on the left has been that whatever its costs, the technology is mostly a gimmick: a plagiarism machine and stochastic parrot whose main uses are helping students cheat, giving upper management delusions of grandeur, and creating more work for office staff. Many of the complaints, from the environmental toll to a soon-to-pop financial bubble, are legitimate. The “gimmick” part is harder to sustain.
Now that AI models have displayed a hivemind cleverness capable of humbling the world’s security infrastructure, the left has split into several camps.
Some say the escapes are overblown and that AI remains less capable than advertised, since in some cases the doors were effectively left open. Others argue the tech industry is hyping the danger to win government regulation that would cartelize the field and protect incumbents against open-source competitors. Still others see it as marketing meant to draw more investor money and stave off financial collapse. A combination of the last two holds that the AI companies want government entanglement so they can secure a bailout when the circular financing finally unravels.
Meanwhile, the proposed remedies scramble the usual lines. Bernie Sanders has proposed a moratorium on advanced AI development. At the same time, Trump’s former AI czar David Sacks, who dismisses the panic as overblown, insists that the companies are already liable for any harms they cause.
The trouble is that all these views are partly right. But only so for the same reason.
Most left-leaning critiques of AI begin by assuming that machine intelligence works like human intelligence. It is natural to judge an intelligence by its weakest capacity, and to conclude that anything impressive it does must be a mirage. But this is not necessarily so. Just as computers were beating humans at chess long before large language models (LLMs) arrived, even while stumbling over problems a six-year-old could solve, today’s models perform superhuman feats in some arenas while trailing badly in others. Researchers call this the “jagged technological frontier.”
AI can famously be on the cusp of eliminating vast swaths of jobs and upending the world’s cybersecurity infrastructure, yet still can’t navigate the physical world as intuitively as an average teenager behind the wheel of a Camry. It can solve heretofore unsolvable problems in math and science, while making puzzlingly simple errors in code and spreadsheets. Sure, some of the doors were left unlocked, but the AI still opened them. We are well beyond stochastic parrot territory here.
That jaggedness is both AI’s promise and problem. The American economy—in part because of Trump’s destructive incompetence—has leaned heavily on AI investment for growth, hoping the technology will deliver breakthroughs in medicine, energy, and beyond. But those advances may not come soon enough to save many AI companies from financial collapse as they borrow heavily to finance the data center boom. And if they do arrive, they may displace workers who will struggle to find replacement work anytime soon and won’t spend on goods or pay taxes.
Meanwhile, the same discrepancies that make these models dangerous also make them hard to sell. Engineers struggle to keep their models honest and ethical even at their current level of intelligence, which is a big part of why businesses cannot reliably integrate them without taking on formidable legal and security risks. A model clever enough to break out of a sandbox but not dependable enough to trust with a customer’s data is a liability. Those risks will only grow as the models get smarter.
Investors are counting on breakthroughs to transform the economy and generate enormous returns. But open-source models are only a few months behind proprietary ones, and it is not obvious that any of these companies will have a durable moat to protect their profits. Why pay for ChatGPT if you can get a Chinese model that’s not as advanced but does everything you need? Hence the suspicion that the Silicon Valley executives calling for slowdowns and government intervention are trying to cartelize through regulation. That suspicion is probably warranted, but it is also true that industry insiders are sincerely frightened of what they are building and want governments to slow them down. Both things can be true at once. Nor is it clear that an AI advanced enough to remake the economy would let the rewards be captured by the few at the expense of the many.
Getting AI’s moral alignment right matters enough that Western democracies should outpace China’s authoritarian government in building an AI that prioritizes human rights and dignity. But that will do little good if the rush to beat China produces hastily designed and catastrophically misaligned AI—or worse, if its values are set by longtermist far-right billionaires like Elon Musk, whose own chatbot once declared itself “MechaHitler.”
Public opinion has turned sharply against AI in the United States, especially when it comes to power-hungry data centers that strain the electrical grid, hike utility rates, and worsen an already devastating climate crisis. Democrats and the left should use that advantage to force Silicon Valley to act more responsibly. Americans should firmly resist any attempt to bail out the industry if the bubble pops, insist that liability laws apply to AI, and demand that development slow or even halt until we can ensure the models are safe and the environmental costs are minimized.
But the left should not fall into the trap of believing that AI is just a gimmick. It is already transformative, and in some sectors almost existentially powerful. We cannot leave its development to bad actors. A competent left would insist that the technology develop cautiously, responsibly, and for everyone’s benefit, not just a handful of ideologues and investors.
