# Open Source AI Definition

> Source: https://aiwiki.ai/wiki/open_source_ai_definition
> Updated: 2026-11-17
> Fact-checked: 2026-11-17
> Categories: AI Policy & Regulation, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Open Source AI Definition." aiwiki.ai, 17 Nov 2026. https://aiwiki.ai/wiki/open_source_ai_definition
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

The **Open Source AI Definition** (**OSAID**) is a document published by the [Open Source Initiative](https://aiwiki.ai/wiki/open_source_initiative) (OSI) that states what an artificial intelligence system must make available, and under what terms, to be called open source. Version 1.0 was released on 28 October 2024 at the All Things Open conference in Raleigh, North Carolina, at the end of a process OSI began in 2022.[1][2] The definition requires that a system grant four freedoms (use, study, modify and share) and that its developer provide the "preferred form to make modifications", which for [machine learning](https://aiwiki.ai/wiki/machine_learning) systems means three things: detailed information about the training data, the complete code used to train and run the system, and the model parameters.[1]

As of November 2026, version 1.0 is the only stable version. OSI has published no revision and no new release candidate, although in October 2026 it began a series of meetings to work out "whether and how the OSAID should evolve".[36] The definition asks for a description of the [training data](https://aiwiki.ai/wiki/training_set) and not for the data itself, and this has been its most disputed feature. Free-software advocates, among them Bradley Kuhn of the Software Freedom Conservancy, say the requirement is too weak, [Meta](https://aiwiki.ai/wiki/meta_ai) rejected the definition, and some analysts argue that a single yes-or-no test does not fit AI.[11][15][16] The broader subject is covered at [open-source AI](https://aiwiki.ai/wiki/open_source_ai).

## Background

OSI, founded in 1998, maintains the Open Source Definition (OSD), a ten-point test that it applies to software licences.[2][40] The second point of the OSD says: "The source code must be the preferred form in which a programmer would modify the program."[40] That sentence has no obvious equivalent for a trained model. OSI's FAQ describes machine-learning systems as blending software with data, configuration options, documentation and new artifacts such as weights, and says that when the project started nobody had a clear answer to what the preferred form for modifying an AI system is.[3]

A second motive was the way the label was being used. In June 2023 OSI's executive director at the time, Stefano Maffulli, objected to language models, tools and services that were announced as "open" or "Open Source" while carrying restrictions the OSD does not allow.[11] OSI's own account adds a regulatory reason: partners it consulted in early 2023 worried that the term would become generic and that European Union regulators would write a definition without community input, so the board set a deadline of October 2024.[4]

## Drafting process

### Research and co-design

OSI describes the work in five phases.[4]

| Phase | Period | What happened |
|---|---|---|
| Preliminary research | July 2022 to December 2023 | "Deep Dive: AI" podcasts and panels in 2022, then 18 webinars with 37 speakers in September 2023[4] |
| Four freedoms | October to November 2023 | In-person workshops in Monterey, Raleigh and Addis Ababa drafted the wording of the freedoms[4] |
| System analysis | February to March 2024 | Four online workgroups studied [BLOOM](https://aiwiki.ai/wiki/bloom), OpenCV, [Llama 2](https://aiwiki.ai/wiki/llama_2) and Pythia and voted on which components are needed to exercise the freedoms[4] |
| System validation | May to July 2024 | Volunteers reviewed 13 systems that described themselves as open[4] |
| Training data workshop | September to October 2024 | 18 specialists from 15 countries met in Paris on 10 and 11 October 2024[4] |

A kickoff meeting took place on 21 June 2023 at Mozilla's headquarters in San Francisco, with OSI board members and people from the Mozilla Foundation, Creative Commons, the Wikimedia Foundation, the Internet Archive, Linux Foundation Europe and OSS Capital.[11][13] The public co-design process opened at All Things Open in October 2023.[4] That autumn OSI hired the co-design firm Do Big Good, whose Mer Joyce facilitated the workshops, and from January 2024 it held online town halls roughly every two weeks.[4][13] A grant from the Alfred P. Sloan Foundation supported outreach to participants around the world.[4]

The system-analysis phase produced the list of required components. Workgroup members started from the component list of the [Model Openness Framework](https://aiwiki.ai/wiki/model_openness_framework), then in pre-release, and voted on whether each component was needed to use, study, modify and share a system. Training, validation and testing code, [inference](https://aiwiki.ai/wiki/inference) code, model architecture, model parameters and supporting libraries came out as required; data preprocessing code as likely required; training datasets only as "maybe required". OSI wrote these results into draft 0.0.6.[4] Critics later attacked both the method and the outcome (see below).

### Who took part

OSI counts more than 50 co-design volunteers from nearly 30 countries.[4] It identified six stakeholder groups: system creators, licence creators, regulators, licensees, end users and "subjects", meaning people affected by a system's output.[5] According to OSI, participants joined in a personal capacity; their employers included the Mozilla Foundation, the [Linux Foundation](https://aiwiki.ai/wiki/linux_foundation), Creative Commons, the Wikimedia Foundation, [EleutherAI](https://aiwiki.ai/wiki/eleutherai), [Hugging Face](https://aiwiki.ai/wiki/hugging_face), GitHub, [Microsoft](https://aiwiki.ai/wiki/microsoft), [Google](https://aiwiki.ai/wiki/google), Amazon, Meta, Intel and Samsung.[5] The Llama 2 workgroup, for example, included two Meta employees, a Debian developer and an official from the French government's digital directorate.[4]

### Drafts and release candidates

OSI published each draft for comment. The dates below are those of its drafts index.[6]

| Version | Date | Notes |
|---|---|---|
| 0.0.2 | 2023-10-11 | Earliest draft in the index |
| 0.0.3 | 2023-10-27 | |
| 0.0.4 | 2023-12-13 | |
| 0.0.5 | 2024-01-30 | |
| 0.0.6 | 2024-03-09 | Added the components recommended by the workgroups[4] |
| 0.0.7.1 | 2024-04-12 | |
| 0.0.8 | 2024-04-24 | |
| 0.0.9 | 2024-08-22 | Stated that models and weights are covered; checklist split into its own document[7][10] |
| 1.0-RC1 | 2024-10-02 | Data information, code and parameters became a mandatory list; copyleft-style conditions allowed[8] |
| 1.0-RC2 | 2024-10-21 | Tightened the data wording; added a paragraph on the legal status of parameters[9] |
| 1.0 | 2024-10-28 | Stable release; definition text identical to RC2[1][9] |

The OSI board, which was due to vote on 27 October 2024, approved the text that month.[11][34]

## The definition

### Four freedoms

The definition says that an open source AI is a system made available under terms, and in a way, that grant the freedoms to:[1]

- "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."

A footnote credits the Free Software Definition as the source of the four freedoms. The text adds that they apply both to a complete system and to its separate elements, and that exercising them depends on access to "the preferred form to make modifications to the system".[1]

### Preferred form of making modifications

For machine-learning systems the preferred form "must include all the elements below":[1]

| Element | What the text requires | Legal terms | Contents the text names |
|---|---|---|---|
| Data Information | "Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system" | "OSI-approved terms" | Must include: a complete description of all training data; a list of public data and where to get it; a list of data obtainable from third parties, including for a fee |
| Code | "The complete source code used to train and run the system" | "OSI-approved licenses" | Must include, where used: data processing and filtering code, training code with its arguments and settings, validation and testing code, [tokenizers](https://aiwiki.ai/wiki/tokenization), [hyperparameter](https://aiwiki.ai/wiki/hyperparameter) search code, inference code, model architecture |
| Parameters | "The model parameters, such as weights or other configuration settings" | "OSI-approved terms" | May include: [checkpoints](https://aiwiki.ai/wiki/checkpoint) from key stages of training and the final [optimizer](https://aiwiki.ai/wiki/optimizer) state |

The terms applied to these elements "may contain conditions that require any modified version to be released under the same terms as the original", so copyleft-style conditions are compatible with the definition.[1]

### Scope and limits of the text

The requirements are the same whether the thing released is called a system, a model or a set of weights. Anything described as an open source model or open source weights must come with the data information and code used to derive the [parameters](https://aiwiki.ai/wiki/parameter).[1] The definition borrows its meaning of "AI system" from the OECD's 2024 recommendation on artificial intelligence.[1]

The text takes no position on whether parameters can be protected by [copyright](https://aiwiki.ai/wiki/copyright). It says the definition "does not require a specific legal mechanism for assuring that the model parameters are freely available to all", and OSI's FAQ explains that this is why parameters fall under "terms" while code falls under "licenses".[1][3]

The definition is silent on safety, ethics and risk. OSI's FAQ says responsible development and regulation are "a separate conversation".[3] A companion checklist maps the definition onto the 17 components of the Model Openness Framework. Datasets or equivalent data documentation, four kinds of code, the model architecture and the parameters are required; evaluation data and code, [model cards](https://aiwiki.ai/wiki/model_card) and sample outputs are optional. OSI warns that the checklist is closely tied to [generative AI](https://aiwiki.ai/wiki/generative_ai) and "is not an operating manual".[5][12]

## The data information compromise

The definition does not require a developer to publish the training set. It requires a description detailed enough for a skilled person to build a "substantially equivalent" system, plus lists of the public and purchasable data with their sources.[1] A developer therefore has to disclose where the data came from, its scope and characteristics, how it was selected, how it was labelled, and how it was processed and filtered, and has to do so for unshareable data too.[1]

OSI's FAQ sorts training data into four classes and attaches a different duty to each.[3]

| Class | Example | OSI's stated expectation |
|---|---|---|
| Open | Data that can be copied, modified and reshared | "This must be shared" |
| Public | Data others can inspect while it stays online | Disclosed with full details of where to obtain it |
| Obtainable | Data available from third parties, including for a fee | Disclosed with full details of where to obtain it |
| Unshareable | Personal or otherwise legally restricted data | Described in detail |

The duty to share open data appears in the FAQ and in OSI's later summaries, which speak of the complete training data or, where it cannot be distributed, detailed information about it. The definition text itself asks only for a listing of public data and where to obtain it.[1][3][34]

OSI gives two reasons for the design. The first is legal: laws that permit training on data often do not permit redistributing it, privacy rules protect health and other personal records, and much Indigenous knowledge is governed by arrangements that do not fit open licensing. The FAQ's example is a hospital that could build a dataset "with identical structure using its own patient data" from a good description.[3] The second is a judgment about practice. The workgroups rated training and data-processing code as more important for modifying a system than the data, and OSI concluded that a full-data rule "would relegate Open Source AI to a niche of AI trainable only on open data".[3] The FAQ states that the requirement, like the OSD, "doesn't mandate full reproducibility and transparency but enables them".[3]

The wording hardened as the drafts progressed. Draft 0.0.9 asked for information allowing a skilled person to "recreate a substantially equivalent system using the same or similar data" and gave its contents as examples, and OSI's announcement said it had decided that "defining training data as a benefit, not a requirement, is the best way to go".[7][10] RC1 turned the examples into a list that "must" be provided and added the two listings. RC2 changed "a detailed description" to "the complete description" and "data cleaning" to "data processing and filtering".[8][9]

## Which systems comply

During the validation phase volunteers examined 13 systems: Arctic, BLOOM, [Falcon](https://aiwiki.ai/wiki/falcon), [Grok](https://aiwiki.ai/wiki/grok), Llama 2, the LLM360 models, Phi-2, Mistral (Mixtral in the results), [OLMo](https://aiwiki.ai/wiki/olmo), OpenCV, Pythia, [T5](https://aiwiki.ai/wiki/t5) and Viking. They worked from drafts 0.0.6 to 0.0.8. OSI reports that most reviewers could not finish because the documentation they needed was not public, and that LLM360's review was a self-assessment.[4]

| Result reported by OSI | Systems |
|---|---|
| Passed | Pythia (EleutherAI), OLMo ([Allen Institute for AI](https://aiwiki.ai/wiki/allen_institute_for_ai)), Amber and CrystalCoder (LLM360), T5 (Google)[3][4] |
| Would probably pass with different licence or legal terms | BLOOM (BigScience), [StarCoder2](https://aiwiki.ai/wiki/starcoder) (BigCode), Falcon (Technology Innovation Institute)[3][4] |
| Did not pass | Llama 2 (Meta), Grok ([xAI](https://aiwiki.ai/wiki/xai), listed by OSI as X/Twitter), Phi-2 (Microsoft), [Mixtral](https://aiwiki.ai/wiki/mixtral) ([Mistral AI](https://aiwiki.ai/wiki/mistral_ai))[3][4] |

OSI says the failing systems lacked required components, had legal agreements incompatible with open source principles, or both.[3] It attaches a caveat to the whole list: "These results should be seen as part of the definitional process, a learning moment; they are not certifications of any kind." OSI states that it will keep reviewing legal documents, as it does for software licences, and will not certify individual AI systems.[3] One hypothesis the phase was designed to test, in OSI's words, was whether dropping the dataset requirement would keep Pythia, "whose dataset is legally challenged in the US", inside the definition without admitting Grok, Phi or Llama.[4]

## Endorsements

The launch list named organisations in each stakeholder group, among them the EleutherAI Institute, [Common Crawl](https://aiwiki.ai/wiki/common_crawl), LLM360, LINAGORA, the Mozilla Foundation, Mercado Libre, SUSE, Kaiyuanshe, the Eclipse Foundation, Bloomberg, the Open Infrastructure Foundation, the French government's digital directorate (DINUM), Nextcloud, the Digital Public Goods Alliance, OpenForum Europe and the open source programme office of [Carnegie Mellon University](https://aiwiki.ai/wiki/cmu). Individual endorsers included Sayash Kapoor, Arvind Narayanan and [Percy Liang](https://aiwiki.ai/wiki/percy_liang).[4] By December 2024 OSI counted more than 20 endorsing organisations and more than 100 individuals.[14]

Some endorsements were qualified. Ayah Bdeir of Mozilla said in OSI's press release that the treatment of training data "may not be a perfect place to be", and the Digital Public Goods Alliance said it would build on the definition when updating its own standard, which ended up stricter (see below).[2][37]

## Criticism

### Training data

The most common objection is that a system cannot be studied or modified in any deep sense without its training data. Before the release, julia ferraioli wrote that without data the definition guarantees little beyond using and distributing a system and building on top of it through [transfer learning](https://aiwiki.ai/wiki/transfer_learning) and [fine-tuning](https://aiwiki.ai/wiki/fine_tuning). Tom Callaway argued that the data exemption was "a loophole that makes data optional" and would devalue the term in every other context.[11]

Bradley M. Kuhn of the Software Freedom Conservancy wrote three days after the release that "the OSAID fails to require reproducibility by the public of the scientific process of building these systems". He argued that OSI should have issued recommendations and not a definition, that it had moved too fast under industry pressure, and that writers, artists and other creators of training material had been left out of the drafting.[16] The Free Software Foundation said on 22 October 2024 that its own criteria for free machine-learning applications would require the software and "the raw training data and associated scripts" to grant users the four freedoms.[17]

Others wanted more from the data clause for practical reasons. Luca Antiga, chief technology officer of [Lightning AI](https://aiwiki.ai/wiki/lightning_ai), told TechCrunch that a model can meet every requirement although its training data is not freely available, and that by not dealing with the licensing of training data OSI was leaving "a gaping hole".[15] Open Future, which took part in the Paris workshop, called the outcome "a weaker but broader standard" and "probably more pragmatic", while noting that open-data advocates saw the missing data requirement as a form of [openwashing](https://aiwiki.ai/wiki/openwashing).[21]

### Process

OSI's own board report records complaints that the co-design process "was not democratic and ultimately unfair", that voting was the wrong method and that the selection of volunteers was biased. It also notes that the Llama 2 workgroup had a voting option the other groups lacked; OSI says re-tabulating without it gave the same result.[4] Kuhn argued that anchoring the work on existing systems biased it toward accepting the practices of first movers.[16]

The dispute carried into OSI's March 2025 board election, in which Kuhn and Richard Fontana ran on a platform that included repealing the OSAID. Neither was seated: both returned amended versions of OSI's board agreement, and OSI left their votes out of the published results.[20]

### Industry and analysts

Meta, whose employees took part in the drafting, rejected the result. A spokesperson told TechCrunch: "There is no single open source AI definition, and defining it is a challenge because previous open source definitions do not encompass the complexities of today's rapidly advancing AI models." The company defended the Llama licence and its acceptable use policy as safeguards and said it was cautious about disclosing training data details.[15] According to Maffulli, Google and Microsoft had agreed to stop calling models open source when they were not fully open, and Meta had not.[15] OSI's board report says some companies felt the pace left no time to form an official position.[4]

Stephen O'Grady of the analyst firm RedMonk argued before the release that the term open source should not be stretched to cover AI at all, because a definition that needs explaining loses the clarity that made the OSD work.[11] In June 2026 he wrote that "adoption has been minimal": of 40 openly released models in RedMonk's sample, none met either open source AI tier of the G7 typology described below, whose criteria OSI says broadly follow the OSAID.[31][33]

Bruce Perens, who wrote the Open Source Definition, said in September 2025 that the new definition had not reduced [openwashing](https://aiwiki.ai/wiki/openwashing) and that "Most AI vendors have not accepted OSI as an authority, with statements by Meta being the most glaring example." In the same article Amanda Brock, chief executive of OpenUK and a former OSI director, argued that openness can only be defined per component (software, data, models, weights).[42]

### OSI's response

Maffulli told LWN in October 2024: "We have not compromised on our principles". His argument was that both camps, those wanting every component and those satisfied with weights and architecture, held legitimate but less workable positions, and that practitioners in the process had identified what is actually needed to fork a system.[11] OSI has also said that anyone may set a higher bar on top of the definition, as some standards add requirements to the OSD.[3]

## Relationship to other definitions and frameworks

### Open weights

Under the OSAID, a release of weights alone is not open source AI, even under an OSI-approved licence such as [Apache 2.0](https://aiwiki.ai/wiki/apache_license) or [MIT](https://aiwiki.ai/wiki/mit_license), because the code and data information are missing.[22] OSI uses [open weights](https://aiwiki.ai/wiki/open_weights) for that narrower category. Two separate proposals define it. Heather Meeker's Open Weights definition, on GitHub since 2023, treats weights as a different kind of artifact from source code.[22][23] The Open Weight Definition from the Open Source Alliance, a draft at version 0.3 dated 21 January 2025, adapts the ten OSD criteria to weights and says "open weight" should replace "open source" when the training data is not available.[24][25] The [open-weight licence comparison](https://aiwiki.ai/wiki/open_weight_license_comparison) covers the licences themselves.

### Model Openness Framework

The Model Openness Framework, a Linux Foundation project set out in a paper by Matt White and six co-authors first posted in March 2024, ranks releases in three classes by how many of 17 components are published under open licences: Class III (Open Model), Class II (Open Tooling) and Class I (Open Science).[26] The OSAID took its component vocabulary from the framework but uses it differently. OSI describes the definition's requirements as "a bar overlayed on the MOF range of classes": the framework grades releases, and the definition draws one line across the grades.[4][12]

### EU AI Act

The [EU AI Act](https://aiwiki.ai/wiki/eu_ai_act) has its own test. Article 53(2) exempts from some documentation duties the providers of models "released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model" whose parameters, architecture information and usage information are public, unless the model poses systemic risk.[27] Recital 102 explains the notion of a free and open-source licence.[28] Nothing in the article asks for training code or for data information in the OSAID's sense.

The European Commission's guidelines on general-purpose AI models, published on July 18, 2025 and issued in final form as Communication C(2025) 7719 on November 19, 2025, interpret the exemption without referring to OSI or the OSAID. They rule out non-commercial and research-only licences and user-scale thresholds and require that the model is not monetised. Unlike the OSAID, they accept "specific, safety-oriented terms that reasonably restrict usage in applications or domains where such use would pose a significant risk to public safety, security, or fundamental rights".[29] OSI has presented the definition to European audiences as an answer to openwashing under the Act, for example in a FOSDEM 2025 talk by its EU policy analyst Jordan Maris titled "Why Europe needs the OSAID: Openwashing and the AI act".[13][34]

### G7 typology

On 29 May 2026 the G7 digital and technology ministers approved a non-binding "Vision on AI Openness Opportunities and Shared Language", for which OSI acted as a knowledge partner.[30][31] It states that "The openness of an AI is not binary" and sets out four tiers.[30]

| Tier | Summary |
|---|---|
| Open Source AI with Open Data | Weights, deployment code, training code and full training data under an open source licence |
| Open Source AI | As above, but data information may replace training data "where sharing is legally or technically impossible" |
| Open Weights AI | Weights and deployment code under an open source licence |
| Weights Available AI | Weights and deployment code under a licence with use restrictions |

The glossary's "Training Data Information" follows the OSAID's wording closely. OSI noted that the second tier is narrower than its own definition, since it allows the data exception only for legal or technical impossibility.[30][31]

### Stricter standards and new licences

Several bodies have set a higher bar. Debian developers debated a general resolution in April 2025, proposed by Mo Zhou, under which models released without their original training data would not meet the Debian Free Software Guidelines; it was later withdrawn.[18][19] The Digital Public Goods Alliance's 2025 update to its standard requires open training data for an AI system to count as a digital public good.[37][38]

Licences are a separate matter from the definition. [OpenMDW](https://aiwiki.ai/wiki/openmdw), a permissive licence for models and their artifacts that grew out of the Model Openness Framework, does not require any particular component to be released.[39] In November 2026 OSI reported that OpenMDW and another AI-specific licence, ModelGo, were awaiting its approval.[36]

## After version 1.0

At the release Maffulli called version 1.0 "a starting point", and OSI set up a committee to follow how it was applied and propose amendments.[2][15] OSI's plan for 2025 was promotion and education.[4] In early 2025 it published, with Open Future, the white paper "Data Governance in Open Source AI", which grew out of the Paris workshop.[32]

In September 2025 OSI reported gradual movement toward the definition, citing [DeepSeek-R1](https://aiwiki.ai/wiki/deepseek_r1) and [IBM Granite](https://aiwiki.ai/wiki/ibm_granite) as models releasing more of their training code, and EleutherAI's [Common Pile](https://aiwiki.ai/wiki/common_pile) as an example of new open datasets. It also said the term was still widely used for releases of weights alone.[34]

OSI's leadership changed during this period. Maffulli's departure was announced on 15 September 2025, and Duane O'Brien started as executive director on 13 April 2026. O'Brien wrote that work on the definition should continue in the open, "grounded in community input, transparent process, and a commitment to definitions that are practical, durable, and widely understood".[43][44]

In June 2026 OSI launched an Open Source AI Fellowship with Duke University's Sanford School of Public Policy. Its announced activities include workshops on the definition's impact, "including an exploration of potential updates or revisions".[35] The first Open Source AI Salons followed in Raleigh on 21 October 2026 and Barcelona on 27 October 2026, with further meetings scheduled for Paris in November 2026 and Brussels in February 2027.[36][41] OSI's report on the first two says participants agreed that the four freedoms should be kept and that data is the bottleneck, and disagreed on whether graded levels of openness are desirable. It adds that although more models now carry OSI-approved licences, "relatively few appear to meet the OSAID's data information requirements".[36]

## See also

- [Open-source AI](https://aiwiki.ai/wiki/open_source_ai)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [OpenMDW](https://aiwiki.ai/wiki/openmdw)
- [Open-Weight LLM License Comparison](https://aiwiki.ai/wiki/open_weight_license_comparison)
- [EU AI Act](https://aiwiki.ai/wiki/eu_ai_act)

## References

1. Open Source Initiative. "The Open Source AI Definition - 1.0". opensource.org, October 2024. https://opensource.org/ai/open-source-ai-definition
2. Open Source Initiative. "The Open Source Initiative Announces the Release of the Industry's First Open Source AI Definition". opensource.org, 28 October 2024. https://opensource.org/blog/the-open-source-initiative-announces-the-release-of-the-industrys-first-open-source-ai-definition
3. Open Source Initiative. "OSAID FAQs: Answers to Frequently Asked Questions". opensource.org. https://opensource.org/ai/faq
4. Open Source Initiative. "Final Board Report" (Open Source AI Definition). opensource.org. https://opensource.org/ai/final-board-report
5. Open Source Initiative. "Open Source AI Process". opensource.org. https://opensource.org/ai/process
6. Open Source Initiative. "Drafts of the Open Source AI Definition". opensource.org. https://opensource.org/ai/drafts
7. Open Source Initiative. "The Open Source AI Definition - draft v. 0.0.9". opensource.org, 22 August 2024. https://opensource.org/ai/drafts/open-source-ai-definition-draft-v-0-0-9
8. Open Source Initiative. "The Open Source AI Definition - 1.0-RC1". opensource.org, 2 October 2024. https://opensource.org/ai/drafts/the-open-source-ai-definition-1-0-rc1
9. Open Source Initiative. "The Open Source AI Definition - 1.0-RC2". opensource.org, 21 October 2024. https://opensource.org/ai/drafts/the-open-source-ai-definition-1-0-rc2
10. Open Source Initiative. "Community input drives the new draft of the Open Source AI Definition". opensource.org, 22 August 2024. https://opensource.org/blog/community-input-drives-the-new-draft-of-the-open-source-ai-definition
11. Joe Brockmeier. "OSI readies controversial Open AI definition". LWN.net, 25 October 2024. https://lwn.net/Articles/995159/
12. Open Source Initiative. "Checklist to evaluate machine learning systems". opensource.org. https://opensource.org/ai/checklist
13. Open Source Initiative. "OSAID Timeline". opensource.org. https://opensource.org/ai/timeline
14. Nick Vidal. "2024 end-of-year review: Open Source AI Definition v1.0". Open Source Initiative, 17 December 2024. https://opensource.org/blog/2024-end-of-year-review-open-source-ai-definition-v1-0
15. Kyle Wiggers. "We finally have an 'official' definition for open source AI". TechCrunch, 28 October 2024. https://techcrunch.com/2024/10/28/we-finally-have-an-official-definition-for-open-source-ai/
16. Bradley M. Kühn. "Open Source AI Definition Erodes the Meaning of 'Open Source'". Software Freedom Conservancy, 31 October 2024. https://sfconservancy.org/blog/2024/oct/31/open-source-ai-definition-osaid-erodes-foss/
17. Free Software Foundation. "FSF is working on freedom in machine learning applications". fsf.org, 22 October 2024. https://www.fsf.org/news/fsf-is-working-on-freedom-in-machine-learning-applications
18. Debian Project. "General Resolution: Interpretation of DFSG on Artificial Intelligence (AI) Models". debian.org, 2025. https://www.debian.org/vote/2025/vote_002
19. Joe Brockmeier. "Debian debates AI models and the DFSG". LWN.net, 25 April 2025. https://lwn.net/Articles/1018497/
20. Joe Brockmeier. "OSI election ends with unsatisfying results". LWN.net, 21 March 2025. https://lwn.net/Articles/1014603/
21. Alek Tarkowski and Paul Keller. "The Open Source AI Definition is a step forward in defining openness in AI". Open Future, 1 November 2024. https://openfuture.eu/blog/the-open-source-ai-definition-is-a-step-forward-in-defining-openness-in-ai/
22. Open Source Initiative. "Open Weights: not quite what you've been told". opensource.org. https://opensource.org/ai/open-weights
23. Open Weights project. "Open Weights: A definition and license for open source compatible licensing of neural network weights". GitHub. https://github.com/Open-Weights/Definition
24. Open Source Alliance. "The Open Weight Definition", version 0.3. openweight.org, 21 January 2025. https://openweight.org
25. OpenWeight project. "Open Weight Definition (OWD)" (README, draft). GitHub. https://github.com/OpenWeight/owd
26. Matt White, Ibrahim Haddad, Cailean Osborne, Xiao-Yang Yanglet Liu, Ahmed Abdelmonsef, Sachin Varghese and Arnaud Le Hors. "The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence". arXiv:2403.13784, 20 March 2024 (revised 18 October 2024). https://arxiv.org/abs/2403.13784
27. European Union. "Article 53: Obligations for Providers of General-Purpose AI Models", Regulation (EU) 2024/1689. EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/53/
28. European Union. "Recital 102", Regulation (EU) 2024/1689. EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/recital/102/
29. European Commission. "Guidelines on the scope of obligations for providers of general-purpose AI models under the AI Act". digital-strategy.ec.europa.eu, published 18 July 2025; current text C(2025) 7719 final, 19 November 2025. https://digital-strategy.ec.europa.eu/en/library/guidelines-scope-obligations-providers-general-purpose-ai-models-under-ai-act
30. G7 Digital and Technology Ministers. "G7 Vision on AI Openness Opportunities and Shared Language". Paris, 29 May 2026. G7 Information Centre, University of Toronto. https://www.g7.utoronto.ca/ict/260529-vision.html
31. Jordan Maris. "Open Source Initiative Helps G7 Deliver Vision On AI Openness". Open Source Initiative, 1 June 2026. https://opensource.org/blog/open-source-initiative-helps-g7-deliver-vision-on-ai-openness
32. Alek Tarkowski. "Data Governance in Open Source AI". Open Source Initiative, 3 February 2025. https://opensource.org/datagovernance
33. Stephen O'Grady. "The G7 on Open Source vs Open Weights". RedMonk tecosystems, 2 June 2026. https://redmonk.com/sogrady/2026/06/02/g7-open-weights/
34. Open Source Initiative. "Report from OSS EU 2025 and AI_dev: What's next for OSAID". opensource.org, 10 September 2025. https://opensource.org/blog/report-from-oss-eu-2025-and-ai_dev-whats-next-for-osaid
35. OSI Policy Team. "Open Source AI Fellowship Announced at UN Open Source Week". Open Source Initiative, 23 June 2026. https://opensource.org/blog/open-source-ai-fellowship-announced-at-un-open-source-week
36. Gabriel Toscano. "Open Source AI Salons: Shaping the OSAID, Together". Open Source Initiative, 10 November 2026. https://opensource.org/blog/open-source-ai-salons-shaping-the-osaid-together
37. Liv Marte Nordhaug. "The Role of Open Data in AI systems as Digital Public Goods". Digital Public Goods Alliance, 29 October 2024. https://digitalpublicgoods.net/blog/the-role-of-open-data-in-ai-systems-as-digital-public-goods
38. Lea Gimpel and Ricardo Mirón Torres. "Navigating Alignment in AI Systems as DPGs". Digital Public Goods Alliance, 23 June 2026. https://www.digitalpublicgoods.net/blog/navigating-alignment-in-ai-systems-as-dpgs
39. OpenMDW project. "OpenMDW License Agreement v1.1: Overview and Purpose" (README). GitHub. https://github.com/OpenMDW/OpenMDW
40. Open Source Initiative. "The Open Source Definition", version 1.9. opensource.org. https://opensource.org/osd
41. Open Source Initiative. "Open Source AI Salons". opensource.org. https://opensource.org/ai-salon
42. Marco Fioretti. "Avoiding OSAID: Brock and Perens Reflect on a Year of Open Source AI Debate". FOSS Force, 30 September 2025. https://fossforce.com/2025/09/avoid-osaid-brock-and-perens-reflect-on-a-year-of-open-source-ai-debate/
43. Open Source Initiative. "OSI charts next phase for the organization with executive director search". opensource.org, 15 September 2025. https://opensource.org/blog/osi-next-phase-executive-director-search
44. Duane O'Brien. "Hello From The New Executive Director". Open Source Initiative, 23 April 2026. https://opensource.org/blog/hello-from-the-new-executive-director

