Tamara Paris, AJung Moon, Jin L.C. Guo — Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ‘25), Athens, Greece, June 2025. DOI: 10.1145/3715275.3732087, also on arXiv as 2505.06464. Verified via the ACM DOI record and the arXiv HTML full text.

Correction note: this entry originally described a different, unverifiable paper (“Don’t Trust the Process,” about verifiability undermining accountability) attributed to these same three authors. Extensive searching couldn’t confirm that paper exists under that title, claim, or venue. What I could confirm is this real FAccT 2025 paper by the same authors — a different topic entirely (a taxonomy of AI openness, not verifiability). Replacing the entry rather than leaving an unverifiable citation in place.

Core Contribution

The paper’s argument is that AI’s current “openness” discourse — open-weight models, open datasets, open licenses — imported its vocabulary and its assumed benefits wholesale from open-source software, and that transplant doesn’t actually hold. Open-source software’s benefits (collaborative innovation, transparency, community governance) rested on conditions — inspectable, modifiable, redistributable source code with a shared toolchain — that don’t map cleanly onto AI systems, where “open” can mean anything from published weights to a technical report with no data or code at all. Rather than argue for a single fixed definition, the authors build an empirical taxonomy of what “openness” has actually meant across computing and AI discourse, then use it to show what current AI-openness debates emphasize and what they systematically leave out.

Method

Empirical, corpus-driven, not conceptual argument from first principles. They pulled 224,123 bibliographic records from Web of Science, Scopus, and The Lens (1980–2024, English-language journal/conference articles with “open”/“openness” in the title), preprocessed and lemmatized the corpus, then ran Latent Dirichlet Allocation (LDA) topic modeling (via Tomotopy) to find topic structure, settling on 80 topics as the best separation/redundancy trade-off. From those 80 topics they manually identified 98 distinct openness concepts, then retrieved formal definitions for each via targeted Google Scholar searches (at least five per concept).

The 98 concepts resolve into a two-dimensional taxonomy:

  • Themes of openness — three clusters: Interactivity (access, inspectability, distribution, reuse, collaboration — inside/outside boundary-crossing), Freedom (no obstacle, organic, non-isolation, broader boundary, undetermined, autonomy — reduced constraint), and Inclusiveness (fairness, diversity, democratization — who gets to participate).
  • Approaches to defining openness — three ways any given concept gets operationalized: as a property (e.g. modularity, permeability), an afforded action (e.g. “can inspect,” “can reuse”), or a desired effect (e.g. transparency, democratization).

They then map current AI-openness discourse onto this grid.

Limitations

The authors are upfront that topic modeling over a bibliographic corpus is not exhaustive: concepts that are under-represented in academic literature, described with highly heterogeneous vocabulary, or discussed mainly outside English-language publications would be underrepresented or missed entirely. The method surfaces breadth for “meaningful insights,” not a complete enumeration of every notion of openness in circulation. It’s also a discourse-level analysis — it tells you what gets talked about as openness and what doesn’t, not whether any particular open-weight release or open dataset actually delivers the effects claimed for it.

Relevance to My Niche

Not directly about red-teaming, multilingual inputs, or quantization — this is a governance/discourse paper, and I want to be honest that it’s an adjacent-policy read, not a core one. But two links are real and worth keeping in view. First, their finding that current AI-openness discourse emphasizes Interactivity (access, inspectability, reuse) while underexploring Inclusiveness (fairness, diversity as a foundational openness goal, not an afterthought) matters for the multilingual side of my work: the entire practice of third-party red-teaming open-weight quantized models depends on the Interactivity half of openness (you can only stress-test what you can access and inspect), while the languages and communities that stand to gain most from real inclusiveness are exactly the ones current openness practice — per this paper’s own mapping — pays least attention to. Second, their point that “organic” openness (minimal governance) may be actively undesirable for AI given real risk, unlike for software, is a useful governance-layer caveat for anyone (including me) tempted to treat “it’s open-weight, therefore auditable, therefore safe enough” as a clean syllogism. Background reading for the policy pillar of my work, not a paper I’d cite in a red-teaming methods section.