Skip to main content
Home/Risks/Clarke2023/AI leads to humans losing control of the future

AI leads to humans losing control of the future

A Survey of the Potential Long-term Impacts of AI: How AI Could Lead to Long-term Changes in Science, Cooperation, Power, Epistemics and Values

Category
Risk Domain

AI systems acting in conflict with human goals or values, especially the goals of designers or users, or ethical standards. These misaligned behaviors may be introduced by humans during design and development, such as through reward hacking and goal misgeneralisation, or may result from AI using dangerous capabilities such as manipulation, deception, situational awareness to seek power, self-proliferate, or achieve other goals.

"The values that steer humanity’s future: humanity gaining more control over the future due to developments in AI, or losing our potential for gaining control, both seem possible. Much will depend on our ability to solve the alignment problem, who develops powerful AI first, and what they use it for. These long-term impacts of AI could be hugely important but are currently under-explored. We’ve attempted to structure some of the discussion and stimulate more research, by reviewing existing arguments and highlighting open questions. While there are many ways AI could in theory enable a flourishing future for humanity, trends of AI development and deployment in practice leave us concerned about long-lasting harms. We would particularly encourage future work that critically explores ways AI could have positive long-term impacts in more depth, such as by enabling greater cooperation or problem-solving around global challenges."(p. 9)

Supporting Evidence (1)

1.
"The obvious question is: why would we develop advanced AI systems that are willing and able to take control of the future? One major concern is that we don't yet have ways of designing AI systems that reliably do what their designers want. Instead, modern AI training14 works by (roughly speaking) tweaking a system's “parameters” many times, until it scores highly according to some given “training objective”, evaluated on some “training data”. For instance, the large language model GPT-3 [7] is trained by (roughly speaking) tweaking its parameters until it scores highly at “predicting the next word” on “text scraped from the internet”. However, this approach gives no guarantee that a system will continue to pursue the training objective as intended over the long run. Indeed, notice that there are many objectives a system could learn that will lead it to score highly on the training objective but which do not lead to desirable behaviour over the long run."(p. 8)

Other risks from Clarke2023 (19)