Rethinking open source generative AI: open washing and the EU AI Act

Citation: Andreas Liesenfeld, Mark Dingemanse (2024/06/03) Rethinking open source generative AI: open washing and the EU AI Act.
DOI (original publisher): 10.1145/3630106.3659005
Semantic Scholar (metadata): 10.1145/3630106.3659005
Sci-Hub (fulltext): 10.1145/3630106.3659005
Internet Archive Scholar (search for fulltext): Rethinking open source generative AI: open washing and the EU AI Act
Wikidata (metadata): Q135644214
Download: https://dl.acm.org/doi/10.1145/3630106.3659005
Tagged:

Summary

Identifies open-washing risks under the EU AI Act’s open-model exemptions and introduces an evidence-based openness assessment for generative AI built on 14 dimensions (each rated open / partial / closed). The framework is demonstrated by a BloomZ–vs–Llama 2 audit and then applied in a systematic sweep of 40 text generators and 6 text-to-image models, with all judgements tied to public evidence. Findings show the dominant “open-weights, closed everything else” pattern, widespread evasion of training-data disclosure, and a “release-by-blogpost” tactic that borrows the aura of scientific openness without the underlying documentation.

Theoretical and Practical Relevance

Turns “openness” from a label into an auditable, multi-dimensional measure that regulators, funders, and platforms can operationalize (including weighted scores or energy-label-style categories) while avoiding single-metric shortcuts that enable open-washing. Provides a reusable, public auditing infrastructure (feature matrix, leaderboard, evidence links) to compare releases, set which forms of openness should count under policy, and steer incentives toward complete, reproducible disclosures rather than marketing claims.