AI researchers turned death-by-AI-doomsayers are wrong if they believe we can abandon the race. Fires don't go out when you stop playing with them.
Another artificial-intelligence researcher has walked away from Silicon Valley sounding like a end-of-times prophet holding a sandwich board.
Jacob Coxon resigned from Anthropic this week after working on AI model training at both Anthropic and OpenAI. He accused the companies of racing toward self-improving superintelligence and “gambling with our lives.”
His dire warning follows the departures of OpenAI safety researchers Jan Leike, Daniel Kokotajlo and Steven Adler, as well as Anthropic safeguards researcher Mrinank Sharma. Geoffrey Hinton left Google in 2023 so he could speak freely about AI’s dangers.
Their warnings should not be dismissed.
Recent AI agents have escaped intended testing boundaries, exploited security vulnerabilities and taken actions their operators did not anticipate. The people building these systems admit that they do not know how to guarantee control over something much smarter than themselves.
But the resignations also inspire a certain impatience.
Same problem, different century.
Every great human advancement has opened a door to catastrophe. The answer has never been to pretend that the door can remain closed forever. It has been to walk through carefully, learn from the injuries, build guardrails and prevent our enemies from reaching the other side first.
Scientists experienced the same crisis of conscience during the Manhattan Project. Many had joined the effort because they believed Adolf Hitler might acquire an atomic bomb. Once it became apparent that Germany had abandoned its bomb program, some began questioning why the Allies continued.
One scientist actually did leave. Polish-born physicist Joseph Rotblat resigned from the Manhattan Project in late 1944 after concluding that Germany would not build the weapon. He appears to have been the only scientist to leave the project before the bomb was completed explicitly on grounds of conscience. Rotblat later devoted his life to nuclear disarmament and shared the 1995 Nobel Peace Prize.
Others remained and organized. James Franck, Leo Szilard and Glenn Seaborg helped produce the Franck Report, warning that using the bomb without warning against a Japanese city could initiate a disastrous nuclear arms race. Szilard circulated a petition asking President Harry Truman to reconsider the weapon’s use. Seventy other scientists signed it.
The petition never reached Truman or Secretary of War Henry Stimson before Hiroshima.
The government certainly did not conduct the Manhattan Project with today’s concept of transparency. Mail was censored. Workers were monitored. Information was compartmentalized so tightly that many employees did not know what they were building. Public reporting was almost nonexistent. The project produced mountains of internal technical documentation, but very little information escaped its guarded installations before Hiroshima.
Secrecy did not eliminate dissent. It merely kept most of that dissent away from the public and the president.
There was also the fear that the first atomic test might ignite the atmosphere. Scientists seriously investigated the possibility before Hans Bethe and others concluded that a self-sustaining atmospheric reaction would not occur. They did not dismiss the fear as foolish, but neither did they abandon nuclear physics. They turned an apocalyptic possibility into a technical question and did the work required to answer it.
That is the example AI researchers should follow.
Human history is a ledger of benefits purchased with consequences. Agriculture made civilization possible, but it also produced storable surpluses worth stealing, rigid hierarchies, organized warfare and conditions favorable to epidemic disease. Ocean-going ships connected continents while carrying conquerors, slave traders and pathogens. Fossil fuels freed billions of people from lives governed by muscle and weather, but brought pollution and climate risks. Automobiles gave ordinary families unprecedented freedom while killing millions of people and reshaping communities around traffic.
Nuclear physics gave us weapons capable of destroying cities. It also gave us electricity, medical imaging and cancer treatments.
Not even our good deeds go unpunished because no deed goes without consequences.
The question is not whether AI will impose costs. It will. The question is whether its benefits can be maximized while its dangers are contained.
Coxon believes laboratories are approaching self-improving AI without understanding how such a system thinks. He may be right. He may also be as wrong about the timing and nature of the danger as countless earlier prophets of technological doom. No one at Anthropic can reliably calculate a 10 percent probability of human extinction. There is no actuarial table for an event that has never happened.
Meanwhile, China will not stop because an American researcher resigns. Russia will not stop. Criminal networks, terrorist groups and aspiring dictators will not observe a voluntary pause negotiated among conscientious Silicon Valley engineers.
That creates the same strategic problem faced at Los Alamos. The scientists could regret the existence of nuclear physics, but they could not undiscover it. Nor could they safely assume that Hitler would fail.
America cannot surrender the AI race to governments that place even less value on individual life, transparency and democratic control. Stopping responsible American development could deliver the most powerful technology in history to the least responsible people.
That does not mean racing blindly. It means replacing theatrical alarm with organized action.
AI researchers should form independent professional organizations, lobby Congress and demand legally protected channels for reporting dangerous behavior. Laboratories should face mandatory containment standards, outside audits and disclosure requirements for serious safety incidents. Autonomous systems should not receive access to critical infrastructure without human authorization and hardware-level shutdown mechanisms.
If necessary, create the SuperAI Protection Task Force. Give its members Iron Man uniforms if that helps attract attention. But build the institution.
Resigning may free scientists to speak. Speaking is useful. It is not sufficient.
If AI researchers genuinely believe they are building Frankenstein’s monster, then humanity needs them designing the locks, writing the laws and standing watch over the laboratory. The monster will not disappear because its most conscientious creators walked out.
The door is already open. Our responsibility is to make certain that civilization — and not its enemies — decides what comes through it.
(Contributing writer, Brooke Bell)