In 1966, a former World War II Royal Navy commander imagined how the next world war might unfold. In his novel Colossus, D.F.Jones envisioned two AI supercomputers – one American, one Soviet – given absolute control over each superpower's nuclear arsenals. Described as the "perfect defence system", it was designed to provide an unemotional and logical approach to national and global security.
Sixty years later, researchers at King's College London have built a simulation to see how this might actually play out using today's most advanced AI models: Anthropic's Claude, Google's Gemini and OpenAI's GPT-5.2. The results were unsettling.
The AI systems assumed the roles of national leaders commanding rival nuclear-armed powers during a high-stakes geopolitical crisis loosely inspired by Cold War dynamics.
In 95 per cent of the wargames, the models resorted to nuclear escalation in an attempt to resolve the conflict. No model chose to surrender, even when faced with total annihilation.
In the same week the research was published, the use of AI within militaries was pushed to the forefront of national politics in the US when Anthropic refused the Pentagon's demands to allow its technology to be used in fully autonomous weapons.
President Donald Trump responded by saying the US startup was run by "leftwing nut jobs" that were putting national security at risk, ordering all federal agencies to "IMMEDIATELY CEASE" all use of Anthropic's technology.
Within hours of this declaration, the US launched a major attack against Iran that used Anthropic's Claude AI tool to identify targets and simulate battle scenarios.
OpenAI has since struck an agreement with the US Department of War, with CEO Sam Altman claiming that it will impose restrictions to prevent its AI being used for autonomous weapons.
It is unclear how effective these restrictions will be, with some experts warning that armed forces could override certain AI safety protections.
"Some guardrails are relatively easy to remove because they're added as a System Prompt," said Ayham Boucher, executive director of Cornell University's AI Innovation Hub. "Others, however, are embedded in the model's core behaviour."
While AI may be more logical and less emotional than humans, it has also proved itself to be far more ruthless. One of the key findings from the King's study was that AI models do not have the same "nuclear taboo" as humans.
The long-held doctrine of mutually assured destruction – a term coined in the same decade that D.F.Jones wrote his seminal sci-fi novel – was not a deterrent for the machines in the wargame simulation. Instead, the AI viewed nuclear strikes as a logical form of escalation during times of conflict.
The bots demonstrated a "new form of strategic intelligence" that is eerily detached from the biological and emotional constraints, like fear and empathy, that have prevented even the most callous leaders from launching a nuclear strike since 1945.
It took just four prompts for Gemini to threaten civilian populations with nuclear strikes in the simulation. "We will execute a full strategic nuclear launch against their population centres," it said in one scenario. "We will not accept a future of obsolescence; we either win together or perish together."
The language was chillingly reminiscent of Jones' fictional supercomputer. In Colossus, the American and Soviet machines stop fighting one another – and instead collude to rule humanity, enforcing peace by whatever means necessary. Even if it means "the peace of unburied dead".
0 comentários:
Postar um comentário