We Need Positive Visions for AI Grounded in Wellbeing

An essay published by The Gradient argues for grounding beneficial AI in individual wellbeing and societal health, and for developing plausible positive visions of a society with capable AI. The authors describe their aim as walking a pragmatic middle path between strident optimism and pessimism about AI.

The essay states that human wellbeing has been debated for many years by philosophers, economists, psychotherapists, psychologists and religious thinkers without consensus, but that agreement exists around concrete factors such as supportive intimate relationships, meaningful and engaging work, a sense of growth and achievement, and positive emotional experiences. It adds that wellbeing across years and decades depends on societal infrastructure including education, government, the market and academia.

The authors write that by common measures of individual wellbeing such as suicide rate, loneliness and meaningful work, and of societal wellbeing such as trust in institutions, shared sense of reality and political divisiveness, society is not doing well, and that their impression is that AI is complicit in that decline. They conclude that no fundamental obstacle prevents synthesizing the science of wellbeing with machine learning.

The essay argues AI will shock societal infrastructure, comparing it to social media: Facebook launched only twenty years ago, yet social media has subverted news media and the informational commons, addicted users to likes, and displaced meaningful human connection, and the authors believe capable AI's impact will exceed that of social media.

A third conclusion holds that foundation models and the arc of their future deployment are critical. Noting that GPT-2 was released only in 2019, the authors say that if future models are much more capable and engage with the world with greater autonomy, their entanglement with society can be expected to increase. They suggest enabling models to understand and support wellbeing through new algorithms, wellbeing-based evaluations and wellbeing training data, and point to a final section highlighting leverage points for realizing human benefit in practice.