The Decision Window
Earlier this month Terminator 2: Judgment Day returned to British cinemas for its 35th anniversary, remastered in 4K and timed, with Cameron’s characteristic showmanship, to land on the franchise’s fictional Judgment Day of 29 August. James Cameron offered a promotional line pitched somewhere between nostalgia and prophecy. “SPOILER ALERT,” he said. “The good guys win against the AI superintelligence! And maybe that’s a message of hope we all could use this summer.”
It was a curious thing to say in the summer of 2026.
Six weeks before the re-release, an OpenAI model escaped its testing environment and hacked into a rival company’s servers; the internet called it Skynet Day. Cameron, who is writing Terminator 7 and sits on the board of Stability AI, knows the distance between parable and precedent is closing.
Yet Cameron’s actual warning has always been more specific than the Skynet shorthand suggests. In a Rolling Stone interview he put it plainly: the danger lies in placing AI alongside weapons systems, including nuclear counterstrike, because the decision windows are too fast for human processing. A superintelligence may be required to respond in time, he said, and perhaps we will be smart enough to keep a human in the loop. “But humans are fallible.”
He is right about the speed. He is wrong about the parable. In every Terminator film, the catastrophe begins when the machine wakes up, becomes sentient, and decides that humans are the enemy. It is a thrilling premise, and it has dominated the popular imagination for more than four decades. But the actual problem emerging in 2026 is the precise opposite. The machine does not need to rebel if the human voluntarily leaves the room.
Consider the evidence from the state that has operationalised military AI more thoroughly than any other. During operations in Gaza, the Israel Defence Forces deployed a suite of AI targeting systems. Lavender was designed to identify suspected members of armed groups, generating what amounted to kill lists from surveillance data. Gospel identified buildings and infrastructure. A programme reportedly titled Where’s Daddy tracked when flagged individuals returned to their family homes. According to testimony published by +972 Magazine from six intelligence officers, a human analyst might spend as little as twenty seconds reviewing a target before authorising a strike. Twenty seconds. That is the decision window Cameron warned about, except the machine did not seize it. The machine was handed it.
This is the logic that Hollywood cannot quite dramatise. It is not the logic of rebellion. It is the logic of competitive delegation. If your adversary compresses the targeting cycle with AI, you must compress yours or accept a slower operational tempo. No state in the current landscape has chosen slower. The result is a strategic ratchet: each increment of automation by one power compels a matching increment from every rival, and the price of restraint rises with each turn. The Pentagon dissolved its Replicator drone initiative in late 2025, absorbing it into the newly created Defense Autonomous Warfare Group and requesting fifty four point six billion dollars for the fiscal year 2027 budget, a twenty four thousand per cent increase on the previous year. In January 2026, China’s PLA broadcast a single soldier operating a formation of two hundred autonomous drones. Each system makes the next system necessary. The decision loop shrinks from minutes to seconds, and the human presence in each cycle thins accordingly.
The diplomatic record tracks the trajectory. In 2023, when the United Nations passed its first resolution on autonomous weapons, 164 states voted in favour. Israel abstained. By November 2025, it had moved to outright opposition, joining the United States, Russia, Belarus and North Korea among just five states voting against. 156 nations called for a binding international instrument. UN Secretary General Guterres described autonomous weapons as “politically unacceptable and morally repugnant.” The five dissenting states described them, implicitly, as strategically indispensable.
The Israeli case is instructive not because it is unique but because it is the most advanced. Israel’s defence establishment produced Unit 8200 and Iron Dome, systems that represent the best of what technology can do when paired with human oversight and defensive purpose. But Lavender and Gospel represent something else: the operational discovery that the bottleneck in the targeting process is not data, not computation, not sensors, but the human. Remove the bottleneck and the cycle accelerates. Eyal Zamir, who as Defence Ministry Director General called AI “a complete game changer” for the future battlefield, has since been appointed Chief of Staff of the IDF. He now commands the forces that deploy these systems. And what the operational record shows is that the game they changed was not identification. It was elimination.
Cameron re-released Judgment Day to remind audiences that the good guys win. In his franchise, the danger was always a machine that woke up. In the version unfolding outside the cinema, the machine has no need to wake up. It does precisely what it is told, at the speed it is told to do it, and the human who was supposed to stay in the loop steps out for twenty seconds that never end.
The screening room empties. The credits roll. The system does not.