Perspectives for Open Source AI

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

The OSAID's approach to training data openness has been highly criticized by the community.

“Such approach to the issue of training data openness has been highly criticized by the community.”

May combine two claims.

The term "open source" is being used widely in the context of AI, both in legal texts such as the EU AI Act and by private companies, with varying levels of transparency and openness.

“the term “open source” is being used widely in the context of AI, both in legal texts, such as the EU AI Act, and by private companies, that use the term “open source” for AI systems with varying levels of transparency and openness.”

• cites [53]Not matched

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

Classified as probably not a substantial claim.

The EU AI Act does not provide a precise definition of AI systems released under free and open source licenses.

“however it does not provide a precise definition of such systems”

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

By the time of writing, the AI Foundation Model Transparency Act has not progressed after being introduced in the House of Representatives and referred to the House Committee on Energy and Commerce.

“However, by the time of writing, the bill has not progressed any further after it had been introduced in the House of Representatives and referred to the House Committee on Energy and Commerce.102”

• cites [102]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). May combine two claims.

The AI Foundation Model Transparency Act does not specify what is understood by "alternative provisions" and does not provide a definition for open source foundation models.

“The bill does not specify what is understood by “alternative provisions”, nor does it provide a definition for open source foundation models.”

• cites [101]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). May combine two claims.

As seen from the draft, many concepts and details in the cited provisions still remain unclear, such as the definitions "free and open-source", "certain code modules", or to what extent the liability may be reduced and how such liability reduction depends on the level of compliance with national standards.

“As seen from the draft, many concepts and details in the cited provisions still remain unclear (which may be due both to translation inaccuracies and to the fact that the above text is just an interim draft), such as the definitions “free and open-source”, “certain code modules”, or to what extent the liability may be reduced and how such liability reduction depends on the level of compliance with”

• cites [129]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).

Unlike earlier versions of SB 1047, the TFAIA simply mentions the open source community as one of the stakeholders whose input should be taken into account by the Department of Technology when making recommendations on updating definitions used in the TFAIA.

“Unlike earlier versions of SB 1047, the TFAIA simply mentions open source community as one of the stakeholders whose input should be taken into account by the Department of Technology when making recommen-”

• cites [112]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). May combine two claims.

Although the provision leaves copyright holders the option to prohibit free use of copyrighted works for open source AI training, the wording "developers of open-source foundation models are encouraged to provide appropriate compensation to copyright holders" leaves broad room for interpretation as to whether such compensation is mandatory and as to the amount of compensation.

“Although this provision leaves the copyright holders the option to prohibit free use of copyrighted works for open source AI training, the wording “developers of open-source foundation models are encouraged to provide appropriate compensation to copyright holders” leaves broad room for interpretation as to whether such compensation is mandatory and as to the amount of compensation.”

• cites [129]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).

The draft (version 1.0) of the Chinese model law contained very limited provisions on open source AI, generally proclaiming support and encouragement of open source development platforms and open source communities (Article 16 of Chapter 2).

“The draft contained very limited provisions on open source AI, just generally proclaiming support and encouragement of open source development platforms and open source communities (Article 16 “Construction of Compute Infrastructure” of Chapter 2 “Support and Promotion of AI”).”

May combine two claims.

In September 2024, a number of safety and transparency bills concerning AI technologies were signed by the California Governor, including AB-2013 on generative AI training data transparency.

“In September 2024, a number of safety and transparency bills concerning AI technologies were signed by the Governor 108 , including AB-2013 109 on generative AI training data transparency.”

• cites [108]Not matched

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

• cites [109]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).

Article 83 of the model law provides for exemption from liability for open source AI.

“Furthermore, Article 83 provides for exemption from liability for open source AI”

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

According to Executive Order 'Preventing Woke AI in the Federal Government', new guidance for governmental departments and wholly owned government corporations should permit vendors to avoid requiring disclosure of specific model weights or other sensitive technical data where practicable (Section 4).

“according to Executive Order “Preventing Woke AI in the Federal Government” 105 , a new guidance to be prepared for governmental departments and wholly owned government corporations should permit vendors to “avoid requiring disclosure of specific model weights or other sensitive technical data where practicable” (Section 4).”

• cites [105]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). May combine two claims.

Version 2.0 of the model law already included specific provisions on open source AI that remained unchanged in version 3.0, except for the article numbering, although the definition of open source AI was introduced only in version 3.0.

“Although the definition of open source AI was introduced only in version 3.0, already version 2.0 of the model law included specific provisions on open source AI that remained unchanged in version 3.0, except for the article numbering.”

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

The EU AI Act uses hyphenation in the term 'open-source'.

“The EU AI Act uses hyphenation in the term “open-source”, as, for example, is stated in Merriam-Webster Dictionary”

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

Article 18 of the model law highlights strong encouragement of open source AI projects by the state.

“In particular, Article 18 highlights strong encouragement of open source AI projects by the state”

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

Data privacy laws are another legal challenge for opening and sharing of AI training datasets.

“Data privacy laws are another legal challenge for opening and sharing of AI training datasets.”

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

Companies like Amazon, Meta, and Google have constructed massive facilities dedicated to housing IT infrastructure for data processing, management, and storage.

“For instance, companies like Amazon, Meta, and Google have constructed massive facilities dedicated to housing IT infrastructure for data processing, management, and storage.78”

• cites [78]Not matched

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

May combine two claims.

Hardware plays an important role in AI training and can impair the four freedoms (use, study, modify, share) of open source AI.

“which plays an important role in AI training and which can impair the four freedoms (use, study, modify, share) of open source AI.”

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

The AI system objectives element of the AI Act definition fails to fulfil any delimiting function.

“Another element of the definition that apparently fails to fulfil any delimiting function is the AI system objectives.”

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

The EU abandoned its definition of AI system and adopted the one elaborated by the OECD, suggesting some political context influencing this decision.

“the fact that the EU abandoned its definition of AI system and adopted the one elaborated by the OECD suggests some political context influencing this decision.”

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

The OSAID v. 1.0 was released on 28 October 2024.

“The OSAID v. 1.0 was released on 28 October 2024.”

• cites [56]Not matched

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

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

Based on Recital 102, GPAI models released under free and open source licenses do not per se ensure high levels of transparency and openness; they should be considered to ensure them provided that certain information is made publicly available: parameters (including weights), information on model architecture, and information on model usage.

“Based on the wording of the first sentence cited above, GPAI models released under free and open source licenses do not per se ensure high levels of transparency and openness.They should be considered to ensure high levels of transparency and openness provided that certain information about them is made publicly available, namely: parameters (including weights), information on model architecture,”

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

Limiting open source AI systems only to systems trained on publicly available data would significantly and unreasonably limit the amount of available datasets and the potential of open source AI, excluding it from areas such as healthcare or any other area dealing with private or personal data.

“At the same time, limiting open source AI systems only to systems trained on publicly available data would significantly and unreasonably limit the amount of available datasets and the potential of open source AI, excluding it from such areas as healthcare or any other area dealing with private or personal data.67”

• cites [67]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).

The reasons for the chosen terminology "general purpose AI models" remain unclear.

“However, the reasons for the chosen terminology remain unclear.”

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

From a legal standpoint, data information has been one of the most controversial elements of the OSAID.

“From a legal standpoint, data information has been one of the most controversial elements.”

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

The OSI has set a goal to overcome the data privacy challenge in order not to exclude open source AI systems from data-sensitive areas.

“However, the OSI has set a goal to overcome this challenge in order not to exclude open source AI systems from data-sensitive areas.”

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

Another discrepancy between the professional community's understanding of open source and the AI Act's term 'AI models released under free and open source license' is the monetization of AI components.

“Another discrepancy between the understanding of open source by the professional community and the term of AI models released under free and open source license under the AI Act is the monetization of AI components.”

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

The final version of the EU AI Act includes controversial and arguable provisions that undermine the concept of open source and the future of open source AI systems.

“the final version of the Act includes a number of controversial and arguable provisions that undermine both the concept of open source and the future of open source AI systems”

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

Recital 102 states that software and data released under a free and open-source licence can contribute to research and innovation and provide significant growth opportunities for the Union economy.

“Software and data, including models, released under a free and open-source licence that allows them to be openly shared and where users can freely access, use, modify and redistribute them or modified versions thereof, can contribute to research and innovation in the market and can provide significant growth opportunities for the Union economy.”

• cites Recital 102Not 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).

In 2018, the contribution of open source to the European economy was estimated between €65 and €95 billion.

“in 2018, the contribution of open source to the European economy was estimated between €65 and €95 billion”

• cites [43]Not matched

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

May combine two claims.

Similar to the FSF, the OSI does not restrict selling or using open source software for commercial purposes.

“Similar to the FSF, the OSI does not restrict selling or using open source software for commercial purposes.80”

• cites [80]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). May combine two claims.

The final part of the AI system definition, "that can influence physical or virtual environments", appears to be more illustrative and does not provide any specific guidance on differentiating AI systems from non-AI systems.

“the final part of the definition, “...that can influence physical or virtual environments” appears to be more illustrative and does not provide any specific guidance on differentiating AI systems from non-AI systems.”

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

In the case of software, anyone who has a computer can use, study, modify, and share the source code; in the case of AI systems, data centers requiring massive memory storage, energy consumption, and computing power are necessary to train and run an AI system.

“Indeed, in case of software, anyone who has a computer can use, study, modify, and share the source code.In case of AI systems, data centers requiring massive memory storage, energy consumption, and computing power are necessary to train and run an AI system.”

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

A number of legal challenges arise in the context of the openness of training datasets, mainly related to differences in intellectual property and privacy laws in different legal systems.

“a number of legal challenges arise in the context of the openness of training datasets, mainly related to the differences in intellectual property and privacy laws in different legal systems.”

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

In the provisional agreement reached by the Parliament and the Council in December 2023, the term "foundation models" was replaced with "general purpose AI models".

“In the provisional agreement reached by the Parliament and the Council in December 2023, the term “foundation models” was replaced with “general purpose AI models”.”

• cites [35]Not matched

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

Classified as probably not a substantial claim.

The lack of a precise definition of open source AI systems in the EU AI Act incentivizes the creation of "openish" AI systems.

“thus incentivizing the creation of “openish” AI systems”

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

With respect to text and data mining for commercial use, Member States shall provide either an exception or limitation to the reproduction right of copyright holders (Article 4).

“With respect to text and data mining for commercial use, Member States shall provide either an exception or limitation to the reproduction right of the copyright holders (Article 4).”

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

It remains unclear why open source AI models should be excluded from the privileged treatment under the AI Act only due to their use for commercial purposes.

“Thus, although the AI Act does not explicitly exclude monetized AI models from the definition of AI models released under free and open source license, it remains unclear why open source AI models should be excluded from the privileged treatment under the AI Act only due to their use for commercial purposes.”

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

Some experts say that tying regulations to a fixed FLOP count or equivalent cannot be applied by regulators today because as algorithmic efficiency improves, computing costs decline, so an established threshold would quickly expand to cover many more models than just the largest.

“some experts say that tying regulations to some fixed FLOP count or equivalent cannot be applied by regulators today, because, “as algorithmic efficiency improves, computing costs decline”, so established threshold would quickly expand to cover many more models than just the largest.”

• cites [41]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).

Defining open source AI has been unanimously recognized as a challenging task.

“Defining open source AI has been unanimously recognized as a challenging task.”

• cites [52]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 S2 (checkable).

The unclarity of the terms "sufficiently detailed information", "a skilled person", and "a substantially equivalent system" might lead to minimal disclosure of training datasets, undermining the whole concept of open source.

“The unclarity of the used terms “sufficiently detailed information”, “a skilled person” 59 , “a substantially equivalent system” might lead to minimal disclosure of training datasets, undermining the whole concept of open source.”

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

A cost-benefit ratio of open source is above 1:4.

“a cost-benefit ratio of open source is of above 1:4”

• cites [43]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).

Some characteristics of AI systems as worded in the AI Act definition still remain unclear.

“some of the characteristics of AI systems as worded in the definition still remain unclear.”

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

Firms would need to spend 3.5 times more on software than they currently do if OSS did not exist.

“firms would need to spend 3.5 times more on software than they currently do if OSS did not exist.”

• cites [44]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). May combine two claims.

Open sourced AI models are more safe and secure compared to closed models, because the former may be trained in a publicly verifiable way, on publicly available data, allowing others to study the model's behavior and minimize the risk of vulnerabilities.

“open sourced AI models result more safe and secure as compared to closed models, since the former may be trained in a publicly verifiable way, on publicly available data, allowing others to study the model's behavior and minimize the risk of vulnerabilities.”

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

Under the OSAID, open source AI is an AI system made available under terms and in a way that grant freedoms to use, study, modify, and share the system.

“OSAI is an AI system made available under terms and in a way that grant the following freedoms: • use the system for any purpose and without having to ask for permission, • study how the system works and inspect its components, • modify the system for any purpose, including to change its output, • share the system for others to use with or without modifications, for any purpose.”

• cites OSAID v. 1.0Not 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). May combine two claims.

Training a model with that amount of computing power costs between US$50 million and $100 million, and this threshold should capture models such as OpenAI's GPT-4 and Meta's LLaMA.

“training a model with that amount of computing power costs between US$50 million and $100 million, and this threshold should capture models such as OpenAI's GPT-4 and Meta's LLaMA.”

• cites [30]Not matched

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

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

The definition of open source was initially elaborated for software, just referring to the source code, and is hardly applicable to AI, which covers many more components and processes.

“Since the definition of open source was initially elaborated for software, just referring to the source code, it is hardly applicable to AI, which covers many more components and processes.”

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

GPAI models with systemic risk are those with cumulative computation used for training greater than 10^25 FLOPs.

“GPAI models with systemic risk, which have high impact capabilities, meaning that the cumulative amount of computation used for their training measured in floating point operations (FLOPs) is greater than 10 25 (Article 51).”

• cites Article 51Not 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). May combine two claims.

The OSI describes four types of training data-open training data, public training data, obtainable training data, and unshareable non-public training data-each with its own legal frameworks allowing different freedoms of distribution.

“The OSI comments that the discussions held within the open source community have helped describe four types of data, each with its own legal frameworks allowing different freedoms of distribution:”

• cites [77]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). May combine two claims.

The "general-purpose AI model" is defined via a term "AI model" that is not defined in the AI Act, as the latter defines only "AI system".

“the “general-purpose AI model” is defined via a term “AI model” that is not defined in the AI Act, as the latter defines only “AI system”.”

• cites Article 3(63)Not matched

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

May combine two claims.

The list of exceptions from the overall exemption for AI systems released under free and open source licenses can be summarized as: high-risk AI systems, AI systems under Article 5 (prohibited practices), AI systems under Article 50 (transparency obligations), and GPAI models with systemic risk (Articles 53(2) and 54(6)).

“Thus, the list of exceptions from the above overall exemption for AI systems released under free and open source licenses can be summarized as follows: • high-risk AI systems, • AI systems that fall under Article 5 (prohibited AI practices), • AI systems that fall under Article 50 (transparency obligations for providers and deployers of certain AI systems), • GPAI models with systemic risk (”

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

Versions of SB 1047 preceding 19 August 2024 specifically addressed open source AI models, defining them as an artificial intelligence model that is made freely available and that may be freely modified and redistributed.

“Interestingly, the versions of SB 1047 preceding 19 August 2024 specifically addressed open source AI models, defining them as follows: (n) “Open-source artificial intelligence model” means an artificial intelligence model that is made freely available and that may be freely modified and redistributed.”

• cites [114]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). May combine two claims.

The FTC uses the term "open-weights models" and states that, despite nascent terminology, they have the potential to be a positive force for innovation and competition, just like open source software.

“The Federal Trade Commission (FTC) admits that there is an active discussion on what “open” and “open source” should mean in the context of AI, but uses the term “open-weights models”, which, “despite the nascent terminology, ... have the potential to be a positive force for innovation and competition”, just like open source software.99”

• cites [99]Not matched

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

May combine two claims.

A common feature of the EU and U.S. approaches to open source AI is that the training dataset disclosure requirement is applicable to open source AI models, but contrary to the OSAID by the OSI, does not constitute a prerequisite for the AI model to be considered open source.

“concerns open source AI, a common feature of the EU and U.S. approaches is that the training dataset disclosure requirement is applicable to open source AI models, but, contrary to the understanding of the OSAID by the OSI, does not constitute a prerequisite for the AI model to be considered open source.”

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

The NTIA Report concludes that current evidence is not sufficient to definitively determine either that restrictions on open-weight models are warranted, or that restrictions will never be appropriate in the future.

“the Report concludes that “current evidence is not sufficient to definitively determine either that restrictions on such openweight models are warranted, or that restrictions will never be appropriate in the future.””

• cites [98]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).

The AI Foundation Model Transparency Act would direct companies to provide consumers and the FTC with information on the model's training data, model training mechanisms, and whether user data is collected in inference.

““direct companies to provide consumers and the FTC with information on the model's training data, model training mechanisms, and whether user data is collected in inference””

• cites [101]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). May combine two claims.

In his 2024 veto message, Governor Newsom commented that by focusing only on the most expensive and large-scale models, SB-1047 establishes a regulatory framework that could give the public a false sense of security about controlling this fast-moving technology, whereas smaller specialized models may emerge as equally or even more dangerous than the models targeted by SB 1047.

“In 2024, in his veto message, the Governor commented that “by focusing only on the most expensive and large-scale models, SB-1047 establishes a regulatory framework that could give the public a false sense of security about controlling this fast-moving technology”, whereas smaller, “specialized models may emerge as equally or even more dangerous than the models targeted by SB 1047.”116”

• cites [116]Not matched

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

May combine two claims.

President Biden's 2023 Executive Order 14110 on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence was the key legal document regulating AI at the federal level before being revoked by President Trump on his first day in office in January 2025.

“President Biden's 2023 Executive Order 14110 on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence 96 was the key legal document regulating AI at the federal level before being revoked by President Trump on his first day in office in January 2025.97”

• cites [96]Not matched

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

• cites [97]Not matched

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

May combine two claims.

Under Recital 104, GPAI models released under a free and open source license have to comply with producing a summary about the content used for model training and putting in place a policy to comply with Union copyright law, including reservation of rights for text and data mining under Article 4 of the Digital Single Market Directive.

“Additionally, under the cited above Recital 104, GPAI models that are released under a free and open source license have to comply with the following transparency and legal requirements: • produce a summary about the content used for model training, • put in place a policy to comply with Union copyright law, including with the reservation of rights by the rightholders in case of text and data”

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

Annex XII lists information that should be disclosed to providers of AI systems intending to integrate the GPAI model into their AI systems, including a general description and a description of model elements and development process.

“As regards the second obligation set forth in Article 53(1)(b), Annex XII lists the following information that should be disclosed to providers of AI systems intending to integrate the GPAI model into their AI systems:”

May combine two claims.

Versions of SB 1047 preceding 19 August 2024 mentioned open source AI in the context of the Board of Frontier Models and the Frontier Model Division, including an advisory committee for open source AI with specified powers.

“The versions of SB 1047 preceding 19 August 2024 further mentioned open source AI in the context of the Board of Frontier Models, which is a state entity that would be established by the bill to oversee the development of such models, and the Frontier Model Division, which would be directly supervised by the Board of Frontier Models.”

• cites [114]Not matched

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

May combine two claims.

In 2025, California Governor Newsom signed SB 243, AB 853, and SB 53 (TFAIA) into law.

“In 2025, Governor Newsom signed into law several new bills regulating AI, which include SB 243 110 (the Companion Chatbot law), AB 853 111 (amending California AI Transparency Act), and SB 53 112 , known as Transparency in Frontier Artificial Intelligence Act (TFAIA).”

• cites [110]Not matched

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

• cites [111]Not matched

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

• cites [112]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).

The TFAIA has been named a "landmark AI law that establishes the nation's first comprehensive framework for transparency, safety and accountability in the development and deployment of advanced AI models."

“The TFAIA has been named a “landmark AI law that establishes the nation's first comprehensive framework for transparency, safety and accountability in the development and deployment of advanced AI models.”113”

• cites [113]Not matched

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

May combine two claims.

According to Corporate Europe Observatory, in 2023, 185 out of 277 AI-related meetings with members of the European Parliament were held with industry and trade associations, with Google, Microsoft, American Chamber of Commerce, and OpenAI being the top lobbyists.

“According to Corporate Europe Observatory, in 2023, the Big Tech's focus was on influencing the Parliament's position, which is reflected in the number of meetings with members of the European Parliament on AI: 185 out of 277 AI-related meetings were held with industry and trade associations, with Google, Microsoft, American Chamber of Commerce, and OpenAI being the top lobbyists.83”

• cites [83]Not matched

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

May combine two claims.

America's AI Action Plan, published in July 2025, addresses "Open-source and open-weights AI" and includes policy recommendations to help the Federal government create a supportive environment for open models.

““Open-source and open-weights AI” are addressed in America's AI Action Plan published in July 2025 104 , which includes a number of policy recommendations that would help the Federal government create a supportive environment for open models.”

• cites [104]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). May combine two claims.

The exemption granted to open source GPAI model providers established in third countries can hardly be explained and goes contrary to the other provisions of the AI Act.

“Whereas the tiered regulatory approach to GPAI models may be justified by the argument that the focus should be on the risks posed and not on particular technology 92 , the exemption granted to open source GPAI model providers established in third countries can be hardly explained and goes contrary to the other provisions of the AI Act.”

• cites [92]Not matched

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

May combine two claims.

Unlike in the EU AI Act, open source AI is not exempt from the application of the TFAIA, so open source AI model developers are subject to the obligations imposed by the bill if such models fall within the definition of covered models.

“Unlike in the EU AI Act, open source AI is not exempt from the application of the TFAIA, so open source AI model developers are subject to the obligations imposed by the bill if such models fall within the definition of covered models.”

• cites [112]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). May combine two claims.

Although the EU has pioneered in enacting a comprehensive legal framework in the area of AI, other jurisdictions have also undertaken a number of legislative initiatives to regulate AI.

“Although the EU has pioneered in enacting a comprehensive legal framework in the area of AI, other jurisdictions have also undertaken a number of legislative initiatives to regulate AI.”

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

Meta suggested that the exemption should cover not only AI components provided under free and open source license, but also under similarly permissive licences; this suggestion did not find its way into the final text of the AI Act.

“Apparently, Meta was setting hopes on utilizing the open source exemption more broadly, as it also suggested that the exemption should cover not only AI components provided under free and open source license, but also under similarly permissive licences that: (i) provide open access to models; (ii) further the goal of fostering collaboration and innovation; and (iii) permit downstream users to use”

May combine two claims.

Another exemption for GPAI models released under free and open source license is stipulated in Article 54(6); it is reserved only for providers established in third countries and concerns the obligation to appoint an authorized representative established in the Union.

“Another exemption for GPAI models released under free and open source license is stipulated in Article 54(6).However, it is reserved only for providers that are established in third countries and concerns the obligation to appoint an authorized representative which is established in the Union (Article 54(6)).”

May combine two claims.

Despite the explicit exemption, the AI Act, in practical terms, is fully applicable to open source AI projects, unless the providers are established in a third country pursuant to Article 54.

“This results in the overall conclusion that, despite the explicit exemption, the AI Act, in practical terms, is fully applicable to open source AI projects, unless the providers are established in a third country, pursuant to Article 54.”

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

California, the leader in AI technologies development across the U.S., has been one of the most active states in introducing AI regulations.

“California, the leader in AI technologies development across the U.S. 95 , has been one of the most active states in introducing AI regulations.”

• cites [95]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).

The AI Foundation Model Transparency Act would direct the FTC to consider including "alternative provisions" for open source foundation models when establishing transparency standards.

“At the same time, the FTC, when establishing the standards to improve the transparency of foundation models, would be directed to consider including “alternative provisions” for open source foundation models (Section 3(g)(1)).”

• cites [101]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).

In U.S. federal government terms, for a foundation model to be open, the only element that must be disclosed is the weights.

“So, in the terms used by the U.S. federal government, in order for a foundation model to be open, the only element to be disclosed is the weights, meaning numerical parameters within an AI model that help determine the model's outputs in response to inputs (Section 3(u) of Executive Order 14110).”

• cites [98]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).

The TFAIA was approved a year after SB 1047, a previous bill addressing safety and security for frontier AI models, had been vetoed.

“It was approved a year after SB 1047 114 , a previous bill addressing safety and security for frontier AI models, had been vetoed.115”

• cites [114]Not matched

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

• cites [115]Not matched

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

May combine two claims.

Out of 97 total AI-related meetings with senior Commission officials, 84 were held with industry and trade associations, twelve with civil society, and just one with academics or research institutes.

“As to AI-related meetings with senior Commission officials, out of 97 total meetings, 84 were held with industry and trade associations, twelve with civil society, and just one with academics or research institutes.84”

• cites [84]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).

The NTIA Report defines an open foundation model as a dual-use foundation model whose model weights have been released openly to the public, either by allowing users to download them from the Internet or through other mechanisms.

“A dual-use foundation model with widely available model weights (or, in this Report, an open foundation model) is a dual-use foundation model whose model weights have been released openly to the public, either by allowing users to download them from the Internet or through other”

• cites [98]Not matched

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

May combine two claims.

With the change of administration in 2025, President Biden's executive actions were revoked and new executive orders addressing AI were signed by President Trump.

“With the change of administration in 2025, the executive actions taken by President Biden were revoked and new executive orders addressing AI have been signed by President Trump.103”

• cites [103]Not matched

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

May combine two claims.

Such unequal treatment of domestic and third-country providers of GPAI models, as well as very limited practical exemptions from the obligations under the AI Act, are most likely to disincentivize European GPAI model providers to open source their models, thus undermining the development of open source AI in the EU.

“Such unequal treatment of domestic and third-country providers of GPAI models, as well as very limited practical exemptions from the obligations under the AI Act, are most likely to disincentivize European GPAI model providers to open source their models, thus undermining the development of open source AI in the EU.”

Classified as probably not a substantial claim.

Articles 53(2) and 54(6) of the AI Act reiterate the provisions of Recital 104 in the part on excluding GPAI models with systemic risk from the overall exemption under Article 2(12).

“Articles 53(2) and 54(6) of the AI Act reiterate the provisions of Recital 104 in the part on excluding GPAI models with systemic risk from the overall exemption under Article 2 (12).”

May combine two claims.

Similar to data privacy regulation, the U.S. federal government has been reluctant to enact a federal law regulating AI technologies, which has led to states taking the initiative, despite the Trump administration explicitly forbidding state AI regulations that conflict with federal policy.

“Similar to the situation with regulating data privacy, the U.S. federal government has been reluctant to enact a federal law regulating AI technologies, which has led to the states taking the initiative in their hands, despite the Trump administration explicitly forbidding state AI regulations that conflict with the federal policy.94”

• cites [94]Not matched

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

May combine two claims.

Pursuant to Article 53(2), the obligations to draw up and keep up-to-date technical documentation and to draw up, keep up-to-date and make available information and documentation to providers of AI systems do not apply to GPAI models released under free and open source license.

“In particular, pursuant to Article 53(2), the following obligations do not apply to the GPAI models released under free and open source license: • draw up and keep up-to-date the technical documentation of the model, including its training and testing process and the results of its evaluation, which shall contain, at a minimum, the information set out in Annex XI for the purpose of providing it,”

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

In September 2023, Meta shared its position on the AI Act and proposed to strike out the exception of foundation models from the overall exemption for open source AI.

“In September 2023, Meta shared its own position on the AI Act.90 The very first amendment proposed by Meta was to strike out the exception of foundation models from the overall exemption for open source AI.”

• cites [90]Not matched

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

May combine two claims.

The Stanford 2025 AI Index Report notes that while the U.S. maintains its lead in quantity, Chinese models have rapidly closed the quality gap: performance differences on major benchmarks shrank from double digits in 2023 to near parity in 2024.

“The report specifically notes that, while the U.S. maintains its lead in quantity, Chinese models have rapidly closed the quality gap: performance differences on major benchmarks shrank from double digits in 2023 to near parity in 2024.”

• cites [119]Not matched

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

May combine two claims.

Considering the jurisprudence on the use of copyrighted works for AI training in the EU and the U.S., where the EU shows a clear inclination towards prohibiting such use and the U.S. leaves AI developers to rely on fair use, such regulatory exemption stands out by its pro-innovation character.

“Considering the jurisprudence on the use of copyrighted works for AI training both in the EU and the U.S., where, in the case of the EU, there is a clear inclination towards prohibiting such use 134 and, in the case of the U.S., the AI developers are left to rely on the flexible concept of fair use, 135 such regulatory exemption stands out by its pro-innovation character.”

• cites [134]Not matched

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

• cites [135]Not matched

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

May combine two claims.

China embraced open source software as early as the end of the 20th century, and, fueled by government investments, open source is now also thriving in cloud computing and AI.

“China embraced open source software as early as at the end of the 20 th century, and, fueled by government investments, open source is now also thriving in cloud computing and AI.121”

• cites [121]Not matched

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

May combine two claims.

The model law distinguishes two liability exemption possibilities: (1) in case code modules are provided for AI R&D on a free and open-source basis, disclosing the functions and risks, such providers are exempt from legal liability, and (2) providers of free and open source AI systems are either exempt from liability or are subject to reduced legal liability in case they have established an AI合规治理

“The model law thus distinguishes two liability exemption possibilities: (1) in case code modules are provided for AI R&D on a free and open-source basis, disclosing the functions and risks, such providers are exempt from legal liability, and (2) providers of free and open source AI systems are either exempt from liability or are subject to reduced legal liability in case they have established an”

• cites [129]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).

The draft law encourages developers of open-source foundation models to provide appropriate compensation to copyright holders by reasonable means, so as to promote the dissemination and use of copyrighted works.

“Developers of open-source foundation models are encouraged to provide appropriate compensation to copyright holders by reasonable means, so as to promote the dissemination and use of copyrighted works.”

• cites [129]Not matched

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

May combine two claims.

Considering that the language on creating incentives for developers of open source AI models was removed from the final version of SB 1047, California seems reluctant to provide political or economic incentives for open source AI and aims at imposing the same safety and transparency obligations on developers of both open source and closed source AI, without any exemption for the former.

“However, considering that the language on creating incentives for developers of open source AI models was removed from the final version of SB 1047, California seems reluctant to provide political or economic incentives for open source AI and aims at imposing the same safety and transparency obligations on developers of both open source and closed source AI, without any exemption for the former.”

• cites [114]Not matched

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

Classified as probably not a substantial claim.

A notable paper published in October 2025, prepared by leading AI experts representing, among others, AI Governance Center of Aliresearch (Alibaba Group) and DeepSeek-AI, examines the regulatory landscape that enhances AI openness.

“A notable paper published in October 2025, prepared by leading AI experts representing, among others, AI Governance Center of Aliresearch (Alibaba Group) and DeepSeek-AI, examines the regulatory landscape that enhances AI openness.136”

• cites [136]Not matched

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

May combine two claims.

Chinese groups, constrained by GPU shortages, are willing to experiment on wild ideas to improve their models, which is bearing results and has led to more efficient models that are cheaper to train and use, which can appeal to budget-conscious clients and help Chinese firms find a niche market alongside the AI giants.

“Chinese groups, constrained by GPU shortages, are “willing to experiment on wild ideas” to improve their models, which is bearing results and has led to “more efficient models that are cheaper to train and use, which can appeal to budget-conscious clients and help the Chinese firms find a niche market alongside the AI giants.”127”

• cites [127]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). May combine two claims.

For both large companies, such as Alibaba, and start-ups, such as DeepSeek and 01.AI, open sourcing AI models offers faster commercialization and global recognition.

“Besides, for both large companies, such as Alibaba, and start-ups, such as DeepSeek and 01.AI, open sourcing AI models offers faster commercialization and global recognition.123”

• cites [123]Not matched

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

May combine two claims.

According to Stanford's 2025 AI Index Report, 15 notable AI models originated from China-based institutions in 2024, which is the second-best result in the world, after 40 AI models originating from the U.S. and followed by three European AI models.

“According to Stanford's 2025 AI Index Report, 15 notable AI models originated from China-based institutions in 2024, which is the second-best result in the world, after 40 AI models originating from the U.S. and followed by three European AI models.119”

• cites [119]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).

Exempt from AB 412's requirement are AI developers that have trained their AI models exclusively on datasets that the developer makes publicly available at no cost to users on its website.

“Exempt from this requirement are, among others, AI developers that have trained their AI models exclusively on datasets that the developer makes publicly available at no cost to users on its website.”

• cites [117]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).

As of June 2026, the most recent open weight model Qwen 3.5 is often considered by the community to outperform U.S. models, such as Google DeepMind's Gemma 4.

“As of June 2026, the most recent open weight model Qwen 3.5 is often considered by the community to outperform U.S. models, such as Google DeepMind's Gemma 4. 125”

• cites [125]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).

A few months after the introduction of OpenAI's ChatGPT, China announced it would work on a model law on AI.

“A few months after the introduction of OpenAI's ChatGPT, China also announced it would work on a model law on AI. 129”

• cites [129]Not matched

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

May combine two claims. The quote only partly states this claim.

According to that paper, China's approach to promoting AI openness is based on six regulatory pillars: presumptive regulatory exemptions for open-source AI, including frontier AI models; presumptive exemption for AI scientific research; fast and efficient adjudication of AI cases via Internet courts; contextualized and evolving approach to AI-assisted scientific research; active development of科学

“According to the paper, China's approach to promoting AI openness is based on six regulatory pillars: presumptive regulatory exemptions for open-source AI, including frontier AI models; presumptive exemption for AI scientific research; fast and efficient adjudication of AI cases 137 (copyright cases concerning model training, platform liability, IP protection of model parameters) via, inter alia,”

• cites [136]Not matched

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

• cites [137]Not matched

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

• cites [138]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).

Article 25 "Fiscal and Procurement Support" encourages governments at all levels, as well as state-owned enterprises and public institutions, to procure open-source artificial intelligence products and services that conform to national standards.

“Government support for open source AI is further stated in Article 25 “Fiscal and Procurement Support”, encouraging governments at all levels, as well as state-owned enterprises and public institutions, “to procure open-source artificial intelligence products and services that conform to national standards.””

• cites [129]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).

Article 90 of version 3.0 of China's draft model law provides definitions of AI and foundation model.

“Article 90 of version 3.0 provides for the definitions as used throughout the text, including definitions of AI and foundation model:”

• cites [129]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).

In version 3.0 of the draft law, the definition of open source AI was added in Article 90(b).

“Importantly, in version 3.0 of the draft law, the definition of open source AI was added (Article 90(b)):”

• cites [129]Not matched

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

May combine two claims.

It is evident that China sees great potential in open source AI and is aiming at establishing significant incentives for providers of open source AI systems, particularly those who make their systems available for the purpose of AI R&D.

“However, it is evident that China sees great potential in open source AI and is aiming at establishing significant incentives for providers of open source AI systems, particularly those who make their systems available for the purpose of AI R&D.”

• cites [129]Not matched

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

May combine two claims.

Overall, U.S. federal and state (particularly Californian) regulations provide little, if any, incentives for open source AI projects, which is most likely to serve as a disincentive for developers to open source their AI models.

“Overall, the U.S. federal and state (particularly, Californian) regulations provide little, if any, incentives for open source AI projects, which is most likely to serve as a disincentive for developers to open source their AI models.”

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

AB 412 faces strong opposition by the tech community in California claiming that the obligations to be imposed on AI developers are not feasible.

“The bill faces strong opposition by the tech community in California claiming that the obligations to be imposed on AI developers are not feasible.118”

• cites [118]Not matched

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

May combine two claims.

The definition of AI in China's draft model law is similar to the ones applied by the EU and the U.S., whereas for foundation models China takes the more flexible approach of not stipulating exact computing power standards in a legal act, but delegating this power to a dedicated newly-established authority.

“Thus, the definition of AI is similar to the ones applied by the EU and the U.S., whereas, as regards foundation models, China is taking the more flexible approach of not stipulating the exact computing power standards in a legal act, but rather delegating this power to a dedicated newly-established authority.”

• cites [129]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). May combine two claims.

The draft definition acknowledges that open source AI models differ by the degree of openness and that it is impractical to establish precise disclosure requirements, but a better approach is to set the goal that such openness is meant to pursue.

“However, this definition acknowledges that open source AI models differ by the degree of openness and that it is impractical to establish precise disclosure requirements, but a better approach is to set the goal that such openness is meant to pursue:”

• cites [129]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).

California is eager to consider the opinion of the open source community in enacting policies on regulations concerning AI.

“Thus, as made clear from the wording of the TFAIA, California is eager to consider the opinion of the open source community in enacting policies on regulations concerning AI.”

• cites [112]Not matched

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

May combine two claims.

The understanding of open source AI in the U.S. significantly differs from the OSI's OSAID or the EU's "AI system released under free and open source licence", as the regulations, both at federal and state level, either refer to models for which solely the weights are widely available or use "open source" without providing a clear definition of the term.

“In sum, the understanding of open source AI in the U.S. significantly differs from the OSI's OSAID or EU's “AI system released under free and open source licence”, as the regulations, both at federal and state level, either refer to models for which solely the weights are widely available or use “open source” without providing a clear definition of the term.”

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

One of the five members of the Board of Frontier Models would be a member of the open source community, appointed by the Governor, subject to Senate confirmation.

“Besides, one of the five members of the Board of Frontier Models would be a member of the open source community, appointed by the Governor, subject to Senate confirmation.”

• cites [114]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).

Article 26 provides another economic incentive: preferential tax treatment for the R&D of open source AI, such as super-deductions for R&D expenditures, to be stipulated by the National AI Administrative Authority.

“Another economic incentive is a preferential tax treatment for the R&D of open source AI, such as super-deductions for R&D expenditures, that should be stipulated by the National AI Administrative Authority (Article 26).”

• cites [129]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).

China has enacted multiple regulations concerning AI technologies.

“China has enacted multiple regulations concerning AI technologies.128”

• cites [128]Not matched

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

May combine two claims.

As regards the elements that should be made publicly available, the provisions of the AI Act deviate significantly from the OSAID, omitting requirements to make publicly available the source code used to train and run the system and the information on data used to train the system.

“However, as regards the elements that should be made publicly available (in order to have access to the preferred form to make modifications to the system), the provisions of the AI Act deviate significantly from the OSAID, omitting requirements to make publicly available (1) the source code used to train and run the system and (2) the information on data used to train system.”

Classified as probably not a substantial claim.

Compared to the OSAID elaborated by the OSI, the term used in the AI Act represents a mixture of the freedoms that the user (deployer) should have and the elements that should be made publicly available.

“Compared to the OSAID elaborated by the OSI, the term used in the AI Act represents a mixture of (1) the freedoms that the user (deployer) should have, and (2) the elements that should be made publicly available.”

Classified as probably not a substantial claim.

The authors of that paper conclude that these six pillars constitute a policy portfolio that enhances an innovation-and openness-friendly environment for AI developers.

“The authors conclude that these six pillars constitute a policy portfolio that enhances innovation-and openness-friendly environment for AI developers.”

• cites [136]Not matched

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

May combine two claims.

The training datasets disclosure requirement enacted for generative AI systems in California mirrors the obligations for providers of GPAI models under the EU AI Act, since the latter does not exempt GPAI models released under a free and open source license from the obligation to make publicly available a sufficiently detailed summary about the content used for training (Article 53(1)(d)).

“The training datasets disclosure requirement enacted for generative AI systems in California mirrors the obligations for providers of GPAI models under the EU AI Act, since the latter does not provide an exemption for GPAI models released under free and open source license from the obligation to make publicly available a sufficiently detailed summary about the content used for training of the GPAI”

May combine two claims.

In sum, China's draft model law is aiming at providing significant economic, legal, and political incentives for open source AI, which demonstrates that China is planning to fully employ the potential of open source AI.

“In sum, China's draft model law is aiming at providing significant economic, legal, and political incentives for open source AI, which demonstrates that China is planning to fully employ the potential of open source AI.”

• cites [129]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).

Introducing the exemption for open source AI only at the later stages of the legislative process might be the reason why the final version of the regulation does not define the term "AI systems released under free and open-source licences."

“Introducing the exemption for open source AI only at the later stages of the legislative process might be the reason why the final version of the regulation does not define the term “AI systems released under free and open-source licences.””

• cites [51]Not matched

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

May combine two claims.

Despite the fact that the published version is only an interim draft, the revealed provisions on open source AI demonstrate China's intention to create a favorable legal regime for open source AI projects in China, by way of providing exemptions and reductions in liability, particularly for AI R&D, preferential tax treatment, supporting open source AI products in the course of public procurement,

“Despite the fact that the published version is only an interim draft, the revealed provisions on open source AI demonstrate China's intention to create a favorable legal regime for open source AI projects in China, by way of providing exemptions and reductions in liability, particularly for AI R&D, preferential tax treatment, supporting open source AI products in the course of public procurement,”

• cites [129]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).

Such approach is even more surprising as the EU is proclaiming its goals to strengthen its digital sovereignty and competitiveness in innovation.

“Such approach is even more surprising as the EU is proclaiming its goals to strengthen its digital sovereignty and competitiveness in innovation.”

Classified as probably not a substantial claim.

The AI Act does not clarify what should be included in the 'information on model architecture' and 'information on model usage', particularly whether source code should be included in such information.

“Furthermore, the AI Act does not clarify what should be included in the “information on model architecture” and “information on model usage”, particularly whether source code should be included into such information.”

May combine two claims.

Taking into account the OSAID developed by the OSI, most, if not all, of these disclosures either are necessary in order for an AI system to be considered open source in accordance with the understanding of OSAI by the open source community or would otherwise need to be made under the AI Act.

“Furthermore, taking into account the OSAID developed by the OSI, most, if not all, of these disclosures either are necessary in order for an AI system to be considered open source in accordance with the understanding of OSAI by the open source community or would otherwise need to be made under the AI Act.”

• cites [75]Not matched

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

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

Based on Recital 102, a GPAI model is distributed under a free and open source license if the parameters (including weights), information on model architecture, and information on model usage are made publicly available, or if the license allows users to run, copy, distribute, study, change and improve software and data, including models, under the condition that the original provider is credited.

“Assuming that both of the cited sentences include requirements that free and open source licenses should satisfy, then, based on Recital 102, a GPAI model is distributed under free and open source license, if: 1. the following is made publicly available about the GPAI model: • parameters, including weights, • information on model architecture, • information on model usage, or 2. such license”

• AssumptionThe authors state this as an assumption: Both of the cited sentences of Recital 102 include requirements that free and open source licenses should satisfy.

May combine two claims.

The U.S. federal government and California, at least to the moment, have not provided any benefits for open source AI, imposing the same obligations on providers of both open and closed source AI models.

“The U.S. federal government and California, at least to the moment, have not provided any benefits for open source AI, imposing the same obligations on providers of both open and closed source AI models.”

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

This fact highlights the importance that the OSI is attaching to the openness of training datasets for the OSAID.

“This fact highlights the importance that the OSI is attaching to the openness of training datasets for the OSAID.”

• cites [75]Not matched

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

Classified as probably not a substantial claim.

None of the cited provisions of the AI Act provides a clear definition of what is precisely understood under a GPAI model released under a free and open source license.

“Thus, none of the cited provisions of the AI Act provides a clear definition of what is precisely understood under a GPAI model released under free and open source license.”

Classified as probably not a substantial claim.

Based on Articles 53(2) and 54(6) of the AI Act, a GPAI model is distributed under a free and open source license if the license allows for the access, usage, modification, and distribution of the model, and the parameters (including weights), information on model architecture, and information on model usage are made publicly available.

“Based on the above text of Articles 53(2) and 54(6), a GPAI model is distributed under free and open source license, if: 1. such license allows for the access, usage, modification, and distribution of the model, and 2. the following is made publicly available with respect to the GPAI model: • parameters, including weights, • information on model architecture, • information on model usage.”

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

Taken together, the cited provisions of the AI Act are inconsistent and confusing.

“Furthermore, taken together, the cited provisions are inconsistent and confusing.”

May combine two claims.

The requirement to disclose training data information is only a transparency obligation and, under the AI Act, does not form a prerequisite for a GPAI model to be considered 'released under a free and open-source licence'.

“such requirement to disclose training data information is only a transparency obligation and, under the AI Act, does not form a prerequisite for a GPAI model to be considered “released under a free and open-source licence”.”

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

There is a direct contradiction between the term as used in the AI Act and the OSAID suggested by the OSI, as the latter puts a strong emphasis on the openness of information on the data used for the training of AI systems.

“Thus, there is a direct contradiction between the term as used in the AI Act and the OSAID suggested by the OSI, as the latter puts a strong emphasis on the openness of information on the data used for the training of AI systems.”

May combine two claims.

The open source exemption was included in the provisional agreement on the AI Act reached by the Council and the Parliament on 9 December 2023 as Article 2(5g) with almost the same wording as in the final version of the Regulation published on 21 May 2024.

“It was included in the provisional agreement on the AI Act reached by the Council and the Parliament on 9 December 2023 as Article 2(5g) with almost the same wording as in the final version of the Regulation published on 21 May 2024.51”

• cites [51]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).

The previous Release Candidate 1 version of the OSAID published on 2 October 2024 included the words "a detailed description of all data used for training", which has been replaced with "the complete description of all data used for training".

“The previous, Release Candidate 1, version of the OSAID published on 2 October 2024 included the words “a detailed description of all data used for training”, which has been replaced with “the complete description of all data used for training””

• cites [75]Not matched

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

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

Considering the very limited exemptions for open source AI systems in the AI Act and a more favorable regime for third-country providers, the EU's approach to open source raises a lot of questions.

“Considering the very limited exemptions for open source AI systems in the AI Act and a more favorable regime for third-country providers, the EU's approach to open source raises a lot of questions.”

• cites [11]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 S2 (checkable). May combine two claims.

Similar to open source software, open source AI possesses significant economic potential and, despite security concerns against open sourcing highly capable AI models, offers benefits that, with necessary security mechanisms in place, outweigh the possibility of potential harm.

“Similar to open source software, open source AI possesses significant economic potential and, despite security concerns against open sourcing highly capable AI models, offers benefits that, with necessary security mechanisms in place, outweigh the possibility of potential harm.”

• cites [139]Not matched

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