IN RECENT months, there has been a flood of claims that superintelligent artificial intelligences could one day lead to the destruction of humanity. Sam Altman, CEO of Open AI, has said the technology could lead to 鈥溾��, while AI pioneer Geoffrey Hinton resigned from Google citing his fear that AI . What鈥檚 more, the Center for AI Safety has on the risk of extinction the technology poses.
These claims are scary 鈥� but not because they are true. They are scary because they are significantly reshaping and redirecting conversations about the impacts of AI and what it means for AI to be developed responsibly.
Advertisement
The narrative that AI might one day develop superhuman intelligence and pose a threat to humanity is a familiar plot line of movies, and has firm roots within the public image of AI. This narrative has always been close at hand in media reporting, but 2023 has proven to be the year that it also found its way into policy and regulatory discussions around AI.
This is troubling, as little to no evidence is given to substantiate the emotive and sensational claims being made, and the limited evidence that is pointed to typically doesn鈥檛 stand up to scrutiny. For example, claims that large language models are demonstrating so-called emergent properties (abilities which they were not trained for but which have somehow emerged) typically centre on unanticipated 鈥� but nevertheless explainable 鈥� functions of the models. When Alphabet CEO Sundar Pichai Google鈥檚 Bard chatbot had responded to prompts in Bengali despite not being trained to do so, the more plausible explanation came from ex-Google employee Margaret Mitchell, who pointed out Bard鈥檚 development included incorporating aspects of its forerunner PaLM, which .
This is yet another reminder that AI models only do what they are programmed to do. Claims of emergent properties are being used to deflect calls for transparency and to close down discussions of developers鈥� responsibilities.
The loudest voices shouting about existential risk are coming from Silicon Valley. This may seem at odds with big tech鈥檚 stake in driving innovation and investment in AI, yet these narratives protect its interests by diverting attention away from big tech鈥檚 current actions to instead speculate about hypothetical future abilities of AI.
By suggesting AI might develop its own intelligence, the focus shifts to how we might hold the technology accountable in the future, rather than how we can hold big tech accountable today. All of this creates an illusion of inevitability, suggesting we are observing the evolution of AI, rather than a series of conscious and controllable decisions by organisations and people.
It is important to consider not just what is being said, but who is being listened to. We have seen Altman being warmly greeted by regulators across Europe, while the voices of people who are negatively affected by AI are barely heard. This is symptomatic of what academic Kate Crawford has termed AI鈥檚 鈥溾��: claims of existential risk are largely coming from affluent white men in positions of power, those least likely to experience the harms of AI today (but with most responsibility for causing them).
We are at a critical moment where governments are grappling with the challenges of effectively regulating AI. Countering narratives of superintelligence and existential risk is crucial to ensure that these discussions centre on the voices of affected communities, rather than the vested interests of big tech, and that they engage with the real impacts of AI, rather than the fantasy and distraction of artificial general intelligence.
Mhairi Aitken is an ethics fellow at The Alan Turing Institute, UK
