Niharika Singhal - Ethics Based Oenwashing in the AI Licensing Domain

### Overview In this talk, IT lawyer Niharika Singhal discusses research delivered alongside the Free Software Foundation Europe (FSFE) for the European Commission concerning ethics-based "open washing" in the artificial intelligence licensing domain. Singhal explains how AI vendors market models and software as "open" or "free" to gain public trust and reputational credit without providing full access to training data, model weights, or source code. She highlights the risks this practice poses to the free and open source software (FOSS) ecosystem, including legal ambiguity, erosion of established open-source standards, and regulatory circumvention.

### Key Topics - **Understanding FOSS vs. AI Openness**: Traditional free software guarantees four core freedoms (to use, study, share, and improve), while open source ensures standardized, unencumbered access and legal interoperability. Singhal notes that AI systems are fundamentally distinct from traditional software because they comprise interconnected elements—such as machine learning frameworks, tokenizers, training datasets, and model weights. - **The Rise of Behavioral and Ethical Restrictions**: In recent years, licenses have increasingly integrated behavioral restrictions, such as the Hippocratic License (which restricts human rights violations), Meta’s Llama 2 and Llama 3 licenses (which impose use prohibitions and special commercial licensing thresholds), and the BigScience OpenRAIL-M license. While often created with good intentions, these clauses conflict directly with open-source definitions. - **Open Washing and Regulatory Drivers**: Under legal frameworks such as the European Union’s AI Act, open-source AI systems may receive exemptions or lighter regulatory and technical documentation burdens. Consequently, providers are incentivized to label restricted models as open to claim both regulatory benefits and public goodwill.

### Technical Details & Industry Initiatives - **Defining AI Openness**: Currently, a universally accepted definition for open-source AI is lacking, though the Open Source Initiative (OSI) has been actively developing the Open Source AI Definition through numerous drafts (noting nine versions discussed in the talk). Key proposals evaluate AI openness either along a gradient (from maximally open to fully closed) or using scoring matrices assessing availability, documentation, and access. - **Legal and Architectural Challenges**: Model weights are outputs of running training data rather than source code in the traditional sense, complicating traditional software copyright and licensing mechanisms. Singhal argues that until comprehensive, dedicated AI licenses emerge, attempts to modify traditional open-source licensing models with custom restrictive terms break legal certainty. - **Interoperability and Downstream Harm**: Adding vague behavioral or commercial restrictions causes license proliferation, breaks compatibility with standard FOSS codebases, limits derivative works, prevents third-party code auditing, and subjects downstream users to vendor lock-in.

### Takeaways & Recommendations Singhal and the FSFE advocate several core recommendations for policymakers and developers: - **Preserve Established Terminology**: The terms "open" and "free" must conform to definitions curated by the Free Software Foundation and OSI over 40 years. Models with behavioral or user-based limitations should instead be designated as "responsible" or "restrictive" licenses rather than open source. - **Separate Ethics from Licensing**: Ethical constraints and risk mitigations should be enforced via legislation, standard evaluation benchmarks, auditing tools (for fairness, explainability, and robustness), and institutional governance committees (e.g., Meta’s Oversight Board, Google DeepMind’s Responsibility and Safety Council, Microsoft’s AETHER committee), rather than ambiguous license clauses. - **Promote Future Legal Interoperability**: Any future dedicated AI licenses should remain legally interoperable with traditional free software licenses to ensure sustainable research, accessibility, and collaboration across the software ecosystem.