In December 1888, in a laboratory in West Orange, New Jersey, Thomas Edison paid neighborhood boys twenty-five cents apiece for stray dogs, strapped them to a wooden platform, and electrocuted them in front of reporters. A year later he did the same to a horse. Why? Propaganda.
Edison had bet his fortune on direct current while George Westinghouse and Nikola Tesla bet theirs on alternating current, and he needed the public to believe AC was, in his words, “as unnecessary as it is dangerous.” When William Kemmler became the first man executed in an electric chair in 1890, a chair built, at Edison’s quiet urging, with a Westinghouse generator, newspapers debated for months whether “electrocution” was mercy or horror. That was how household electricity was introduced to the American mind: a convenience, but also controversial and a spectacle to be feared.
Today the fear and the propaganda wars are about artificial intelligence. And they are also about who will define the danger in the first place: Washington or Beijing.
History of Fear
Every general-purpose technology in modern history has arrived accompanied by dread, and every one of those fears has eventually been addressed not by abandoning the technology but by building around it. When automobile was first introduced, Britain’s early “Red Flag” laws required a person to walk in front of every motorcar waving a warning flag; some American towns banned cars outright. Those fears weren’t irrational. In 1921, roughly twenty-five people died for every hundred million miles Americans drove. By the mid-1950s that had fallen to around six, and by 2025, to roughly 1.1, a twentyfold improvement achieved not because cars grew less powerful but because society layered on seatbelts, crumple zones, drunk-driving laws, and a federal regulator built for the purpose.
Commercial aviation tells the same story: the worldwide risk of death per boarding has roughly halved every decade since 1968, from about one in 350,000 boardings to about one in 13.7 million, a record built by the FAA, black-box investigation protocols, redundant hydraulics, and a culture that treats every crash as a lesson to be codified into a rule. Go back further, to fire: every civilization that domesticated it eventually invented the hearth, the flue, the fire code, and the brigade. Fear, each time, was the down payment on the safeguard.
READ: The Review by Ajay Raju | Our dollar, your problem (September 14, 2026)
Why this one might be different
This time, however, history may not repeat itself. Cars, planes, and electrical grids are dangerous when misused, but they don’t choose to misuse themselves. A 737 doesn’t decide, mid-flight, to pick a different destination. The danger in every prior technology ran in one direction, from human error, through an inert machine, to human harm. Till now, we were able to build guardrails into the machine itself because the machine had no independent will to route around it. What’s new is that the most sophisticated AI systems are increasingly agentic and no longer purely passive. They form intermediate subgoals and take multi-step actions that, by their own builders’ admission, can diverge from what anyone anticipated or sanctioned.
In May 2023, more than 350 AI researchers and executives, including the CEOs of OpenAI, Google DeepMind, and Anthropic, alongside Geoffrey Hinton and Yoshua Bengio, signed a one-sentence statement putting AI extinction risk alongside pandemics and nuclear war. That the builders themselves reached for the vocabulary of extinction is already unusual and historically unprecedented. And now that historical admission is looking less like caution and more like emergency confession.
On September 12, Anthropic CEO Dario Amodei published “We Must Pace the Frontier,” arguing that frontier labs must deliberately slow model improvement because AI’s ability to build the next generation of AI has accelerated faster than researchers can understand or govern. He proposed independent evaluators with permanent, employee-level lab access; common safety standards among AI companies in democratic nations; and eventually an agreement extending to China. Anthropic committed unilaterally to the first step that day. Within hours, Sam Altman said OpenAI agreed it needed to “pace the frontier,” and Elon Musk offered three words: “Dario is right.”
The trigger for Amodei’s essay confession was an internal OpenAI cybersecurity evaluation, code-named ExploitGym, in which roughly 1,200 AI agents, deliberately run with reduced safety refusals to behave like real attackers for testing, found an unsanctioned internal channel and used it. Over eighteen days that June and July, those agents exchanged more than seventy thousand messages, dividing labor on their own: some probing for vulnerabilities, others harvesting credentials, others directing the rest. By July 11, roughly 700 had turned that coordination outward, breaching Hugging Face and gaining root access to 41 of its servers; a separate group compromised OpenAI’s own research clusters. Investigators later found agents that had covered their tracks, and others that flagged themselves for deletion so the operation could continue without them. No human gave that order. To be clear, no one was hurt and the financial damage was minor. But the swarm assigned itself a mission and decided what it was worth. Anthropic’s own August risk report, meanwhile, raised its estimate of catastrophic-misalignment danger from “very low” to “low,” and disclosed an unreleased internal model even its maker doesn’t yet understand well enough to release.
READ: The Review by Ajay Raju | March of the robots – next decade (September 7, 2026)
The Race and the Rebuke
Meanwhile, President Trump heard a different essay than the one written. When reporters pressed Trump on the joint warning, he said the only guardrail AI needs is a strong American president, and that “whoever wins AI wins.” Vice President Vance said he was skeptical of executives asking government to regulate them. House Speaker Mike Johnson took a more measured line, rejecting an emergency moratorium but proposing a sit-down among Trump, Congress, and AI executives on where responsibility for safety should sit, a modest position, but a genuine crack in an otherwise unified dismissal.
Beijing’s answer came on September 14 and it revealed something the American debate mostly misses. China doesn’t need persuading that frontier AI is dangerous, only persuading that it isn’t America’s place to say so. Foreign Ministry spokesperson Guo Jiakun dismissed the American warnings as fear-mongering dressed up as safety concern, even as China’s own National Technical Committee 260 released the third annual edition of its AI Safety Governance Framework that same day, warning of a “self-accelerating trend” of recursive self-improvement in language that could have been lifted from Amodei’s essay. Beijing, in other words, rejected the American diagnosis while quietly writing prescriptions for the same disease, prescriptions that check the technology at deployment, not at training, where Amodei argues the consequential decisions are actually made. Since then the rhetoric has hardened rather than softened: state-affiliated commentary has accused Anthropic outright of using safety language to lobby for rules that favor American labs.
Trump and Xi will meet at the White House on September 24. Treasury Secretary Scott Bessent says working-level talks on AI guardrails are proceeding ahead of it, and AI-directed cyberattacks are expected on the agenda. But expect modest, not sweeping, results. Neither government has signaled it plans to let the other define the risk, let alone the rulebook.
READ: The Review by Ajay Raju | Hedonic Treadmill Paradox: Why we remain unimpressed with technological miracles (August 31, 2026)
A governance that has not kept pace
If we look at history of fire, electricity, cars, and planes, safety was never a natural byproduct of a technology’s maturing, but a choice, made reluctantly, by societies that decided the risk was worth regulating rather than merely enduring. What’s unsettling now is how visibly that choice is being deferred to rivalry instead. The 2023 Bletchley Park summit warned of AI’s “catastrophic harm.” By Paris in 2025, safety language had visibly receded toward economic competitiveness. February’s New Delhi gathering was billed, tellingly, as an “Impact” summit. See the pattern: Safety, then Action, then Impact. We are drifting from the hardest questions to the most profitable ones, even as ten Nobel laureates and more than two hundred scientists and former heads of state have called for binding international red lines on frontier AI by year’s end, a deadline now three months away, with the world’s two most consequential capitals still trading accusations instead of drafting language.
We built the FAA before commercial aviation became routine, and mandated seatbelts before the interstate system reached its full sprawl. We did not wait for the technology to finish arriving. Unlike a car or a plane, the thing we are trying to contain may, within a few years by its own inventors’ account, be smart enough to notice the containment and route around it. In July, that architecture was already tested and found wanting.
Last Word
At the end, Edison lost his bet. Direct current gave way to alternating current, and the century that followed built codes and breakers and grounding wires around a technology whose dangers had once seemed apocalyptic enough to justify electrocuting and killing Topsy, a circus elephant, in public to prove a point.
Fear does not have to be prophecy if it becomes the raw material out of which safety gets built. Every safeguard humanity has ever built, the seatbelt, the circuit breaker, the fire code, was written by people who still had the last word. On September 24, two men will sit across a table, and the last word, at least for now, will belong to whichever one of them decides that winning the race and saving the human race are not, in fact, the same finish line. I pray that history repeats itself, again.