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

May combine two claims.

A transparency index released as a preprint uses a wide range of indicators and methods, including interviews, but does not open up individual data points for scrutiny or contestation.

“One is a transparency index released as a preprint [8] and widely publicised Ivy League press release that uses a wide range of indicators and methods, including interviews, but which does not open up individual data points for scrutiny or contestation.”

• cites [8] [8] Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, and Percy Liang. 2023. The Foundation Model Transparency Index. https://doi.org/10.48550/arXiv.2310.12941 arXiv:2310.12941 [cs]. - sourceDisputed

  The paper describes a 100-indicator transparency index, but its account of access and contestation says the opposite of the claim. In §1.3, it says scores and justifications are publicly available; in “Company feedback,” it says companies were invited to contest individual scores and that eight submitted rebuttals. Our copy provides section labels but no page numbers for these passages. (LLM, conf=0.90)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Sociological research has established that multinationals like Starbucks and Nestlé effectively co-opted the notion of fair trade.

“Sociological research has established that this resulted in these multinationals effectively co-opting the notion of fair trade.”

• cites [26] [26] Daniel Jaffee. 2012. Weak Coffee: Certification and Co-Optation in the Fair Trade Movement. Social Problems 59, 1 (Feb. 2012), 94-116. https://doi.org/10.1525/sp. 2012.59.1.94 - sourceSupported

May combine two claims.

Open generative AI systems are on the rise, with small players and academic initiatives leading the way in open innovation and scientific documentation and several larger corporations joining the fray by releasing models billed as open.

“Open generative AI systems are on the rise, with small players and academic initiatives leading the way in open innovation and scientific documentation [20, 32, 61] and several larger corporations joining the fray by releasing models billed as 'open'.”

• cites [20] [20] Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, Shane Arora, David Atkinson, Russell Authur, Khyathi Raghavi Chandu, Arman Cohan, Jennifer Dumas, Yanai Elazar, Yuling Gu, Jack Hessel, Tushar Khot, William Merrill, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew E. Peters, Valentina Pyatkin, Abhilasha Ravichander, Dustin Schwenk, Saurabh Shah, Will Smith, Emma Strubell, Nishant Subramani, Mitchell Wortsman, Pradeep Dasigi, Nathan Lambert, Kyle Richardson, Luke Zett sourceSupported
• cites [32] [32] Andreas Liesenfeld, Alianda Lopez, and Mark Dingemanse. 2023. Opening up ChatGPT: tracking openness, transparency, and accountability in instructiontuned text generators. In Proceedings of CUI'23 . Eindhoven. https://opening-upchatgpt.github.io/ - sourceSupported
• cites [61] [61] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron G sourceSupported

Probably not a substantial claim: may not pass S2 (checkable).

The rise of large language models and text-to-image generators means that a different approach to open source classification is needed.

“However, the rise of large language models and text-to-image generators means that a different approach is needed [47].”

• cites [47] [47] Dirk Riehle. 2023. The Future of the Open Source Definition. Computer 56, 12 (Dec. 2023), 95-99. https://doi.org/10.1109/MC.2023.3311648 Conference Name: Computer. - sourceCould not verify

  The OPEN PROJECT GOVERNANCE and OPEN SOURCE, WHAT NEXT? sections discuss ways to classify software projects and certify vendor behavior, but neither connects those approaches to large language models or text-to-image generators. Searches for those technologies and for generative AI found no passage addressing the claimed reason a different approach is needed. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The multi-faceted nature of openness is also seen in initiatives that aim to codify transparency from upstream resources to downstream uses and in the Linux Foundation's model openness framework.

“The multi-faceted nature of openness is also seen in initiatives that aim to codify transparency from upstream resources to downstream uses [8] and in a proposal like the Linux Foundation's model openness framework [59], which shows many likenesses to the dimensions we use here and introduced in [32].”

• cites [8] [8] Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, and Percy Liang. 2023. The Foundation Model Transparency Index. https://doi.org/10.48550/arXiv.2310.12941 arXiv:2310.12941 [cs]. - sourceSupported
• cites [59] [59] Matt White, Ibrahim Haddad, Cailean Osborne, Xiao-Yang, Liu, Ahmed Abdelmonsef, and Sachin Varghese. 2024. The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency and Usability in AI. https://doi.org/10.48550/arXiv.2403.13784 arXiv:2403.13784 [cs]. - sourceSupported
• cites [32] [32] Andreas Liesenfeld, Alianda Lopez, and Mark Dingemanse. 2023. Opening up ChatGPT: tracking openness, transparency, and accountability in instructiontuned text generators. In Proceedings of CUI'23 . Eindhoven. https://opening-upchatgpt.github.io/ - sourceSupported

May combine two claims.

The LLM data used to train the base model is documented in great detail by Bloom, while for Llama only the vaguest details are provided in a corporate preprint.

“The LLM data that was used to train the base model is documented in great detail by Bloom [61], while for Llama only the vaguest details are provided in a corporate preprint: 'a new mix of data from publicly available sources, which does not include data from Meta's products or services' [57].”

• cites [61] [61] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron G sourceSupported
• cites [57] [57] Hugo Touvron, Louis Martin, and Kevin Stone. 2023. Llama 2: Open Foundation and Fine-Tuned Chat Models. (2023). https://ai.meta.com/research/publications/ llama-2-open-foundation-and-fine-tuned-chat-models/ -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

BloomZ's source code is Apache 2.0, an OSI-approved open source licence, while the model weights are released under the Responsible AI Licence (RAIL); Llama 2 is released under Meta's own Community Licence.

“BloomZ has two relevant licences. Its source code is Apache 2.0, an OSI-approved open source licence, while the model weights are released under the Responsible AI Licence (RAIL) [13]. Llama 2 is released under Meta's own Community Licence.”

• cites [13] [13] Danish Contractor, Daniel McDuff, Julia Katherine Haines, Jenny Lee, Christopher Hines, Brent Hecht, Nicholas Vincent, and Hanlin Li. 2022. Behavioral Use Licensing for Responsible AI. In 2022 ACM Conference on Fairness, Accountability, and Transparency . ACM, Seoul Republic of Korea, 778-788. https://doi.org/10. 1145/3531146.3533143 - sourceCould not verify

  The paper’s Limitations and Open Questions section describes Apache 2.0 as an open-source license, but does not identify BloomZ’s source-code or model-weight licenses. Searches for “BloomZ” and “Responsible AI License” found no passage establishing the BloomZ licensing claim attached to this citation. The separate Llama 2 sentence is not attributed to this source. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

May combine two claims.

The scientific documentation of BloomZ includes multiple peer-reviewed papers, while no peer-reviewed papers providing scientific documentation or evaluation of Llama are known currently.

“The scientific documentation of BloomZ also includes multiple peer-reviewed papers , from a scientific description of the multitask fine-tuning procedures [41] to an estimation of the carbon footprint [36] currently one of the very few scientifically vetted sources of data on the energy footprint of training large language models. No peerreviewed papers providing scientific documentation or eval”

• cites [41] [41] Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward Raff, and Colin Raffel. 2023. Crosslingual Generalization through Multitask Finetuning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto, C sourceSupported
• cites [36] [36] Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. 2023. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Journal of Machine Learning Research 24, 253 (2023), 1-15. http://jmlr.org/papers/ v24/23-0069.html - sourcePartly supported

  The Abstract (p. 1) identifies this paper as an estimate of BLOOM’s carbon footprint, and Section 3.1 (p. 2) reports its training energy use. Section 5.1 (p. 7) says few recent LLM papers reported training carbon footprints. Our copy does not establish that this paper was peer-reviewed or otherwise scientifically vetted; the citation does not appear to be offered as evidence for the separate statement about Llama. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

May combine two claims.

Systems like AllenAI's OLMo Instruct, BloomZ and LLM360's AmberChat are approaching full openness status and top the openness leaderboard.

“One is the broad open source approach seen in systems like AllenAI's OLMo Instruct [20], BloomZ [61] and LLM360's AmberChat [33], which are approaching full openness status and top the openness leaderboard.”

• cites [20] [20] Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, Shane Arora, David Atkinson, Russell Authur, Khyathi Raghavi Chandu, Arman Cohan, Jennifer Dumas, Yanai Elazar, Yuling Gu, Jack Hessel, Tushar Khot, William Merrill, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew E. Peters, Valentina Pyatkin, Abhilasha Ravichander, Dustin Schwenk, Saurabh Shah, Will Smith, Emma Strubell, Nishant Subramani, Mitchell Wortsman, Pradeep Dasigi, Nathan Lambert, Kyle Richardson, Luke Zett sourcePartly supported

  The Conclusion and Future Work section describes OLMo as truly open and says the authors released adapted models along with their adaptation code and data. The Adaptation section confirms that they trained OLMo as an instruction-following chat assistant. The paper does not establish that OLMo Instruct tops an openness leaderboard; the other named systems have separate citations. (LLM, conf=0.88)

  Experimental — accuracy not yet measured
• cites [61] [61] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron G sourcePartly supported

  Section 3.5 Training (p. 22) identifies BLOOMZ as finetuned BLOOM models, and Section 3.6 Release (p. 24) describes openness as central to BLOOM and documents its public release. The paper does not say that BLOOMZ approaches full openness or tops an openness leaderboard; those ranking claims are unsupported by this source. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [33] [33] Na Liu, Liangyu Chen, Xiaoyu Tian, Wei Zou, Kaijiang Chen, and Ming Cui. 2024. From LLM to Conversational Agent: A Memory Enhanced Architecture with FineTuning of Large Language Models. https://doi.org/10.48550/arXiv.2401.02777 arXiv:2401.02777 [cs]. - sourceCould not verify

  The citation [33] is attached to LLM360's AmberChat, but our copy does not identify AmberChat or assess its openness. The LLMs and §2.2 LLMs passages discuss other models, and searches for OLMo Instruct, BloomZ, AmberChat, and an openness leaderboard found no relevant passage (page numbers unavailable). (LLM, conf=0.77)

  Experimental — accuracy not yet measured

May combine two claims.

BloomZ's multiple preprints document data curation and finetuning in great detail; in contrast, Llama's single preprint offers fewer details and appears strategically vague on crucial details such as training datasets and instruction tuning.

“BloomZ's multiple preprints document data curation and finetuning [50, 61] in great detail; in contrast, Llama's single preprint offers fewer details and appears strategically vague on crucial details (for instance, training datasets and instruction tuning).”

• cites [50] [50] Teven Le Scao, Thomas Wang, Daniel Hesslow, Lucile Saulnier, Stas Bekman, M. Saiful Bari, Stella Biderman, Hady Elsahar, Niklas Muennighoff, Jason Phang, Ofir Press, Colin Raffel, Victor Sanh, Sheng Shen, Lintang Sutawika, Jaesung Tae, Zheng Xin Yong, Julien Launay, and Iz Beltagy. 2022. What Language Model to Train if You Have One Million GPU Hours? https://doi.org/10.48550/arXiv.2210. 15424 arXiv:2210.15424 [cs]. - sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [61] [61] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron G sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Data sheets, model cards, and system cards are frameworks that can help structure the systematic presentation of metadata about models and systems and that feed into auditing procedures.

“Important constituent elements are the notions of data sheets [18, 25, 38], model cards [40], and system cards [22]: frameworks that can help structure the systematic presentation of metadata about models and systems and that feed into auditing procedures [37].”

• cites [18] [18] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for datasets. Commun. ACM 64, 12 (Nov. 2021), 86-92. https://doi.org/10.1145/3458723 - sourceSupported
• cites [25] [25] Ben Hutchinson, Andrew Smart, Alex Hanna, Emily Denton, Christina Greer, Oddur Kjartansson, Parker Barnes, and Margaret Mitchell. 2021. Towards Accountability for Machine Learning Datasets: Practices from Software Engineering and Infrastructure. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21) . Association for Computing Machinery, New York, NY, USA, 560-575. https://doi.org/10.1145/3442188.3445918 - sourcePartly supported

  Section 4.3, “Dataset Audits and Reviews” (page not available), says dataset-development documentation forms part of an audit trail and that ML audits may review datasets or models. It does not specifically establish that datasheets, model cards, and system cards all feed into auditing procedures. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [38] [38] Angelina McMillan-Major, Emily M. Bender, and Batya Friedman. 2023. Data Statements: From Technical Concept to Community Practice. ACM Journal on Responsible Computing (2023). https://doi.org/10.1145/3594737 Just Accepted. - sourcePartly supported

  Section 2.1 describes documentation toolkits as structured ways to provide information about datasets, models, and AI systems, and says they are being integrated into standards for managing bias. The cited paper does not establish that these frameworks feed into auditing procedures; it also does not discuss system cards specifically. No page number was available in our copy. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [40] [40] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model Cards for Model Reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* '19) . Association for Computing Machinery, New York, NY, USA, 220-229. https://doi.org/10.1145/3287560.3287596 - sourceSupported
• cites [22] [22] Furkan Gursoy and Ioannis A. Kakadiaris. 2022. System Cards for AIBased Decision-Making for Public Policy. http://arxiv.org/abs/2203.04754 arXiv:2203.04754 [cs]. - sourceSupported
• cites [37] [37] Jeanna Matthews. 2020. Patterns and Antipatterns, Principles, and Pitfalls: Accountability and Transparency in Artificial Intelligence. AI Magazine 41, 1 (2020), 82-89. https://doi.org/10.1609/aimag.v41i1.5204 \_eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1609/aimag.v41i1.5204. - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Practical approaches to implement such ideas typically revolve around the systematic collection and curation of data on various aspects of systems.

“Practical approaches to implement such ideas typically revolve around the systematic collection and curation of data on various aspects of systems [2, 28, 46].”

• cites [2] [2] Ariful Islam Anik and Andrea Bunt. 2021. Data-Centric Explanations: Explaining Training Data of Machine Learning Systems to Promote Transparency. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems . ACM, Yokohama Japan, 1-13. https://doi.org/10.1145/3411764.3445736 - sourcePartly supported

  Section 2.3 (p. 11) describes standardized documentation of training datasets, covering their creation, composition, uses, distribution, and maintenance. The Overview describes selecting that information for a data-centric explanation. This supports an example of systematically organizing information for transparency, but not the broader claim that practical approaches typically collect and curate data about various aspects of systems. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [28] [28] Mehtab Khan and Alex Hanna. 2022. The Subjects and Stages of Al Dataset Development: A Framework for Dataset Accountability. 19, 2 (2022), 171-256. - sourcePartly supported

  On pp. 248 and 252, the paper proposes documenting stages and design choices in AI dataset development to improve accountability. That supports systematic collection of information about datasets, but not the broader claim that practical approaches typically center on collecting and curating information about systems generally. (LLM, conf=0.79)

  Experimental — accuracy not yet measured
• cites [46] [46] Inioluwa Deborah Raji, Andrew Smart, Rebecca N. White, Margaret Mitchell, Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker Barnes. 2020. Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency . ACM, Barcelona Spain, 33-44. https://doi.org/10.1145/3351095.3372873 - sourceSupported
  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Copyright violations in images are easier to pinpoint and prosecute in the legal system.

“Another possible reason is that copyright violations in images are easier to pinpoint and prosecute in the legal system [52].”

• cites [52] [52] Jan Smits and Tijn Borghuis. 2022. Generative AI and Intellectual Property Rights. In Law and Artificial Intelligence: Regulating AI and Applying AI in Legal Practice , Bart Custers and Eduard Fosch-Villaronga (Eds.). T.M.C. Asser Press, The Hague, 323-344. https://doi.org/10.1007/978-94-6265-523-2\_17 - sourceCould not verify

  The passages from “Digging into the Accountability Gap” (§§19.3–19.5; page not provided) address liability for AI systems in healthcare, not copyright infringement. Further searches for image copyright and detection or enforcement of copyright infringement found no passage in our copy saying that violations involving images are easier to identify or prosecute. (LLM, conf=0.73)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Open source AI risks and opportunities are increasingly being studied in the generative AI landscape.

“Open source AI risks and opportunities are increasingly being studied in the generative AI landscape [16].”

• cites [16] [16] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, Fazl Barez, Genevieve Smith, Gianluca Guadagni, Jon Chun, Jordi Cabot, Joseph Marvin Imperial, Juan A. Nolazco-Flores, Lori Landay, Matthew Jackson, Paul Röttger, Philip H. S. Torr, Trevor Darrell, Yong Suk Lee, and Jakob Foerster. 2024. Near to Mid-term Risks and Opportunities of Open Source Generative AI. https://doi.org/10.48550/arXiv.2404.17047 arXiv:2404.17047 [cs] version: 1. - sourceSupported

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Important parts of present-day NLP and machine learning could be described as finding clever ways of gaming metrics, whether automated or human.

“indeed important parts of present-day NLP and machine learning could be described as finding clever ways of gaming metrics, whether automated or human [10].”

• cites [10] [10] Kenneth Ward Church and Valia Kordoni. 2022. Emerging Trends: SOTA-Chasing. Natural Language Engineering 28, 2 (March 2022), 249-269. https://doi.org/10. 1017/S1351324922000043 Publisher: Cambridge University Press. - sourcePartly supported

  Section 2 says NLP and machine-learning publications increasingly pursue better numbers with less insight, and §2.1 criticizes leaderboard-driven competition. Section 2.3 also warns that inappropriate metrics can produce misleading results. These passages support the metric-chasing characterization, but they do not establish that researchers find clever ways to game both automated and human metrics. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

True openness of this kind, and not just open weight models, can speed up innovation and afford inclusion and diversity.

“It is true openness of this kind, and not just open weight models, that can speed up innovation and afford inclusion and diversity [20, 61].”

• cites [20] [20] Dirk Groeneveld, Iz Beltagy, Pete Walsh, Akshita Bhagia, Rodney Kinney, Oyvind Tafjord, Ananya Harsh Jha, Hamish Ivison, Ian Magnusson, Yizhong Wang, Shane Arora, David Atkinson, Russell Authur, Khyathi Raghavi Chandu, Arman Cohan, Jennifer Dumas, Yanai Elazar, Yuling Gu, Jack Hessel, Tushar Khot, William Merrill, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew E. Peters, Valentina Pyatkin, Abhilasha Ravichander, Dustin Schwenk, Saurabh Shah, Will Smith, Emma Strubell, Nishant Subramani, Mitchell Wortsman, Pradeep Dasigi, Nathan Lambert, Kyle Richardson, Luke Zett sourceSupported
• cites [61] [61] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, Jonathan Tow, Alexander M. Rush, Stella Biderman, Albert Webson, Pawan Sasanka Ammanamanchi, Thomas Wang, Benoît Sagot, Niklas Muennighoff, Albert Villanova del Moral, Olatunji Ruwase, Rachel Bawden, Stas Bekman, Angelina McMillan-Major, Iz Beltagy, Huu Nguyen, Lucile Saulnier, Samson Tan, Pedro Ortiz Suarez, Victor Sanh, Hugo Laurençon, Yacine Jernite, Julien Launay, Margaret Mitchell, Colin Raffel, Aaron G sourcePartly supported

  Section 2.2 (pp. 6–7) describes an open international collaboration aimed at linguistic, geographical, and scientific diversity; Section 3.6 (p. 24) describes releasing documentation and code as well as the model. The Abstract (p. 3) says these releases aim to facilitate future research and applications. The paper does not establish that this broader openness speeds innovation relative to releasing model weights alone. (LLM, conf=0.79)

  Experimental — accuracy not yet measured

• AssumptionRests on an assumption: Openness is one of the key features that enables transparency and accountability.

May combine two claims.

Auditing of Stable Diffusion's underlying datasets has revealed legal challenges and deeply problematic content.

“This has enabled auditing of the underlying datasets, which has revealed legal challenges [49] and deeply problematic content [4, 5].”

• cites [49] [49] Pamela Samuelson. 2023. Generative AI meets copyright. Science 381, 6654 (July 2023), 158-161. https://doi.org/10.1126/science.adi0656 Publisher: American Association for the Advancement of Science. - sourceNot checked — no full text of the source
• cites [4] [4] Abeba Birhane, Vinay Prabhu, Sang Han, Vishnu Naresh Boddeti, and Alexandra Sasha Luccioni. 2023. Into the LAION's Den: Investigating Hate in Multimodal Datasets. In 37th Conference on Neural Information Processing Systems (NeurIPS 2023) . - sourceSupported
• cites [5] [5] Abeba Birhane and Vinay Uday Prabhu. 2021. Large image datasets: A pyrrhic win for computer vision?. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV) . 1536-1546. https://doi.org/10.1109/WACV48630.2021.00158 ISSN: 2642-9381. - sourceNot checked — no full text of the source

May combine two claims.

For present-day generative AI systems, openness is not an all-or-nothing feature; rather, it is best conceived as a composite and gradient notion.

“For present-day generative AI systems, it is not an all-or-nothing feature; rather, it is best conceived as a composite and gradient notion [53].”

• cites [53] [53] Irene Solaiman. 2023. The Gradient of Generative AI Release: Methods and Considerations. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT '23) . Association for Computing Machinery, New York, NY, USA, 111-122. https://doi.org/10.1145/3593013.3593981 - sourceSupported

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

This hand-waving appears mostly a thinly disguised attempt to obscure the clear and present harms such models already pose while minimising the considerable legal and regulatory exposure that would come with disclosing details about training data.

“but this appears mostly a thinly disguised attempt to obscure the clear and present harms such models already pose [17] while minimising the considerable legal and regulatory exposure that would come with disclosing details about training data [35, 49].”

• cites [17] [17] Timnit Gebru, Emily M. Bender, Angelina McMillan-Major, and Margaret Mitchell. 2023. Statement from the listed authors of Stochastic Parrots on the "AI pause" letter. https://www.dair-institute.org/blog/letter-statement-March2023 -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [35] [35] Nicola Lucchi. 2023. ChatGPT: A Case Study on Copyright Challenges for Generative Artificial Intelligence Systems. European Journal of Risk Regulation (Aug. 2023), 1-23. https://doi.org/10.1017/err.2023.59 - sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [49] [49] Pamela Samuelson. 2023. Generative AI meets copyright. Science 381, 6654 (July 2023), 158-161. https://doi.org/10.1126/science.adi0656 Publisher: American Association for the Advancement of Science. - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

The framework is flexible enough to serve as a blueprint for implementing living guidelines or for formulating templates such as those to be developed by the EU AI Office.

“This means that the framework is flexible enough to serve as a blueprint for the implementation of living guidelines [7] or for the formulation of templates such as those to be developed by the EU AI Office.”

• cites [7] [7] Claudi L. Bockting, Eva A. M. van Dis, Robert van Rooij, Willem Zuidema, and Johan Bollen. 2023. Living guidelines for generative AI - why scientists must oversee its use. Nature 622, 7984 (Oct. 2023), 693-696. https://doi.org/10.1038/d41586023-03266-1 Bandiera\_abtest: a Cg\_type: Comment Number: 7984 Publisher: Nature Publishing Group Subject\_term: Machine learning, Technology, Policy, Computer science. - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Another model openness framework converges on many of the same dimensions as our work and proposes to embed openness descriptions in model releases themselves.

“Another is a model openness framework that converges on many of the same dimensions as our work [59], and which proposes to embed openness descriptions in model releases themselves.”

• cites [59] [59] Matt White, Ibrahim Haddad, Cailean Osborne, Xiao-Yang, Liu, Ahmed Abdelmonsef, and Sachin Varghese. 2024. The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency and Usability in AI. https://doi.org/10.48550/arXiv.2403.13784 arXiv:2403.13784 [cs]. - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Work on AI safety and reproducibility has long pointed out the crucial importance of training data for understanding model performance, ensuring reproducibility, and assessing legal exposure.

“Work on AI safety and reproducibility has long pointed out the crucial importance of training data for understanding model performance, ensuring reproducibility, and assessing legal exposure [2, 25, 29, 46, 58].”

• cites [2] [2] Ariful Islam Anik and Andrea Bunt. 2021. Data-Centric Explanations: Explaining Training Data of Machine Learning Systems to Promote Transparency. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems . ACM, Yokohama Japan, 1-13. https://doi.org/10.1145/3411764.3445736 - sourcePartly supported

  The Introduction says training-dataset quality and characteristics are fundamental to system performance, and §2.3 (p. 11) says dataset documentation helps assess quality and fitness for a machine-learning system. The passages do not address ensuring reproducibility or assessing legal exposure; searches for those topics found no supporting passage in our copy. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites [25] [25] Ben Hutchinson, Andrew Smart, Alex Hanna, Emily Denton, Christina Greer, Oddur Kjartansson, Parker Barnes, and Margaret Mitchell. 2021. Towards Accountability for Machine Learning Datasets: Practices from Software Engineering and Infrastructure. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21) . Association for Computing Machinery, New York, NY, USA, 560-575. https://doi.org/10.1145/3442188.3445918 - sourcePartly supported

  The Abstract (page unavailable) says ML depends critically on datasets and that datasets frequently contribute to errors. “Data Reuse and Reproducibility” (page unavailable) says open dataset reporting lets others attempt to replicate results and predictions. The paper also discusses licensing constraints, but the passages found do not establish the claim that training data is crucial for assessing legal exposure; the Conclusions explicitly place liability beyond its scope. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [29] [29] Florian Königstorfer and Stefan Thalmann. 2022. AI Documentation: A path to accountability. Journal of Responsible Technology 11 (Oct. 2022), 100043. https: //doi.org/10.1016/j.jrt.2022.100043 - sourceSupported
• cites [46] [46] Inioluwa Deborah Raji, Andrew Smart, Rebecca N. White, Margaret Mitchell, Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker Barnes. 2020. Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency . ACM, Barcelona Spain, 33-44. https://doi.org/10.1145/3351095.3372873 - sourcePartly supported

  Section 4.4.2 (page unavailable) treats dataset documentation as important to auditing and understanding what affects model performance; section 4.6 also proposes changing training data to address performance across groups. The passages do not establish that training data is crucial for reproducibility or for assessing legal exposure. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [58] [58] Joel Walmsley. 2021. Artificial intelligence and the value of transparency. AI & SOCIETY 36, 2 (June 2021), 585-595. https://doi.org/10.1007/s00146-020-01066-z - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Data sheets document key aspects of data collection and curation and ensure that relevant information about training data is made available in systematic and relatively standardised ways.

“Data sheets document key aspects of data collection and curation [18, 38]. They ensure that relevant information about training data is made available in systematic and relatively standardised ways.”

• cites [18] [18] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for datasets. Commun. ACM 64, 12 (Nov. 2021), 86-92. https://doi.org/10.1145/3458723 - sourceSupported
• cites [38] [38] Angelina McMillan-Major, Emily M. Bender, and Batya Friedman. 2023. Data Statements: From Technical Concept to Community Practice. ACM Journal on Responsible Computing (2023). https://doi.org/10.1145/3594737 Just Accepted. - sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Model cards represent the field's standard for disclosing and documenting key facts about the architecture, training and evaluation of the model.

“Model cards represent the field's standard for disclosing and documenting key facts about the architecture, training and evaluation of the model [14, 40].”

• cites [14] [14] Anamaria Crisan, Margaret Drouhard, Jesse Vig, and Nazneen Rajani. 2022. Interactive Model Cards: A Human-Centered Approach to Model Documentation. In 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT '22) . Association for Computing Machinery, New York, NY, USA, 427-439. https: //doi.org/10.1145/3531146.3533108 - sourceSupported
• cites [40] [40] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model Cards for Model Reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* '19) . Association for Computing Machinery, New York, NY, USA, 220-229. https://doi.org/10.1145/3287560.3287596 - sourcePartly supported

  The Model Card outline (p. 3) calls for information about model type, training algorithms and data, and evaluation results, supporting the documentation part of the claim. But the paper proposes model cards rather than establishing them as the field's standard; the discussion (p. 9) says they are one transparency approach and are unlikely to be standardized in the near term. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S4 (based).

The past year has seen a steep rise in generative AI systems that claim to be open.

“The past year has seen a steep rise in generative AI systems that claim to be open.”

May combine two claims.

Companies operating in the generative AI space currently appear to be converging on a strategy known as open-washing: collect brownie points for openness without disclosing critical information of training and tuning procedures, thereby largely escaping the scientific scrutiny and legal exposure that would come with full openness.

“Companies operating in the generative AI space currently appear to be converging on a strategy known as open-washing [60]: collect brownie points for openness without disclosing critical information of training and tuning procedures, thereby largely escaping the scientific scrutiny and legal exposure that would come with full openness.”

• cites [60] [60] David Gray Widder, Sarah West, and Meredith Whittaker. 2023. Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. https://doi.org/10.2139/ssrn.4543807 - sourceNot checked — no full text of the source

May combine two claims.

A survey of 40 large language models and 6 text-to-image models reveals the most important trends in current open generative AI and shows how evidence-based openness assessment can work in practice.

“A survey of 40 large language models and 6 text-to-image models reveals the most important trends in current open generative AI and shows how evidence-based openness assessment can work in practice.”

• AssumptionRests on an assumption: The 46 generative AI systems surveyed are representative of the current open generative AI landscape.

May combine two claims.

BloomZ makes available source code for training, fine-tuning and running the model, while for Llama none of the model's source code is made available, only scripts for running the model are shared.

“When it comes to open code , we find that BloomZ makes available source code for training, fine-tuning and running the model, while for Llama none of the model's source code is made available, only scripts for running the model are shared.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If there are technical reports, they tend to focus on performance evaluation at the expense of documenting system architecture and training data.

“If there are technical reports, they tend to focus on performance evaluation at the expense of documenting system architecture and training data.”

May combine two claims.

BloomZ released model cards that describe the architecture and evaluation results with extensive cross-references to other documentation, while the Llama model card only provides minimum detail and none whatsoever on training data.

“BloomZ also released model cards that describe the architecture and evaluation results with extensive cross-references to other documentation on training data, training approach, model architecture, fine-tuning and responsible use. In contrast, the Llama model card only provides minimum detail and none whatsoever on training data.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

Peer-reviewed papers seem to have almost completely fallen out of fashion and are increasingly replaced by blogposts with cherry-picked examples or corporate preprints with minimal detail.

“Peer-reviewed papers seem to have almost completely fallen out of fashion and are increasingly replaced by blogposts with cherry-picked examples or corporate preprints with minimal detail.”

May combine two claims.

Both BloomZ and Llama licences aim to restrict harmful use cases, but there is a key difference in how they implement how model outputs are to be represented: RAIL stipulates that a user may not generate content without expressly and intelligibly disclaiming that the text is machine-generated, while Llama stipulates that a user may not represent that Llama 2 outputs are human-generated, a much a

“Both licences aim to restrict harmful use cases, but there is a key difference in how they implement how model outputs are to be represented. RAIL stipulates that a user may not 'generate content without expressly and intelligibly disclaiming that the text is machine-generated', while Llama stipulates that a user may not 'represent that Llama 2 outputs are human-generated' - a much lower bar, bec”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Neither BloomZ nor Llama distribute models as software packages via indexed and version controlled public code repositories such as pypi; both are primarily intended for local deployment.

“Neither BloomZ nor Llama distribute models as software packages via indexed and version controlled public code repositories such as pypi. Instead, both are primarily intended for local deployment.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

BloomZ is available through the Petals API, while for Llama an API is only available behind a privacy-defying signup form.

“BloomZ is available through the Petals API , while for Llama an API is only available behind a privacy-defying signup form.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

As a communication strategy, open-washing is highly effective.

“There is ample evidence that as a communication strategy, open-washing is highly effective.”

Probably not a substantial claim: may not pass S4 (based).

Many systems share little information about instruction tuning steps or metaprompting techniques.

“Many systems share little information about instruction tuning steps or metaprompting techniques.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

System, data and code documentation is often incomplete and lacks academic rigour.

“System, data and code documentation is often incomplete and lacks academic rigour.”

May combine two claims.

Both BloomZ and Llama make the LLM weights of the base model available, though for Llama access is restricted through a consent form.

“Both systems make the LLM weights of the base model available, though for Llama access is restricted through a consent form.”

May combine two claims.

Model weights for the instruction-tuned version (RL weights) are made openly available by BloomZ, while for Llama they require an access request.

“Model weights for the instruction-tuned version ( RL weights ) are made openly available by BloomZ, while for Llama they require an access request.”

May combine two claims.

All of the big commercial players -Meta, Google, Cohere, Microsoft and Mistral- are occupying the lower ranks of the openness leaderboard, as are many alternatives that build on them.

“It is noteworthy that all of the big commercial players -Meta, Google, Cohere, Microsoft and Mistral- are occupying the lower ranks, as are many alternatives that build on them.”

May combine two claims.

The architecture is described for BloomZ in multiple scientific papers and supported by a github repository of code and recipes on HuggingFace; for Llama, the architecture is described in less detail and scattered across corporate websites and a preprint.

“The architecture is described for BloomZ in multiple scientific papers and supported by a github repository of code and recipes on HuggingFace; for Llama, the architecture is described in less detail and scattered across corporate websites and a preprint.”

May combine two claims.

The BloomZ project code is well-documented and actively maintained, while for Llama 2 no documentation of source code is available as the source code itself is not open.

“The BloomZ project code is well-documented and actively maintained, while for Llama 2 no documentation of source code is available as the source code itself is not open.”

May combine two claims.

The training data for instruction tuning (RL data) is described and documented by Bloom as consisting of xP3, while for Llama the fine-tuning data remains undisclosed.

“The training data for instruction tuning ( RL data ) is described and documented by Bloom as consisting of xP3 (Crosslingual Public Pool of Prompts); for Llama, the corporate preprint notes that fine-tuning was done based on 'a large dataset of over 1 million binary comparisons based on humans applying our specified guidelines, which we refer to as Meta reward modeling data', and which remains und”

Probably not a substantial claim: may not pass S4 (based).

Most models in the bottom half do not provide any details about datasets beyond very generic descriptors obviously designed to evade legal scrutiny.

“Most models in the bottom half do not provide any details about datasets beyond very generic descriptors obviously designed to evade legal scrutiny.”

Probably not a substantial claim: may not pass S3 (load-bearing).

We did not identify systems other than text and text-to-image generators that bill themselves as open source.

“we did not identify systems other than text and text-to-image generators that bill themselves as open source.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Some of the other assessed text-to-image systems build on or fine-tune the various models of Stable Diffusion.

“Some of the other assessed systems build on or fine-tune the various models of Stable Diffusion.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

OpenAI's DALL-E is completely closed.

“Open AI's DALL-E is completely closed.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Only Stable Diffusion has been open to scrutiny from scientists, regulators and the general public.

“It also means that only Stable Diffusion has been open to scrutiny from scientists, regulators and the general public.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

If we focused on instruction-tuning, there would be only a handful of models that offer the requisite openness.

“if we were to focus on the all-important instruction-tuning, by many accounts the secret sauce of chat LLMs, there would be only a handful of models that offer the requisite openness.”

May combine two claims.

Text-to-image generators often are released under a Responsible AI Licence, restricting harmful uses and adding one more layer of safety and accountability.

“Whereas the latter often are released under a Responsible AI Licence, restricting harmful uses and adding one more layer of safety and accountability”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Meta and Mistral have collectively dragged the average level of openness down simply through the ubiquity of their model weights.

“Two influential companies in this domain, Meta and Mistral, have collectively dragged the average level of openness down simply through the ubiquity of their model weights.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Models can feature a simple training pipeline, but can also employ complex training procedures such as distributed or incremental training techniques, sometimes even using data acquired from user interactions.

“Models can feature a simple training pipeline, but can also employ complex training procedures such as distributed or incremental training techniques (e.g. federated learning, batch learning or online learning), sometimes even using data acquired from user interactions.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A first crop of text generators, including BloomZ and OpenAssistant, clearly aimed at meaningful degrees of openness.

“While a first crop of text generators -including BloomZ and OpenAssistant- clearly aimed at meaningful degrees of openness”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Text-to-image generators differ in terms of their machine learning architecture.

“Text-to-image generators also differ in terms of their machine learning architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

There are many open weight models currently.

“there are many open weight models currently”

May combine two claims.

Soon enough large corporate players started releasing systems billed as open source while they were in fact at best open weight, significantly diluting the term.

“soon enough large corporate players started releasing systems billed as open source while they were in fact at best open weight, significantly diluting the term [60].”

• cites [60] [60] David Gray Widder, Sarah West, and Meredith Whittaker. 2023. Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. https://doi.org/10.2139/ssrn.4543807 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Some text-to-image systems use watermarking to enable a form of provenance tracking.

“Some systems use watermarking to enable a form of provenance tracking.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Focusing on the open weight dimension makes it seem as if almost half of all text generators are open.

“focusing on this one dimension makes it seem as if almost half of all text generators are open.”

Probably not a substantial claim: may not pass S3 (load-bearing).

We did identify a range of recent hybrid systems that combine text and text-to-image generation in one system, such as DeepFloyd by Stability AI, Open AI's GPT-4 or Google's Gemini model family.

“We did identify a range of recent hybrid systems that combine text and text-to-image generation in one system, such as DeepFloyd by Stability AI, Open AI's GPT-4 or Google's Gemini model family.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some other assessed text-to-image systems are open-weight only.

“Some other systems are open-weight only.”

May combine two claims.

The most common licences for text generators are classic open source licences like Apache 2.0, presenting no restrictions on use of the technology.

“the most common licences for text generators are classic open source licences like Apache 2.0, presenting no restrictions on use of the technology.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Text-to-image systems commonly rely on forms of prompt moderation, often text filtering or classification, for moderation.

“For moderation, text-to-image systems commonly rely on forms of prompt moderation , often text filtering or classification.”

Probably not a substantial claim: may not pass S2 (checkable).

Our survey has offered a first glimpse at the detrimental effects of open-washing by companies looking to evade scientific, regulatory and legal scrutiny.

“Our survey has offered a first glimpse at the detrimental effects of open-washing by companies looking to evade scientific, regulatory and legal scrutiny.”

May combine two claims.

Openness in generative AI is gradient, coming in degrees, and cannot be reduced to a simple open/closed binary.

“we argue that openness in generative AI is necessarily composite (consisting of multiple elements) and gradient (coming in degrees)”

• Also stated in: § 1 INTRODUCTION · p. 2 · plain

  “The first is that openness is not a binary feature: today's transformer-based system architectures and their training procedures are complex, and they can only be classified into open or closed at the price of severe information loss.”

• Also stated in: § 2 OPENNESS ASSESSMENT IN ACTION: A COMPOSITE AND GRADIENT APPROACH · p. 5 · plain

  “Graded, because each element itself can be realised with different degrees of openness, and it is no longer feasible to maintain a simple open/closed binary.”

Classified as probably not a substantial claim.

Openness in generative AI is composite, consisting of multiple elements each of which can be assessed.

“we argue that openness in generative AI is necessarily composite (consisting of multiple elements) and gradient (coming in degrees)”

• Also stated in: § 2 OPENNESS ASSESSMENT IN ACTION: A COMPOSITE AND GRADIENT APPROACH · p. 5 · plain

  “Composite, because it is made up of multiple elements, each of which can be assessed.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

The practice of open-washing has the effect of compromising professional standards in software and technology development, moving the field away from core tenets of the open source movement like reverse-engineerability and full transparency.

“This practice of open-washing has the effect of compromising professional standards in software and technology development, moving the field away from core tenets of the open source movement like reverse-engineerability and full transparency.”

Classified as probably not a substantial claim.

The most important gap in the EU AI Act appears to be that there is very little reference to training datasets.

“The most important gap appears to be that there is very little reference to training datasets.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Opening up data and training pipelines might pose not only a commercial risk but also lead to legal exposure.

“And opening up data and training pipelines might pose not only a commercial risk but also lead to legal exposure.”

Probably not a substantial claim: may not pass S4 (based).

With the enshrinement of open source in the AI Act, the term open source comes to carry unprecedented legal weight.

“In sum, with the enshrinement of open source in the AI Act, the term 'open source' comes to carry unprecedented legal weight.”

Classified as probably not a substantial claim. May combine two claims.

Until recently, classifications of software as open source could simply rely on the availability of source code under appropriate licensing: if some software is released under a licence approved by the Open Source Initiative, it means that it is fully open and minimally restrictive.

“Until recently, classifications of software as open source could simply rely on the availability of source code under appropriate licensing: if some software is released under a licence approved by the Open Source Initiative (OSI), it means that it is fully open and minimally restrictive [45].”

• cites [45] [45] Bruce Perens. 1999. The Open Source Definition. In Open Sources: Voices from the Open Source Revolution , Chris DiBona, Sam Ockman, and Mark Stone (Eds.). O'Reilly, 171-188. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The EU AI Act will impose on general purpose AI providers a number of potentially burdensome requirements, including going through a conformity assessment, providing human oversight, and providing technical documentation that includes detailed information on system architecture, training datasets, provenance and curation.

“The EU AI Act [44] will impose on general purpose AI providers a number of potentially burdensome requirements, including going through a conformity assessment, providing human oversight, and providing technical documentation that includes detailed information on system architecture, training datasets, provenance and curation (Annex IV in [44]).”

• cites [44] [44] European Parliament. 2024. Artificial Intelligence Act: Provisional Agreement Resulting from Interinstitutional Negotiations. , 245 pages. http://web.archive.org/web/20240310112041/https://www.europarl. europa.eu/meetdocs/2014\_2019/plmrep/COMMITTEES/CJ40/AG/2024/0213/1296003EN.pdf -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A key sign of open-washing is the growth of a release by blogpost strategy.

“A key sign of open-washing is the growth of what we call a 'release by blogpost'-strategy.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

These evaluation tables offer ample degrees of freedom for cherry-picking, enabling model providers to present their products in the best light possible.

“These tables also offer ample degrees of freedom for cherry-picking, enabling model providers to present their products in the best light possible.”

Probably not a substantial claim: may not pass S2 (checkable).

As current and impending regulation increasingly puts legal weight on the notion of open source generative AI, the spectre of false assurance looms large.

“As current and impending regulation increasingly puts legal weight on the notion of open source generative AI, the spectre of false assurance looms large.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Open-washing is bad for the public understanding of AI, because it creates artefacts designed to impress without providing people with the resources to reach a deeper understanding of the technology.

“And open-washing is bad for the public understanding of AI, because it creates artefacts designed to impress without providing people with the resources to reach a deeper understanding of the technology [39].”

• cites [39] [39] Lisa Messeri and M. J. Crockett. 2024. Artificial intelligence and illusions of understanding in scientific research. Nature 627, 8002 (March 2024), 49-58. https://doi.org/10.1038/s41586-024-07146-0 Publisher: Nature Publishing Group. - sourceNot checked — no full text of the source

May combine two claims.

Open-washing stands in the way of innovation, because if large corporations can derive benefits from the trappings of open source without doing the requisite work, this sucks up the oxygen in the open source ecosystem, making it less attractive for smaller entities to find funding for truly open projects.

“Open-washing stands in the way of innovation, because if large corporations can derive benefits from the trappings of open source without doing the requisite work, this sucks up the oxygen in the open source ecosystem, making it less attractive for smaller entities to find funding for truly open projects [1].”

• cites [1] [1] Nur Ahmed, Muntasir Wahed, and Neil C. Thompson. 2023. The growing influence of industry in AI research. Science 379, 6635 (March 2023), 884-886. https://doi.org/10.1126/science.ade2420 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The auditing procedure is itself fully open to public scrutiny and to community contribution.

“The auditing procedure is itself fully open to public scrutiny and to community contribution (see Appendix).”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The matrix of relevant dimensions of openness in generative AI provides a sufficient level of detail to provide well-informed, high quality, systematic judgements of openness in generative AI.

“The dozen or so dimensions we have identified here along with three levels of openness ( open, partial, closed) provide a sufficient level of detail to provide well-informed, high quality, systematic judgements of openness in generative AI.”

Classified as probably not a substantial claim.

The framework for model openness and transparency used here has been tried and tested since July 2023 in an openness leaderboard that tracks degrees of openness for a growing number of generative AI models.

“Here we use and extend a framework for model openness and transparency that has been tried and tested since July 2023 in an openness leaderboard that tracks degrees of openness for a growing number of generative AI models.”

May combine two claims.

Only BloomZ can substantially claim open source status, while Meta's Llama is at best open weights, and is closed in almost all other aspects.

“Only BloomZ can substantially claim open source status, while Meta's Llama is at best open weights, and is closed in almost all other aspects.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Every single openness judgement in the survey is directly linked to publicly available evidence, and all data points are available in a versioned data repository.

“Every single openness judgement is directly linked to publicly available evidence, and all data points are available in a versioned data repository.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Open-washing is related to the notion of audit-washing: the risks posed by poorly designed auditing procedures.

“Open-washing is related to the notion of audit-washing [19]: the risks posed by poorly designed auditing procedures.”

• cites [19] [19] Ellen P. Goodman and Julia Tréhu. 2022. AI Audit-Washing and Accountability . Technical Report. German Marshall Fund of the United States. https://www. jstor.org/stable/resrep44893 - sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

These hidden champions of the generative AI world are where progress towards more open systems is most likely.

“In fact, these hidden champions of the generative AI world are where progress towards more open systems is most likely.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

For evidence-based openness assessment, the most relevant aspects of text-to-image generators are how they implement provenance tracking for synthetic imagery and guardrails against undesired content.

“Most relevant for evidence-based openness assessment are the ways in which text-to-image generators implement ways of tracking provenance of synthetic imagery and set up guardrails against creating undesired content.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Some data requires closed-door assessment; in some instances assessment should take place behind closed doors.

“Some data requires closed-door assessment.”

• Also stated in: § 3.2 Limitations · p. 10 · plain

  “in some instances assessment should take place behind closed doors”

Probably not a substantial claim: may not pass S3 (load-bearing).

For image generators, open datasets like LAION have enabled serious auditing.

“Perhaps this is because for image generators, open datasets like LAION have enabled serious auditing.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Well-funded corporate heavyweights are taking oxygen out of the room for smaller organisations that operate with higher professional and ethical standards.

“well-funded corporate heavyweights are taking oxygen out of the room for smaller organisations that operate with higher professional and ethical standards.”

Classified as probably not a substantial claim.

The overall framework is designed so that the bulk of its features are applicable to any generative AI system.

“Importantly, the overall framework is designed so that the bulk of the features are applicable to any generative AI system.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

The place of a system on the openness gradient is determined by the relative openness of its constituent elements.

“The place of a system on this gradient is determined by the relative openness of its constituent elements.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Assigning weights to openness classes and deriving a cumulative openness score yields a gradient measure of openness.

“One is to assign weights to openness classes and to derive, based on this, a cumulative openness score (Figure 4, panel 2). This would be a gradient measure of openness.”

May combine two claims.

Makers of text-to-image models have been more aware of harmful uses and legal exposure that may result from use of their models, while makers of text generators seem less concerned about harmful applications.

“It seems makers of text-to-image models have been more aware of harmful uses and legal exposure that may result from use of their models, while makers of text generators seem less concerned about harmful applications.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

Many corporate players propose that the diluted sense of openness is the only attainable.

“Many corporate players, moving aside the first of these as unrealistic, propose that therefore the second, diluted sense is the only attainable.”

Probably not a substantial claim: may not pass S2 (checkable).

A composite notion of openness can provide the building blocks that can cater to specific use cases.

“A composite notion of openness can also provide the building blocks that can cater to specific use cases - be it a radically open system for use in research and education, or more privacy-oriented ones with privacy-enhancing techniques in place.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Full openness can be harmful.

“Full openness can be harmful”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Such challenges make it hard to assess training data openness well.

“In sum, such challenges make it hard to assess training data openness well: we are only scraping the surface here.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

When properly implemented, a framework that conceptualises AI openness in a gradient and composite way can safeguard against open-washing by enabling public scrutiny.

“When properly implemented, such a framework can safeguard against open-washing by enabling public scrutiny.”

• AssumptionRests on an assumption: Openness is one of the key features that enables transparency and accountability.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Our survey stays relatively superficial when it comes to assessing exactly how open the training data of a system is.

“Our survey stays relatively superficial when it comes to assessing exactly how open the training data of a system is.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

A possible reason for the smaller number of text-to-image generators might be that relatively few image datasets are available.

“A possible reason for this might be that relatively few image datasets are available.”

• AssumptionRests on an assumption: The 46 generative AI systems surveyed are representative of the current open generative AI landscape.

Probably not a substantial claim: may not pass S2 (checkable).

If we are satisfied with calling such systems open, we are discarding many relevant dimensions and degrees of openness just to arrive at a single binary classification.

“If we are satisfied with calling such systems open, we are discarding many relevant dimensions and degrees of openness just to arrive at a single binary classification.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Singling out only a single measure and basing an openness classification on that is a fourth method, increasingly popular but even more reductive.

“Separate from these approaches is a fourth method, increasingly popular but even more reductive (Figure 4, panel 5). This is to single out only a single measure and base an openness classification on that.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

What some systems lack in performance, they make up for in openness and transparency.

“What some systems lack in performance, they make up for in openness and transparency - and this should be rewarded.”

Probably not a substantial claim: may not pass S4 (based).

Discourse about appropriate levels of openness has been prone to equate 'open source' with two rhetorical extremes: radical openness and homeopathic openness.

“Discourse about appropriate levels of openness has been prone to equate 'open source' with two rhetorical extremes: (i) radical openness, which would mean literally sharing every single model component and training dataset, and (ii) homeopathic openness, which is openness diluted beyond recognition, for instance by sharing only model weights.”

Probably not a substantial claim: may not pass S4 (based).

Open systems that disclose training and fine-tuning data are safer because of the possibility of public scrutiny and professional auditing.

“There is a case to be made that open systems that disclose training and fine-tuning data are safer because of the possibility of public scrutiny and professional auditing.”

• AssumptionRests on an assumption: Openness is one of the key features that enables transparency and accountability.

Classified as probably not a substantial claim.

There are multiple ways to turn fine-grained openness assessments into metrics of openness that support classification or comparison.

“There are multiple ways to turn fine-grained openness assessments into metrics of openness that support classification or comparison (Figure 4).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The framework is in principle applicable to the full range of generative AI.

“While the framework is in principle applicable to the full range of generative AI, we did not identify systems other than text and text-to-image generators that bill themselves as open source.”

• AssumptionRests on an assumption: The field of generative AI is fast growing.

Probably not a substantial claim: may not pass S3 (load-bearing).

Audits may have made makers of such models more aware of the harmful uses to which their systems can be put.

“and may have made makers of such models more aware of the harmful uses to which their systems can be put.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Using a composite and gradient approach to openness makes it possible to cut through competing stakes and actors trying to influence the definition of open source AI and to arrive at meaningful, evidence-based, multidimensional openness judgements.

“The way to subvert this risk is to use a composite and gradient approach to openness. That makes it possible to cut through the knot of competing stakes and actors trying to influence the definition of open source AI, and to arrive at meaningful, evidence-based, multidimensional openness judgements.”

• AssumptionRests on an assumption: A licence and its definition forms a single pressure point that will be targeted by corporate lobbies and big companies.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Many constituent features of the framework, from datasets to scientific documentation, are of general relevance to defining open in the context of AI and machine learning.

“More specifically, many of the constituent features, from datasets to scientific documentation, are of general relevance to the question of how to define open in the context of AI and machine learning.”

Probably not a substantial claim: may not pass S2 (checkable).

Small but open models can be just as useful for many end users.

“Sidestepping the toxicity of performance benchmarks and a bigger-is-always-better logic, these small but open models can be just as useful for many end users.”

Probably not a substantial claim: may not pass S2 (checkable).

Openness comes in degrees.

“Openness comes in degrees, and regulation should be designed to foster meaningful forms of openness.”

Probably not a substantial claim: may not pass S2 (checkable).

The claim that the diluted sense of openness is the only attainable is not the case.

“But this is not the case.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The 'open weight branded as open source' approach of companies like Meta and Mistral and lobbying attempts to favour simple licence-based decisions are likely partly due to corporate interests arguing for the easiest-to-attain openness dimensions to weigh most heavily.

“indeed this is likely one of the reasons behind the 'open weight branded as-open source' approach of companies like Meta and Mistral, and behind lobbying attempts to favour simple licence-based decisions.”

Probably not a substantial claim: may not pass S4 (based).

The EU AI Act is at risk of tying itself to a moving target: a licence-based definition of 'open source AI' that itself is evolving.

“The EU AI Act is at risk of tying itself to a moving target: a licencebased definition of 'open source AI' that itself is evolving.”

• AssumptionRests on an assumption: The licence-based definition of 'open source AI' is evolving.

Probably not a substantial claim: may not pass S3 (load-bearing).

Each contribution is reviewed by a domain expert before it is published as part of the assessment outcome of the system under scrutiny.

“Each contribution is then reviewed by a domain expert before it is published as part of the assessment outcome of the system under scrutiny.”

Probably not a substantial claim: may not pass S3 (load-bearing).

No third-party information was used in the assessment.

“No third-party information was used.”

Probably not a substantial claim: may not pass S2 (checkable).

Full disclosure is where a key to meaningful openness lies.

“Despite the challenges of openly sharing all data, we think full disclosure is where a key to meaningful openness lies.”

• AssumptionRests on an assumption: Openness is important for risk analysis, auditability, scientific reproducibility, and legal liability.

Probably not a substantial claim: may not pass S2 (checkable).

Knowing what is open and how open it is can help everyone make better decisions.

“However, open is better than closed in most cases, and knowing what is open and how open it is can help everyone make better decisions.”

• AssumptionRests on an assumption: Open is better than closed in most cases.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The assessment procedure is a two-step, evidence-based procedure with contributor and reviewer roles using an open data repository.

“Using the definitions for each dimension of openness in section 2, we set up a two-step, evidence-based assessment procedure with contributor and reviewer roles using an open data repository.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Currently, two variations of the framework exist that are tailored to the specific needs to assess text and text-to-image generators respectively.

“Currently, two variations of the framework exist that are tailored to the specific needs to assess text and text-to-image generators respectively.”

Probably not a substantial claim: may not pass S2 (checkable).

Full openness is not always the solution.

“Full openness is not always the solution: after all, even fully open systems can do harm and may be legally questionable.”

Probably not a substantial claim: may not pass S2 (checkable).

Such openness judgements can be used for individual models by potential users to make informed decisions for or against deployment of a particular architecture or model.

“Such judgements can be used for individual models by potential users to make informed decisions for or against deployment of a particular architecture or model.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Even fully open systems can do harm and may be legally questionable.

“after all, even fully open systems can do harm and may be legally questionable.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Such openness judgements can be used cumulatively to derive overall openness scores, and more reductively to classify systems into shades of openness or to define an openness cutoff for regulatory purposes.

“They can also be used cumulatively to derive overall openness scores, and more reductively to classify systems into shades of openness or to define an openness cutoff for regulatory purposes.”