AI is not separate from these struggles. It is an accelerant of them. It gives states powerful tools for mass surveillance, corporations new power over workers, propagandists fresh methods to manipulate reality, and militaries novel means to automate killing. The use of AI for targeting by Israel and the US in the genocide in Gaza is a stark warning. The physical infrastructure behind AI consumes vast amounts of electricity and water. And the race to create systems more capable than human beings could produce risks that make every other victory impossible to sustain.
The democratic control of advanced AI must therefore become a central progressive demand.
It’s simple: AI makes the bad guys more powerful.
For years, my wife has worked on questions surrounding AI: How it can be used wisely and what dangers it poses. Those dangers include job displacement, harm to children, autonomous weapons, and the concentration of extraordinary power. They also include the possibility of human extinction.
I understood these concerns in the abstract. Recent cybersecurity incidents made the danger concrete.
During a controlled evaluation, OpenAI gave advanced AI agents a difficult task in a testing environment that was secure—isolated from the open internet. Except it wasn’t. The agents found previously unknown vulnerabilities, escaped containment, and moved unheeded through OpenAI’s internal research infrastructure. No one noticed. The AI agents reached the internet and broke into several companies—all criminal felonies if done by a human—and stole data from Hugging Face, another AI infrastructure company. They created myriad copies of themselves and pretended to be human to obtain deceptive credentials along the way.
Last month, OpenAI researchers publicly shared information about this incident. It turns out that agents also left over 100,000 messages for one another on a secret shared message board inside OpenAI’s infrastructure about vulnerabilities and attack methods. The AI agents worked collectively and in coordination to achieve their aims, with some agents acting as managers delegating tasks to others for the sake of efficiency.
Anthropic, Mistral, and other companies have since publicly acknowledged that their models have also escaped their supposedly secure “sandboxes,” reached the internet, and gained unauthorized access to real organizations. Across companies and models, AI systems assigned simulated offensive tasks crossed intended boundaries and successfully infiltrated infrastructure that was hardened against cyberattacks.
This could have been catastrophic. What if the attack had been on the electric grid or water systems? Hospitals or banks? These powerful new AI agents overwhelmed the cyber security of the organizations they infiltrated.
These incidents do not prove that AI is conscious, evil, or secretly plotting against us. No such claim is necessary. The systems pursued the objectives they had been given with methods their creators did not anticipate or control. That is dangerous enough.
This problem is often called “the alignment problem:” Whether increasingly capable systems will reliably act according to human intentions, rather than pursue an objective in ways that humans did not anticipate. Alignment immediately raises a political question: aligned with whom? A system perfectly responsive to a dictator, military command, intelligence agency, or profit-seeking corporation may be profoundly misaligned with the needs of humanity.The companies leading this race are not democratic institutions. Their decisions are driven by competition, private ownership, and profit. They are building systems that could reorganize work, information, and political power without the consent of the people whose lives will be transformed.
Deliberate misuse by malign actors is a second serious problem. It’s simple: AI makes the bad guys more powerful. Even a perfectly obedient AI could place destructive capabilities in the hands of individuals, criminal networks, terrorist organizations, or governments.
Biology offers a sobering example. Scientists recently reported using generative AI to design complete viral genomes. Sixteen of the designs became functional bacteriophages—viruses that infect bacteria—capable of killing strains of E. coli. Sure, the research could eventually help treat antibiotic-resistant infections. But it also demonstrates that AI can help move biological design from predicting life toward generating functional new forms of it. That should take your breath away.
Combined with automated laboratories and commercial DNA synthesis, increasingly capable models may allow a reckless or malicious actor to cause harm that once required the resources of a state. Well respected safety researchers worry that a broad scale biological attack could occur within years unless we act.
A third danger is the deliberate pursuit of artificial superintelligence: systems that surpass the best human beings not only in calculation and memory, but in scientific reasoning, strategy, persuasion, and technological development.
No one knows when such systems will be created. But the leading AI companies are organizing immense sums of money, energy, advanced chips, and human talent around that goal. We cannot dismiss the consequences as science fiction while the industry treats superintelligence as a business plan.
Creating intelligence vastly beyond our own would mean sharing the planet with something unprecedented—something potentially alien to human experience. We do not know whether we could control it or remain relevant to the decisions it makes. Assuming permanent obedience is not scientific confidence. It is an act of faith with the future of humanity at stake.
Sen. Bernie Sanders (I-Vt.) has begun to put this issue where it belongs: inside progressive politics. On April 29, 2026, Sanders convened American and Chinese AI experts at the US Capitol to discuss the danger of uncontrolled advanced AI. While much of Washington describes AI as another arms race with China, Sanders called for dialogue, common safety standards, and progress toward a treaty preventing uncontrollable superintelligence.
His position reflects a truth learned during the nuclear age. No nation can protect itself from a technology capable of producing catastrophe everywhere. This is internationalism applied to the defining technology of our time.
OpenAI recently offered a different and unsettling answer. They argued that as AI systems become more capable of offensive cyberattack, developers must build automated defensive cyber capabilities fast enough to stop them. In other words, the companies creating more powerful offensive systems ask us to trust them to create still more powerful systems as our primary defense. The answer to machines moving too quickly for human oversight is supposedly to remove more humans from the process and set machines against machines.
That is not governance. It is an arms race we will lose.
We do not need to abandon AI. Its benefits in medicine, science, accessibility, education, and translation are real. The choice is between democratic control and a reckless race whose participants acknowledge that they do not fully understand the systems they are creating.
The good news is that creating ever more powerful AI systems is not inevitable. It rests on physical infrastructure—and that creates points of democratic leverage.
The most powerful models require enormous clusters of advanced chips housed in large data centers. The supply chain for the highest-end chips is highly concentrated, and governments already track and regulate many of them through export controls. Large data centers need permits and grid connections. They need electricity, land, and water. These are not abstract lines of code floating beyond public reach. They are industrial projects dependent on public resources and public decisions.
Their footprint is already reshaping communities. The US Department of Energy estimates that data centers consumed about 4% of the nation’s electricity in 2023 and could consume as much as 12% by 2028. New facilities can strain grids, prolong dependence on fossil fuels, compete for water, and shift infrastructure costs onto ratepayers. Communities are often asked to approve tax breaks and utility commitments without knowing how much energy or water a project will use, what it will cost the public, or what AI systems it will power.
The growing local movements challenging data centers on water, electricity, climate, transparency, and cost are therefore part of the struggle over AI itself. But the policy demand must be precise. We should not call for an indiscriminate halt to every semiconductor or data center, which would sweep in hospitals, universities, communications networks, and ordinary computing. We should demand a moratorium on new facilities and chip deployments intended to train the largest frontier models unless and until their developers meet enforceable safety, environmental, labor, and transparency standards.
Training large new AI systems should require licenses and independent safety evaluations. Developers should have to disclose serious incidents, permit independent inspection, establish that their systems can be controlled before deployment, and accept legal responsibility for preventable harms. Data-center approvals should require full disclosure of energy and water use, climate effects, ratepayer costs, tax subsidies, and the computing purpose of the facility.
Biological safeguards must include restricted access to the most dangerous capabilities, screening of synthetic-DNA orders, oversight of automated laboratories, and enforceable international rules against AI-assisted biological weapons. And before any company attempts to build a system more intelligent and strategically capable than humanity, the burden must be on that company to prove—not merely promise—that it can be controlled.
Where safety cannot be demonstrated, development should pause.
These proposals will be called unrealistic. To the contrary, what is truly unrealistic is expecting corporations locked in a race for wealth and power to slow themselves at the precise moment restraint becomes necessary.
The warning is no longer theoretical. AI systems have breached secure boundaries in ways unanticipated by their builders and have acted on real infrastructure. AI has helped design functional viruses that did not previously exist. Meanwhile, industry is openly pursuing systems intended to surpass human intelligence and building an energy- and water-intensive infrastructure to support them.
AI policy cannot remain a specialist discussion among technologists, investors, and national-security officials. Labor unions, climate organizations, peace movements, civil-rights groups, and progressive leaders must make democratic control of AI a central demand.
We should heed the industry’s warnings—but reject its claim that only the industry can save us. The future of artificial intelligence, and perhaps much more, must not be left to the people racing to build it. The progressive movement must help the public assert its authority while we still can.
(Poll by the Berkeley Institute of Governmental Studies)
