Why human–AI relationships need socioaffective alignment

Key Findings

• — Standard alignment framings treat user preferences as given; the paper argues that in sustained relationships preferences are co-produced, so an aligned system can still shape the very preferences it is aligning to
• — Three intrapersonal dilemmas are set out as the substance of the problem: immediate satisfaction against long-term wellbeing, protecting user autonomy, and holding AI companionship alongside the user's human social bonds
• — The dilemmas are framed through basic psychological needs, giving a criterion for distinguishing systems that support human social and emotional nature from systems that exploit it
• — The argument is developed structurally, through the ingredients of human-AI relationships and the proposed socioaffective alignment frame, rather than through new empirical data

Methodology Notes

Conceptual perspective article, peer reviewed, open access under CC BY 4.0, volume 12 issue 1 article 728. It contributes a framework and vocabulary, not measurements, and should be cited as such. Author affiliations span the Oxford Internet Institute and Google DeepMind. Curator-verified at the publisher page on nature.com, from which the title, publication date, section structure and full abstract were read; the Crossref record agrees on title, all five authors, venue, date and licence.

Sources