On the Societal Impact of Open Foundation Models

From AcaWiki
Jump to: navigation, search

Citation: Sayash Kapoor, Rishi Bommasani, Kevin Klyman, Shayne Longpre, Ashwin Ramaswami, Peter Cihon, Aspen Hopkins, Kevin Bankston, Stella Biderman, Miranda Bogen, Rumman Chowdhury, Alex Engler, Peter Henderson, Yacine Jernite, Seth Lazar, Stefano Maffulli, Alondra Nelson, Joelle Pineau, Aviya Skowron, Dawn Song, Victor Storchan, Daniel Zhang, Daniel E. Ho, Percy Liang, Arvind Narayanan (2024/02/07) On the Societal Impact of Open Foundation Models.
Internet Archive Scholar (search for fulltext): On the Societal Impact of Open Foundation Models
Wikidata (metadata): Q135645242
Download: https://arxiv.org/abs/2403.07918
Tagged:

Summary

Defines open foundation models as those with broadly available model weights and frames their societal impact via five distinctive properties—broader access, greater customizability, potential for local inference, irreversibility of release, and weaker monitoring—which jointly explain both benefits (innovation, competition, dispersion of decision-making, transparency) and risks. The paper contributes a six-point framework for assessing marginal risk (risk beyond closed models or preexisting tech) and surveys seven misuse vectors (e.g., biosecurity, cybersecurity, disinformation, NCII/CSAM), concluding that current evidence is generally insufficient to quantify marginal risk in most cases and clarifying why past debates often talk past each other.

Theoretical and Practical Relevance

Offers policy-usable guidance: keep benefits and marginal-risk analysis distinct; require assumptions and evidence when claiming harm; and avoid obligations that presume centralized control (e.g., developer liability or strict watermarking duties) that open-model developers cannot realistically meet. The authors recommend: (i) developers publish which responsible-AI practices they implement vs. delegate downstream; (ii) researchers prioritize empirical tests of marginal risk; (iii) policymakers assess regulatory burden on open-model developers and target interventions to specific misuse vectors; and (iv) competition authorities measure openness-linked benefits (choice, costs) rather than assume them.