The argument is simple and the authors state it plainly: 'beneficial AI' is meaningless without grounding it in measurable human wellbeing. Written by researchers closer to machine learning than to positive psychology or wellbeing economics, this essay from The Gradient attempts to bridge that gap. Their core claim is that individual wellbeing factors, such as supportive relationships, meaningful work, and positive emotional experience, combined with societal infrastructure like education and government, are the only concrete anchors available for defining what AI should actually optimize for.
The authors point to a specific and uncomfortable data point: Facebook launched 20 years ago, and social media has already subverted the informational commons, displaced human connection, and entrenched addictive feedback loops. Their position is that capable AI will exceed social media's disruption, and GPT-2, released only in 2019 and already considered barely functional, illustrates how fast the capability curve moves. The essay argues that foundation models need wellbeing-based training data and evaluations now, before deployment scale makes course correction exponentially harder.
What makes this worth reading in full is not the conclusion but the working-through: the authors openly document where they got stuck, including the definitional chaos around 'flourishing' across philosophy, psychology, and economics, and how they moved past it. The essay is structured so readers focused on concrete research directions can skip directly to the final section on leverage points. That section, covering new algorithms, model evaluations, and training data grounded in wellbeing science, is where the practical argument lives.
[READ ORIGINAL →]