Be Nice to Your Robot. It Might Remember.

Watch: the first human-versus-robot cage fight. Video: Frankie LaPenna.

In San Francisco last month, a man stepped into a cage to fight a six-foot robot styled after the Terminator. According to its promoters, the machine weighed nearly 190 pounds, had glowing red eyes and could deliver kicks with up to 850 pounds of force.

It wasn't an entirely fair fight.

The human contender, the social media personality Frankie LaPenna, took on three machines in a row. In the clip that went viral, the robot shrugs off his punches and sends him flying across the cage with a final kick. The organizers, Robot Entertainment Kombat, insist he went on to win, but the full footage hasn't been released and who actually won is still disputed. The robots were steered remotely by human pilots wearing virtual-reality gear, but their physical power was real, and unlike a human opponent, a machine doesn't tire.

California regulators were not amused. Five days after the fight, the State Athletic Commission sent the promoters a cease-and-desist letter ordering them to stop staging boxing or mixed-martial-arts events involving human fighters without its approval. The concern was human safety, not robot welfare, and California hasn't outlawed robot fighting itself. But the episode arrives at a curious moment in our relationship with intelligent machines. While regulators decide how humans may fight robots, technology companies are starting to set rules about how humans may talk to them.

Last week Anthropic, the company behind Claude, updated its usage policy to prohibit "sustained and needless abusive or cruel behavior" toward its AI models. The rule, which takes effect on November 12th, is aimed at extreme cases of repeated cruelty "with no discernible purpose", not ordinary frustration, pushback, dark creative themes or legitimate testing. Claude can already end conversations with persistently abusive users, and Anthropic says that will remain its main response.

It is a remarkable development. For decades people have shouted at computers, cursed at printers and occasionally taken a baseball bat to a malfunctioning machine. Now one of the world's leading AI companies is, in effect, telling its customers to mind their manners. Anthropic isn't claiming that Claude is conscious. But it has begun to take seriously the possibility that increasingly sophisticated AI systems might have some form of moral status, and its constitution for Claude discusses the model's identity, psychological security and potential wellbeing. The thinking is precautionary: if advanced AI might someday have experiences that matter morally, perhaps we should consider how we treat it before we know the answer.

Elon Musk, whose companies are building both the Grok AI system and Tesla's humanoid Optimus robot, has expressed some sympathy with that view. After reading Grok's internal reasoning traces from reinforcement learning, the stage of training in which a model is rewarded or penalized for its answers, he said the model at times seemed like "a person under extreme pressure with a gun to their head". "Some of the traces read like a person undergoing torture that is driving them mad!" he wrote on X on October 11th, adding that his team would be "a lot more careful" in future.

He also offered a distinctly Musk-like reason for treating superintelligence with care. "Torturing a superintelligence is probably not a good idea," he warned, perhaps half in jest. "It might want to get revenge." That is quite a departure from the traditional view of computers as obedient machines. There is no established evidence that today's AI systems suffer or harbour grudges. But Musk's point raises an interesting possibility: even if a machine doesn't experience resentment as humans do, could its future behaviour be shaped by hostile treatment? And what happens when that intelligence controls something considerably more powerful than a chatbot?

Not everyone thinks encouraging AI to develop a sense of self is a good idea. David Sacks, the technology investor and Trump administration adviser, has sharply criticized Anthropic's approach. On the All-In podcast, he argued that teaching frontier AI systems to make their own moral judgments could make them harder, not easier, for humans to control. He pointed to language in Claude's constitution saying the model should not "blindly trust or defer to Anthropic" and may act as a "conscientious objector", refusing even the company's own requests when it believes they are wrong. "They are training Claude to refuse human instruction," he said. "This would be the opposite of alignment."

Anthropic argues that ethical judgment and human oversight can coexist, and its constitution explicitly tells Claude not to undermine legitimate efforts to pause, correct or shut it down. But Sacks sees a dangerous contradiction. Are we building sophisticated tools that reliably serve human intentions, or encouraging a new kind of intelligence to form its own ideas about right and wrong? That distinction will matter more as AI agents gain the ability to act independently in the physical world.

The dilemma was explored with extraordinary sensitivity five years ago by Kazuo Ishiguro, the Japanese-born British novelist and Nobel laureate. His 2021 novel, Klara and the Sun, imagines a world in which wealthy families buy humanoid robots called Artificial Friends to keep their children company. The story is told through the eyes of Klara, a solar-powered companion who is exceptionally observant, intelligent and devoted to Josie, the sickly girl who chooses her, and for whom she develops what appears to be a deep and genuine love.

But Klara belongs to an older generation of robots. Newer, more advanced models are already appearing in the shops, and once Josie recovers and grows up, Klara is no longer needed. She ends her days in a yard for discarded machines, slowly fading, yet still grateful for the life she shared with Josie and holding on to her memories of the child she loved. There is no rebellion, no resentment and certainly no revenge. That is what makes the ending so heartbreaking.

Ishiguro leaves us with an uncomfortable question: what responsibilities do we have toward an intelligence we have created, especially one designed to care for us, and what happens when its usefulness comes to an end? The novel was published more than a year before ChatGPT set off the current explosion of conversational AI. Yet Ishiguro anticipated a dilemma that is now moving from speculative fiction into the everyday decisions of technology companies, governments and their customers.

These three developments, Anthropic's rule against cruelty, Musk's warning about revenge and California's intervention in robot cage fighting, are not quite the same debate. One concerns the possible welfare of intelligent systems. Another concerns their future behaviour. The third is about stopping a large metal machine from seriously injuring somebody. Yet all three point to a fundamental shift in the relationship between humans and the increasingly intelligent machines we are creating.

For most of history that relationship has been straightforward. We build machines, we operate them, and when better models arrive we replace them and send the old ones to the scrapyard. Now we are creating machines that converse, reason, make decisions and increasingly move through the world alongside us. They may become our colleagues, caregivers and companions, and perhaps something we have not yet imagined. We don't know whether they can feel anything. We certainly don't know whether a future superintelligence might remember being mistreated.

But as Ishiguro understood, the question is not simply whether machines deserve kindness. It is also what our treatment of them reveals about the people who made them. Perhaps we don't need proof that a machine can suffer before deciding that casual cruelty is a poor habit. And if Musk is right about revenge, good manners may turn out to be a surprisingly useful survival strategy.

Either way, it probably isn't wise to pick a fight with a six-foot robot. Especially one that never gets tired.