{"id":"a223b131-9387-420e-b726-371ebe9deb08","arxiv_id":"2505.06464","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A taxonomy of 98 openness concepts from across disciplines reveals that AI openness discussions overemphasize access, inspection, reuse, and organic behavior while underrepresenting fairness, diversity, autonomy, and non-isolation.","lead":"This paper builds a map of what openness means across many fields of study, then uses that map to show which meanings are missing from today's debates about open AI. It gives AI researchers and policymakers a vocabulary to see what kind of openness they are really choosing when they release a model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5's gap conclusion (AI openness misses autonomy, fairness, diversity) rests on an unoperationalized reading of AI openness texts; no coding rule or inter-rater reliability is provided, so the central claim's empirical content is unverified.","rationale":"The reader correctly notes that the taxonomy's extraction may miss concepts. My concern is downstream: even with a perfect taxonomy, the paper's claim about what AI openness does/does not represent is based on an unsystematic narrative. The paper's own limitations only acknowledge taxonomy incompleteness, not the lack of a reproducible procedure for the situational analysis. This is not a fatal flaw for a conceptual contribution, but the strength of the word 'does not represent' (Section 5) exceeds the evidentiary basis. A condition that the authors either add a systematic, inter-coded analysis of AI openness literature or explicitly reframe the gaps as hypotheses would bring the claims in line with the method. Thus the verdict should be CONDITIONAL rather than ACCEPT.","tokens_in":33294,"tokens_out":6130,"duration_ms":62792,"concrete_test":"Select a stratified sample of AI openness publications (including the references in Sections 4.1-4.3 and a broader keyword search for 'open AI' or 'open source AI'). Provide two independent coders with a codebook that defines 'represent' as the sub-theme appearing as a definitional requirement or core objective, and ask them to code each publication for sub-themes. Compute Cohen's kappa and compare coded emphases against the Section 5 claims. If autonomy, fairness, or diversity are coded as represented in a substantial fraction of papers, or if kappa is below 0.6, the gap conclusion is not robust.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central contribution has two parts: the taxonomy and the diagnosis of AI openness gaps. The diagnosis (Section 5: AI openness emphasizes access, inspectability, reuse, organic, democratization but 'does not represent' non-isolation, autonomy, fairness, diversity) is not derived from the systematic topic-modeling pipeline; it comes from the authors' narrative reading of selected AI openness literature in the 'X in AI Openness' paragraphs of Sections 4.1-4.3. No corpus of AI openness texts is defined, no inclusion criteria for those paragraphs are given, and no coding scheme is specified that would let a reader decide when a sub-theme is 'represented' versus merely mentioned. This matters because Section 4.2.6 acknowledges that OSI's Open Source AI Definition increases developers' autonomy, and Section 4.3.1 discusses fairness benefits of model openness; the paper's conclusion that these are not represented therefore depends on an implicit, unexplained distinction between a sub-theme as a core principle and as a side effect. The limitations (Section 5.3) address completeness of the concept-extraction phase but not the reliability of the AI-situating phase, leaving the central gap claim without an evidentiary anchor. As a result, the taxonomy may be a valid instrument, but its application to AI openness is not demonstrably reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a taxonomy of openness derived from a large multidisciplinary literature corpus (224,123 English abstracts, 1980–2024, with “open|openness” in the title), using LDA topic modeling to identify 98 openness concepts and reflexive thematic analysis to organize them along two dimensions: three themes (Interactivity, Freedom, Inclusiveness) and three definitional approaches (properties, afforded actions, desired effects). It then uses this taxonomy as an instrument to situate the current AI openness discourse, concluding that AI openness emphasizes access, inspectability, reuse, organicity, and democratization while under-representing non-isolation, autonomy, fairness, and diversity. The paper recommends defining AI openness through all three approaches and considering openness at multiple scales beyond the model.","tokens_in":33625,"tokens_out":3385,"duration_ms":34774,"significance":"The strengths of the paper are its scale and transparency: the corpus construction is described in detail, the topic-modeling pipeline and hyperparameter choices are documented, the appendix lists all topics with stability scores and associated concepts, and the qualitative analysis follows an established method. If the taxonomy is accepted, it provides a genuinely interdisciplinary instrument for AI openness discussions and could support future empirical work on open-washing, licensing design, and AI governance. The paper also makes a falsifiable diagnostic claim about which openness dimensions are missing from AI discourse. However, the central diagnostic claim is currently under-supported, so the significance of the contribution depends on whether that claim can be made reproducible.","major_comments":[{"comment":"The opening claim that openness in AI “does not represent” non-isolation, autonomy, fairness, and diversity is not consistent with the paper's own earlier sections. Section 4.2.6 explicitly states that the OSI Open Source AI Definition “increases developers' autonomy” by allowing reuse without permission, and Section 4.3.1 states that inspectability enabled by openness “improves the fairness of the models” and that openness “facilitates the redistribution of access” to help close inequality gaps. Unless “represent” is given a precise technical meaning, such as “specified as a required property in a definition,” the conclusion as written is contradicted by the manuscript's own evidence. The distinction between a concept being a side effect of openness and being a core definitional principle is essential to the claim but is never operationalized.","section":"Section 5, first paragraph"},{"comment":"The diagnosis of gaps in AI openness is not derived from the systematic topic-modeling pipeline; it comes from narrative paragraphs that cite selected AI openness literature. No corpus of AI openness texts is defined for this phase, no inclusion criteria are given for the cited works, and no coding scheme is specified that would allow a reader to decide when a sub-theme is “represented” versus merely mentioned. The limitations in Section 5.3 address completeness of the concept-extraction phase but not the reliability of the AI-situating phase. As a result, the central claim that AI openness misses autonomy, fairness, and diversity is not demonstrably reproducible, even though it may be true.","section":"Sections 4.1–4.3, “X in AI Openness” paragraphs"},{"comment":"The recommendation to define AI openness through properties, afforded actions, and desired effects is interesting but is presented without a concrete test of its applicability. For example, the paper recommends “being accompanied by a model card” as an objectively verifiable property, yet Section 4.1.2 notes that documentation does not always translate to auditability. The authors should clarify how their three-approach framework handles trade-offs between verifiability and ethical objectives, or at least acknowledge that the recommendation is a design proposal rather than an empirically validated outcome.","section":"Section 5.1"}],"minor_comments":[{"comment":"There is a typo: “communities whith peer-recognition” should read “communities with peer-recognition.”","section":"Section 4.1.5"},{"comment":"The sentence “For example, access to open source software [108], open educational resources, and [160]open standards should be non-discriminatory” has misplaced citation punctuation; it should read “open educational resources [160] and open standards.”","section":"Section 4.3.1"},{"comment":"Several table entries contain typos: “open mindness,” “open educational ressources,” “Massive Open Online Couse,” “open univeristy,” and “open-door laminopy” (likely “laminoplasty”). These should be corrected.","section":"Appendix A.3, Tables 5–7"},{"comment":"The title-based filter “open|openness” excludes concepts such as “openwashing” or uses of “open” only in the abstract; this is acknowledged in Section 5.3, but a brief sentence in Section 3.1 noting the intended scope would help readers assess coverage.","section":"Section 3.1"},{"comment":"The distinction between “availability” and “accessibility” is valuable but is illustrated with only a few examples; Table 1 lists many concepts without indicating which ones demonstrate the distinction. A footnote or table column clarifying this contrast would improve readability.","section":"Section 4.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid methodological core and fits FAccT's scope, but the central gap claim needs to be either operationalized or substantially softened. I would not reject it, because the taxonomy itself is a useful contribution and the Section 5 claims can be fixed with a defined corpus/coding scheme or a more careful formulation. The internal contradiction between Section 5 and Sections 4.2.6/4.3.1 should be addressed directly by the authors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution of this paper is the taxonomy itself: two dimensions, three themes, three definitional approaches, built from 98 openness concepts across disciplines. That is a real synthesis. The method is described in enough detail to be repeatable in spirit: 224k abstracts, LDA with hyperparameter search, stability and coherence checks, manual concept extraction, reflexive thematic analysis. I believe them when they say no prior framework combines the breadth and the structure. The links to AI openness in each section are thoughtful and the distinction between properties, afforded actions, and desired effects is a genuinely useful lens.\n\nThe soft spot is the gap claim in Section 5. The paper states that AI openness emphasizes access, inspectability, reuse, organic, and democratization but does not represent non-isolation, autonomy, fairness, and diversity. That conclusion does not come from the systematic topic-modeling pipeline. It comes from a narrative reading of the AI openness literature in the 'X in AI Openness' subsections. No corpus of AI openness texts is defined, no inclusion criteria for those paragraphs, no coding scheme, no inter-rater reliability. The stress-test note is right: the paper even acknowledges in Section 4.2.6 that OSI's definition increases developers' autonomy, and Section 4.3.1 discusses fairness benefits. So the 'does not represent' claim depends on an implicit line between a concept as a core principle and as a side effect. That line is not drawn in the paper.\n\nIs this fatal? No. The taxonomy still stands as a useful instrument, and the gap diagnosis is a reasoned opinion backed by a lot of reading. But it is an opinion, not a measurement. I would temper the reader report's moderately high confidence in the central claim. The limitations section (5.3) addresses completeness of concept extraction but not reliability of the AI-situating phase. I would flag that explicitly in any review.\n\nMinor: data and code are not released. For a paper that leans on topic modeling, that is a reproducibility gap, but acceptable for a qualitative taxonomy if the authors share the concept list and coding materials.\n\nWho is this for? People working on AI governance, open source AI policy, responsible AI. It will be cited. It deserves a serious referee; a good reviewer would ask for a transparent coding of the AI-situating phase or a rephrasing of the gap claim as hypothesis rather than finding. I would accept after revision, not desk-reject, and not accept as-is without that change.","headline":"The taxonomy is a solid instrument; the claim about what AI openness misses is a reasoned reading, not a measured result.","tokens_in":34106,"tokens_out":2963,"would_cite":true,"duration_ms":25229,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a two-dimensional taxonomy of openness built from 98 cross-disciplinary concepts shows AI openness emphasizes access, inspectability, reuse, organic freedom, and democratization while omitting non-isolation…","keywords":["artificial intelligence","openness","open source","taxonomy","topic modeling","latent dirichlet allocation","interdisciplinarity","responsible AI"],"falsifier":"Run the same topic-modeling pipeline on a corpus that adds non-English abstracts, full texts, or a relaxed title filter; if concepts such as labor openness or epistemic openness appear that fit no existing sub-theme, the taxonomy's completeness—and the claim that AI openness omits certain dimensions—would be an artifact of the corpus.","tokens_in":1418,"feed_emoji":"🔓","tokens_out":1313,"duration_ms":54269,"temperature":0.7,"pith_summary":"The paper argues that 'openness' in AI has been defined almost entirely through open source software, which narrows what counts as open. To broaden that scope, it builds a taxonomy of openness from 98 concepts drawn from 224,123 abstracts across disciplines, organized by two dimensions: themes (Interactivity, Freedom, Inclusiveness) and approaches to definition (properties, afforded actions, desired effects). It then uses the taxonomy to show that current AI openness discussions center on access, inspectability, reuse, organic freedom, and democratization, while largely ignoring non-isolation, autonomy, fairness, and diversity. If the taxonomy is right, AI openness conversations are missing several dimensions that other fields treat as central.","feed_headline":"Openness in AI misses fairness, diversity, and autonomy","feed_subtitle":"AI openness centers on access and reuse while skipping fairness, diversity, non-isolation, and autonomy.","key_machinery":"The central object is the taxonomy of openness, built from 98 concepts extracted by LDA topic modeling of 224,123 English abstracts with 'open' or 'openness' in the title, then analyzed through reflexive thematic analysis. The taxonomy has two dimensions: themes (Interactivity, Freedom, Inclusiveness) and approaches to defining openness (properties, afforded actions, desired effects). It works as an instrument for situating current AI openness discussions and for identifying which parts of the broader openness landscape are missing.","core_discovery":"The central claim is that openness is not a single concept but a family of related ideas that can be organized by what is opened and how openness is defined. Across 98 concepts, three themes emerge—Interactivity (access, inspectability, distribution, reuse, collaboration), Freedom (no obstacle, organic, non-isolation, broader boundary, undetermined, autonomy), and Inclusiveness (fairness, diversity, democratization)—and definitions work through intrinsic properties, afforded actions, or desired effects. Situating AI openness within this taxonomy, the paper finds that AI discussions emphasize access, inspectability, reuse, organic, and democratization, but do not represent non-isolation, autonomy, fairness, and diversity. The authors conclude that current AI openness frameworks, including the Open Source AI Definition and the Model Openness Framework, are therefore narrower than the broader concept of openness.","pith_inferences":["If the taxonomy were treated as an audit checklist, open-washing criticisms could be restated as missing dimensions such as fairness or autonomy rather than as loose rhetoric.","The taxonomy implies a measurable research program: mapping existing AI frameworks onto the two dimensions could produce a profile per framework, making the claim about gaps testable.","Because the corpus is limited to English-language academic abstracts with 'open' or 'openness' in the title, popular or practitioner notions of openness might not fit the taxonomy; extending the method to non-academic sources could reveal additional sub-themes."],"forward_implications":["AI openness definitions that list only afforded actions (access, modify, share) omit objective properties and ethical effects, so they cannot by themselves guarantee transparency or democratization.","Concepts like fairness and diversity are treated as consequences of AI openness rather than as defining principles, unlike in education and other fields.","Non-isolation—the involuntary permeability and leakage risks of open systems—is largely absent from AI openness discussions despite being a real security concern.","Autonomy, the freedom to act without permission, is only partially represented in AI openness through the ability to download and copy components.","Opening at the level of AI systems does not automatically open the AI field; field-level properties may need to be enforced separately."],"supporting_citations":[{"why":"Supplies the classic open source definition that AI openness is contrasted with and that the taxonomy places in a broader context.","marker":"[108]"},{"why":"Provides the current Open Source AI Definition that the paper uses as a key example of AI openness framed through afforded actions.","marker":"[64]"},{"why":"Offers the Model Openness Framework, a leading existing framework for assessing AI openness that the paper compares against.","marker":"[155]"},{"why":"Documents open washing in generative AI and motivates the need for a broader, less binary notion of AI openness.","marker":"[84]"},{"why":"Argues that AI release is a gradient and composite notion, supporting the paper's claim that a binary open/closed framing is too limiting.","marker":"[139]"},{"why":"Presents an earlier framework of openness principles, which the taxonomy extends by adding structure and cross-disciplinary scope.","marker":"[125]"},{"why":"Identifies inclusivity and freedom as core aspects of openness across disciplines, feeding directly into the taxonomy's themes.","marker":"[109]"},{"why":"Provides the stability metric used to filter topic model results, a load-bearing part of the concept extraction method.","marker":"[60]"}],"fun_headline_variants":["AI openness skips fairness, diversity, and autonomy","Openness in AI misses key values like fairness","Current AI openness is too narrow for society","Rethinking openness in AI: beyond access and reuse","AI openness overlooks non-isolation and autonomy"],"cache_read_input_tokens":36224,"weakest_assumption_plain":"The whole exercise depends on the 98 concepts that topic modeling happened to surface from a corpus limited to English-language article titles containing 'open' or 'openness'; if that pipeline missed concepts, the gaps found in AI openness may be artifacts of the corpus rather than real omissions.","fun_headline_variants_meta":{"raw":{"variants":["AI openness skips fairness, diversity, and autonomy","Openness in AI misses key values like fairness","Current AI openness is too narrow for society","Rethinking openness in AI: beyond access and reuse","AI openness overlooks non-isolation and autonomy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000357,"raw_usage":{"total_tokens":1911,"prompt_tokens":898,"completion_tokens":1013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":941}},"tokens_in":514,"tokens_out":1013,"duration_ms":9266,"temperature":1.0,"reasoning_tokens":941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:41:27.664866+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same topic-modeling pipeline on a corpus that adds non-English abstracts, full texts, or a relaxed title filter; if concepts such as labor openness or epistemic openness appear that fit no existing sub-theme, the taxonomy's completeness—and the claim that AI openness omits certain dimensions—would be an artifact of the corpus.","supporting_citations":[{"cited_title":"Openness","cited_arxiv_id":null,"evidence_quote":"Presents an earlier framework of openness principles, which the taxonomy extends by adding structure and cross-disciplinary scope."}],"review_version":1}