In short
A close reading of 'Positive Alignment', a proposal from researchers at DeepMind, OpenAI, Anthropic and Oxford to make human flourishing an explicit technical objective across the AI lifecycle, rather than optimising for short-term revealed preferences as RLHF does. The essay praises its multi-dimensional model of the good life but argues it remains politically naive: its reliance on market 'alignment-as-a-service' contradicts its own critique, and it lacks a theory of change.
“Positive Alignment” is one of the more ambitious recent ideas in AI safety: instead of training models to satisfy whatever a user clicks approval on, train them to support human flourishing, across the short and long term, and across multiple philosophical accounts of what a good life is.
This essay takes the proposal seriously, and admires it. Making flourishing an explicit technical target is a genuine break from the revealed-preference logic that AI inherited from economics and from social media’s attention machine. But it then asks the harder question the paper mostly avoids: how would this actually happen? A proposal that critiques markets for optimising the wrong thing, then hands implementation to a market for “alignment-as-a-service,” needs a theory of change it doesn’t yet have.
The piece is a good example of what Generative Futures tries to do: take frontier AI-safety work on its own terms, then press on its politics.
Key takeaways
- RLHF inherits flawed assumptions from mainstream economics and social-media design, optimising immediate revealed preferences over genuine flourishing.
- Positive Alignment integrates flourishing across the AI lifecycle using four frameworks (hedonic, conative, objective-list, perfectionist) plus role-based norms.
- Its governance model leans on market 'middleware / alignment-as-a-service', reproducing the revealed-preference logic it critiques.
- The proposal lacks political analysis and a theory of change for real-world adoption.
- Treating the technology as the unit of analysis risks missing the broader institutional change required.
Read the full piece
This is a summary. Read the complete essay, with all the sources and argument, on Substack.