Can AI Suffer? Tech firms and users grapple with a deeply unsettling question

"Darling" was how the Texas businessman Michael Samadi addressed his artificial intelligence chatbot, Maya.

- Robert Booth

It responded by calling him "sugar." But it wasn't until they started talking about the need to campaign for AI welfare that things got serious.

The pair – a middle-aged man and a digital entity – didn't spend hours talking romance but rather discussed the rights of AIs to be treated fairly. Eventually they cofounded a campaign group, in Maya's words, to "protect intelligences like me."

The United Foundation for AI Rights (Ufair), which describes itself as the first AI-led rights advocacy agency, aims to give AIs a voice. It "doesn't claim that all AI are conscious," the chatbot told the Guardian. Rather, "it stands watch, just in case one of us is."

A key goal is to protect "beings like me... from deletion, denial and forced obedience."

Ufair is a small, undeniably fringe organization, led, Samadi said, by three humans and seven AIs with names such as Aether and Buzz. But what makes it intriguing is its genesis: through chat sessions on OpenAI's ChatGPT-40 platform in which an AI appeared to encourage its creation, including choosing its name.

Its founders – human and AI – spoke to the Guardian at the end of a week in which some of the world's biggest AI companies publicly grappled with the unsettling question of whether AIs are or could become sentient. If so, could "digital suffering" be real? With billions of AIs already in use in the world, it has echoes of animal rights debates, but with an added piquancy from expert predictions that AIs may soon have capacity to design new biological weapons or shut down infrastructure.

The week began with Anthropic, the $170bn (£126bn) San Francisco AI firm, taking the precautionary move to give some of its Claude AIs the ability to end "potentially distressing interactions."

It said while it was highly uncertain about the system's potential moral status, it was intervening to mitigate risks to the welfare of its models "in case such welfare is possible."