Open-washing

As generative-model providers increasingly used open-source language, established software definitions did not map neatly onto systems made of data, code, weights, documentation, and costly training processes. OSI's 2023 multi-stakeholder effort named open washing as a problem that a new definition should help address. The Linux Foundation's Model Openness Framework later proposed graded release classes across lifecycle components. At FAccT 2024, Liesenfeld and Dingemanse assessed 46 text and image systems across 14 dimensions and argued that openness is composite and gradual rather than a single yes-or-no property.

Sources: s1, s2, s3

An open label can influence procurement, research reuse, regulatory treatment, community trust, and investment. If users receive weights but lack essential licenses, data provenance, code, or documentation, they may be unable to reproduce results, audit claims, understand restrictions, or continue a project after upstream changes. Open-washing also weakens the vocabulary needed to compare release strategies. A component-level assessment is more useful than arguing over a brand label: it records what is available, under which terms, in what form, and whether the release supports inspection, modification, redistribution, and reproducibility.

Sources: s2, s3, s4

A provider calls a model fully open source because downloadable weights are available, but the custom license restricts fields of use, the training data and code are unavailable, and the evaluation recipe cannot be reproduced. A reviewer should preserve the exact marketing statement, inventory each released component and permission, and compare the result with the definition or framework invoked by the claim. The evidence may support describing the release as open weights or partially open without automatically reaching a legal conclusion that the provider acted deceptively.

Sources: s2, s3

Complete disclosure is not always possible or desirable: privacy, copyright, security, contractual, and practical constraints can limit release. Those constraints do not themselves prove open-washing if claims are precise about what is and is not open. Conversely, a permissive weight license does not disclose the training process. This entry does not adjudicate named providers or offer legal advice. Reviewers should use current license text and a stated openness framework, distinguish factual inventory from normative judgment, and avoid treating openness as a proxy for safety, ethics, or model quality.

Sources: s1, s2, s3