AI Policy Wiki
Dashboard

Acceptable Use Policies for Foundation Models — Kevin Klyman (AIES 2024)

high confidence · updated 2026-06-06

Stanford CRFM / Harvard Belfer paper analyzing acceptable use policies (AUPs) across major foundation-model developers as a class of governance instruments. AUPs are now the most-deployed cross-developer policy artifact and the primary mechanism for category-level usage restrictions.

"Acceptable Use Policies for Foundation Models" is a paper by Kevin Klyman (Stanford Center for Research on Foundation Models and Harvard Belfer Center), published 22 October 2024 in the Proceedings of the Seventh AAAI/ACM Conference on AI, Ethics, and Society (AIES 2024). It examines acceptable use policies (AUPs) as a class of governance instrument and offers comparative analysis across major closed-weight and open-weight foundation-model developers.

Summary of argument

The paper defines acceptable use policies (AUPs) as the legally binding, terms-of-service-adjacent documents that prohibit specific categories of foundation-model use. Klyman argues that AUPs are the most-deployed cross-developer governance instrument and the primary mechanism for category-level usage restrictions, while remaining understudied as a policy artifact. The paper compares AUPs across major developers, including OpenAI, Anthropic, Google, Meta, and Mistral, and identifies both convergent and divergent patterns in their design.

Relation to other wiki pages

The paper provides a cross-developer comparative frame for several AUP-adjacent governance artifacts tracked elsewhere. OpenAI's OpenAI Model Spec is its most detailed policy artifact in this space, with the IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs (OpenAI, March 2026) paper as a training-side counterpart. Anthropic's safety and transparency frameworks, Anthropic's Responsible Scaling Policy (Version 3.1) and A Framework for AI Development Transparency (Anthropic), operate alongside its AUP, and Klyman's comparison clarifies the layering of AUP, responsible scaling policy, and system card.

The dispute documented in Anthropic Claude Gov + Pentagon Dispute (2025–2026) and Clawed centered on Claude AUP restrictions on autonomous weapons and mass surveillance, a category of usage restriction the paper analyzes directly. The complaints in Raine v. OpenAI, Inc. and Garcia v. Character Technologies, Inc. reference AUP-adjacent design choices. The paper also anchors Kevin Klyman (planned) as an academic voice on foundation-model governance.

Provenance

Author: Kevin Klyman (Stanford Center for Research on Foundation Models and Harvard Belfer Center). Publication: Proceedings of the Seventh AAAI/ACM Conference on AI, Ethics, and Society (AIES 2024), 22 October 2024. It is the first comparative academic primary source on AUPs as a governance-instrument category among the AUP-adjacent pages here.

Relationships