Citation and evidence

Openwashing

29 min full readUpdated 51 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI EthicsAI Policy & RegulationOpen Source AI

Cite this article

Openwashing (also written open-washing or open washing) is the practice of presenting a product, organisation or technology as "open" while withholding the things that openness normally implies. The word is modelled on greenwashing and was coined in 2009, well before the current wave of generative AI, to describe marketing that borrowed the vocabulary of open source and free culture.[1][2] In artificial intelligence it has been applied since 2023 to model releases that are described as "open" or "open source" although the developer publishes only part of the system, typically the trained weights under a licence that restricts use, and keeps back the training data, the training code or detailed documentation.[3][4]

The word is used as an accusation.[5] Meta, a company frequently named, has replied that no single definition of open-source AI exists, and several long-standing open-source practitioners argue that the stricter definitions on offer are unworkable or unnecessary.[6][7] The dispute has practical stakes because the EU AI Act relieves providers of some free and open-source models of certain documentation duties, so the wording of an "open" claim can have legal consequences.[8][9]

Origin of the term

Michelle Thorne, a free-culture advocate who has worked for Creative Commons and Mozilla, introduced the word in a blog post dated 14 March 2009.[1][10] She defined it in one line, "Openwashing: to spin a product or company as open, although it is not", and said it was derived from greenwashing.[1] Her example was not software. It was an advertising campaign by Berlin Partner, a promotional body for the German capital, whose slogan was "be open. be free. be Berlin." and whose terms of use barred visitors from downloading, copying or modifying the content of the campaign website.[1] Thorne did not treat the trend as wholly bad: she called openwashing a side effect of customers wanting transparency and access, and wrote that companies resorting to it were indirectly giving weight to those issues.[1]

The term then spread through education technology and enterprise software. The writer Audrey Watters posted a definition on Twitter in about 2012 that she later called one of her most popular tweets: "Openwashing: n., having an appearance of open-source and open-licensing for marketing purposes, while continuing proprietary practices."[11] Her target was the learning-management vendor Blackboard, which had bought two companies that supported the open-source Moodle platform.[11] In March 2012 Scott Crenshaw, head of the cloud business unit at Red Hat, said that virtualisation vendors such as VMware and Microsoft were calling closed cloud products open as a marketing gimmick, a complaint the trade press reported as a warning about cloud "open-washing".[12] The word also entered academic writing: in 2017 Maximilian Heimstädt published "Openwashing: A decoupling perspective on organizational transparency" in the journal Technological Forecasting and Social Change.[13]

DateEvent
14 March 2009Michelle Thorne's blog post defines "openwashing"[1]
March 2012Red Hat accuses cloud vendors of using "open" as a marketing gimmick[12]
July 2023The Open Source Initiative objects to Meta calling Llama 2 open source; "Opening up ChatGPT" is presented at CUI '23[14][15]
August 2023Widder, West and Whittaker post the preprint "Open (For Business)"[16]
April 2024LF AI & Data describes an "open-washing" trend when introducing the Model Openness Framework[17]
June 2024Liesenfeld and Dingemanse present "Rethinking open source generative AI: open-washing and the EU AI Act" at FAccT '24[3]
28 October 2024The Open Source Initiative releases version 1.0 of the Open Source AI Definition[18]
27 November 2024Nature publishes "Why 'open' AI systems are actually closed, and why this matters"[4]
24 February 2025The European Open Source AI Index is introduced[19]
18 July 2025The European Commission publishes guidelines that set out when a general-purpose AI model counts as free and open-source under the AI Act[9]

Expanded article table

Use in artificial intelligence

The term moved into AI after a series of large "open" language models appeared in 2023. In July of that year the Open Source Initiative (OSI) objected to Meta's description of Llama 2 as open source, and three researchers at Radboud University in the Netherlands published the first of two surveys grading how open such systems were.[14][15] In August 2023 David Gray Widder, Sarah West and Meredith Whittaker posted the preprint "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI", which Liesenfeld and Dingemanse later cited as their source for the term.[16][3]

By 2024 the word was in general use. In April the Linux Foundation's LF AI & Data group wrote that models presented as open source often used bespoke licences with ambiguous terms and lacked datasets, training code and benchmarks, and that this "open-washing" trend "threatens to undermine the very premise of openness".[17] In May The New York Times featured the word in its column on business jargon, glossing it as an accusation that some AI companies use the open-source label too loosely, and noting that the same charge had earlier been aimed at coding projects.[5] It was still current in 2026: a commentary in Tech Policy Press on 25 July, two days before Moonshot AI published the weights of Kimi K3, argued that describing the model as open source was a category error, because even with the weights released its training data and training pipeline would stay closed.[20] Moonshot's model card, published with the weights on 27 July 2026, calls K3 an "open-weight" model and places it under the company's own Kimi K3 License.[48]

Writers do not use the word in exactly the same way. Liesenfeld and Dingemanse apply it to models that are "open in weights only", with the weights under an open licence and most other aspects of the system kept back.[3] Widder, Whittaker and West apply it to systems that offer little more than an API or a downloadable model under use restrictions.[4] The OSI uses it for calling a model open source when its licence fails the Open Source Definition.[21] The European Open Source AI Index has described it as projecting an image of transparency without doing the work needed to count as meaningfully open, and adds a regulatory sense: overclaiming openness, for example by attaching an open-source licence to a single component, in a way that risks unintended claims to exemptions under EU law.[22][23]

Academic literature

Liesenfeld and Dingemanse

"Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators", by Andreas Liesenfeld, Alianda Lopez and Mark Dingemanse, appeared in the proceedings of the 5th International Conference on Conversational User Interfaces (CUI '23, Eindhoven, July 2023).[15] It graded ChatGPT alternatives on code, training data, model weights, data from reinforcement learning from human feedback, licensing, scientific documentation and access methods, and found that many projects billing themselves as open source inherited undocumented data, that few shared their instruction-tuning data and that careful scientific documentation was very rare.[15] The authors later described the accompanying website as the first global openness index for instruction-tuned large language models and said it grew to more than 50 models from over 25 providers.[19]

The follow-up paper, "Rethinking open source generative AI: open-washing and the EU AI Act", was presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24) in Rio de Janeiro in June 2024.[3] It assessed 40 text generators and 6 text-to-image systems that were marketed as open, with ChatGPT as a closed reference point, against 14 dimensions.[3] The authors reported that "many models are 'open weight' at best", that roughly the bottom third of the text generators shared weights and little else, and that the models of the large commercial developers (Meta, Google, Cohere, Microsoft and Mistral AI) sat in the lower ranks, while OLMo Instruct from the Allen Institute for AI, BloomZ from the BigScience workshop and AmberChat from LLM360 came close to full openness.[3]

The paper also described a "release by blogpost" pattern: a model is announced in a corporate post that calls it open and shows selected benchmark tables, without a peer-reviewed paper or full technical documentation.[3] Its central argument is that openness in generative AI is composite (made of several elements) and gradient (a matter of degree), and that declaring a model open on one feature, such as downloadable weights or a permissive licence, is among the most effective methods of open-washing.[3] The authors compared the risk to fair-trade labelling, which they said large companies had adopted and diluted.[3] Nature reported the study in its news pages under the headline "Not all 'open source' AI models are actually open: here's a ranking".[24]

Widder, Whittaker and West

After their 2023 preprint, the same three authors published the Perspective "Why 'open' AI systems are actually closed, and why this matters" in Nature on 27 November 2024 (volume 635, pages 827 to 833). The authors are listed as David Gray Widder of Cornell University's Digital Life Initiative, Meredith Whittaker of the Signal Foundation and the University of Western Australia, and Sarah Myers West of the AI Now Institute.[4] The paper examines what openness can and cannot provide across models, data, labour, development frameworks and computing power, and identifies three things that open AI does offer: transparency, reusability and extensibility.[4]

Its conclusion is broader than a complaint about labels. The authors write that "openness alone does not perturb the concentration of power in AI", because the data, compute and distribution needed to build and deploy large systems remain in the hands of a few companies, and that rhetoric about open AI is often used in ways that increase that concentration.[4] On terminology they say that some systems described as open, citing Meta's Llama 3, offer little more than an API or a model that can be downloaded subject to use restrictions, and they call this "'openwashing' systems that are better understood as closed"; as a contrast they cite EleutherAI's Pythia models, released with source code, training data, documentation and licence terms aligned with the OSI's definition.[4]

Dimensions along which openness is graded

Allegations of openwashing usually turn on which parts of a system are released and on what terms. The framework used in the FAccT paper and in the two Radboud indexes grades each of 14 dimensions as open, partial or closed. The paper's appendix groups them as follows.[3]

GroupDimensionQuestion asked
AvailabilityOpen codeIs the source code for data processing and training available?
AvailabilityBase model dataCan the training datasets of the base model be inspected?
AvailabilityBase model weightsAre the weights of the base model shared?
AvailabilityInstruction-tuning dataAre the datasets used for instruction tuning available?
AvailabilityInstruction-tuning weightsAre the weights of the final tuned model shared?
DocumentationCodeIs the code documented well enough to replicate or modify?
DocumentationArchitectureAre the architecture, training setup and evaluation described?
DocumentationPreprintIs there an archived preprint covering the system?
DocumentationPaperHas a description passed peer review?
DocumentationModel cardIs there a model card covering architecture, training and evaluation?
DocumentationDatasheetIs data collection and curation documented in a datasheet?
AccessPackageIs there an indexed, versioned software package?
AccessAPIIs there an API, and is access to it unrestricted?
AccessLicenceIs the system fully covered by an OSI-approved or responsible-AI (RAIL) licence?

Expanded article table

A variant for text-to-image systems replaces two of these with watermarking and prompt moderation.[3] Other schemes divide the system differently. The Open Source AI Definition names three required elements (data information, code and parameters), and the Model Openness Framework checks 17 components under the headings code, data and documentation.[25][26] The article on open-source AI compares these schemes in more detail.

Commonly cited examples

Meta's Llama licences

Meta's Llama family is a frequently cited case; Liesenfeld and Dingemanse wrote that probably the strongest claims to the open-source label came from Meta and its Llama 2 and Llama 3 models.[3] Meta introduced Llama 2 in July 2023 as "the next generation of our open source large language model" and used almost the same words for Llama 3.[3] In a letter published on 23 July 2024, Mark Zuckerberg called Llama 3.1 405B "the first frontier-level open source AI model".[27]

On 20 July 2023 the OSI's executive director, Stefano Maffulli, published a statement that Llama 2 was not open source. It said Meta was confusing "open source" with "resources available to some users under some conditions", that the licence restricted commercial use for some users and, through its acceptable use policy, restricted use for certain purposes, and that the Open Source Definition forbids discrimination against persons, groups or fields of endeavour.[14] The same statement welcomed Meta's lowering of barriers to access and said the OSI did not question the company's wish to limit competitive use of Llama, only the label.[14] The commercial clause requires services with more than 700 million monthly active users to request a licence from Meta.[18]

In February 2025 the OSI repeated the objection for the Llama 3.x licences, citing the Free Software Foundation's assessment of the Llama 3.1 licence and adding that newer versions excluded people in the European Union.[21] The acceptable use policy for Llama 3.2 withholds the licence rights for its multimodal models from individuals domiciled in the EU and companies with their principal place of business there, with an exception for end users of products that incorporate the models.[28] The OSI post called Meta's use of the term "open washing" and said that this conduct, and the OSI's response to it, predated the Open Source AI Definition.[21]

When the OSI published its AI definition in October 2024, Meta spokesperson Faith Eischen told The Verge: "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."[6] A Meta spokesperson told TechCrunch that the company makes Llama free and openly available, that the licence and acceptable use policy "help keep people safe by having some restrictions in place", and that Meta was taking a cautious approach to sharing details of training data while regulation evolved; the spokesperson pointed to other attempts to define the term by the Linux Foundation and the Free Software Foundation.[18] The Verge reported that Meta cites safety concerns for restricting access to its training data, while critics see a motive in limiting legal liability and protecting a competitive advantage.[6]

In May 2025 ITPro reported fresh accusations of open washing after Meta sponsored a Linux Foundation study of open-source AI adoption. Amanda Brock, chief executive of OpenUK, said that Llama was not open source under any definition, citing the commercial restriction in its licence, while commending Meta's work on Llama as a "step in the right direction" for awareness of open source.[29]

Other cases

The releases below were assessed in the same sources. Not every developer listed was accused of openwashing by name, and Google avoided the term "open source" for Gemma.

Developer and releaseCriticismDeveloper's description or response
Mistral AI, Mistral 3 (December 2025)The European Open Source AI Index wrote that Mistral released weights but not training data and, by conflating "open-weight" with "open-source", engaged in open-washing[30]Mistral released every Mistral 3 model under the Apache 2.0 licence. The announcement used both terms: Mistral Large 3 "joins the ranks of frontier instruction-fine-tuned open-source models" and is "one of the best permissive open weight models in the world"[31]
DeepSeek models, including DeepSeek R1A post dated 25 January 2025 on the index's website, a month before the index was formally introduced, called them open weights at best: more open than some large rivals, with pre-training and instruction-tuning data undisclosed[22]DeepSeek's repository says the company has "open-sourced" the models and places code and weights under the MIT License[32]
Google, Gemma (February 2024)Gemma 7B Instruct is listed in the Radboud table, and the FAccT paper counted Google among the commercial developers in its lower ranks[33][3]Google called Gemma "open models", said such models' terms of use "may not be based on an open-source license", and wrote: "We're precise about the language we're using to describe Gemma models"[34]. Gemma 4, announced on 2 April 2026, was released under the Apache 2.0 licence[49]
Microsoft, Phi modelsNature reported the FAccT authors' view that companies such as Meta and Microsoft describe models as open source while withholding important information about them, and the paper placed Microsoft among the commercial developers in its lower ranks; the OSI's validation exercise separately found that Phi-2 did not meet its AI definition[24][3][35]According to Maffulli, Google and Microsoft agreed after discussions with the OSI to stop using "open source" for models that are not fully open[18]

Expanded article table

The OSI describes its review of models during the drafting of its definition as a learning exercise and not a certification. In that review Pythia, OLMo, Amber, CrystalCoder and T5 passed; Llama 2, Grok, Phi-2 and Mixtral did not; and the OSI said that BLOOM, StarCoder2 and Falcon would probably pass if their licence terms were changed.[35]

The AI Act (Regulation (EU) 2024/1689) gives lighter treatment to free and open-source releases in several places. Article 2(12) takes AI systems released under such licences out of the regulation's scope unless they are placed on the market or put into service as high-risk systems or fall under Article 5 or Article 50.[8] For general-purpose AI models, Article 53(2) exempts providers from two duties, drawing up technical documentation for the authorities and providing documentation to downstream developers, if the model is released under a free and open-source licence that allows access, use, modification and distribution, and if its parameters (including the weights), information on its architecture and information on its usage are public. Article 54(6) similarly waives the duty to appoint an authorised representative in the EU. Neither exception applies to models with systemic risk, and the duties to keep a copyright compliance policy and to publish a summary of training content remain for every provider.[8][9]

Liesenfeld and Dingemanse, writing about a draft of the Act in early 2024, argued that such an exemption makes open-source status attractive to providers as a way to avoid documentation and scrutiny, and that a test based on the licence alone would reward releasing only the weights.[3] Brock made a similar point in 2025, saying that because regulators such as the EU use the term as the basis of exceptions, "the impact of open washing has become a societal one".[29]

The European Commission's guidelines on the obligations of general-purpose AI model providers, published on 18 July 2025, set out how the Commission reads these conditions. They are not legally binding, and the Commission notes that only the Court of Justice of the European Union can give an authoritative interpretation, but they state the reading on which the Commission will base enforcement.[9][36] The conditions are set out in the table below.[9]

ConditionThe Commission's reading in the 2025 guidelines
Licence rightsThe licence must grant access, use, modification and redistribution. If any one of the four is missing, the exemption is lost.
Disqualifying termsExamples given: limits to non-commercial or research use, bans on redistribution, "usage restrictions triggered by user scale thresholds", and a requirement to obtain a separate commercial licence for some uses.
Permitted termsAttribution, share-alike distribution terms, and proportionate, non-discriminatory safety restrictions on uses that pose a significant risk to public safety, security or fundamental rights.
AccessAnyone must be able to obtain the model without payment. User verification is allowed if it does not discriminate unfairly, for example by country of origin.
No monetisationDual licensing with paid commercial use, mandatory paid support, paid or advertising-funded exclusive hosting, and access conditioned on processing personal data all count as monetisation. Optional paid services do not.
Public informationWeights, architecture information and usage information must be published in a form that enables access, use, modification and distribution.
Systemic riskThe exemption never applies to a model with systemic risk, which the Act presumes when cumulative training compute exceeds 10^25 floating-point operations.

Expanded article table

The guidelines name no company or licence. They differ from the OSI's Open Source Definition, which does not let a licence discriminate against fields of endeavour, in accepting narrowly drawn safety restrictions.[14][9] According to the Commission, the obligations for providers of general-purpose AI models applied from 2 August 2025, its enforcement powers from 2 August 2026, and models already on the market before August 2025 must comply by 2 August 2027.[37]

Responses

Definitions and frameworks

The OSI released version 1.0 of the Open Source AI Definition on 28 October 2024, and it was still the current version in November 2026.[18][25] It requires the freedoms to use, study, modify and share a system, and access to the preferred form for modifying it: the complete training and inference code, the model parameters, and "data information" detailed enough for a skilled person to build a substantially equivalent system.[25] The OSI has no means of enforcing it. Maffulli told TechCrunch he hoped the AI community would refuse to recognise misuse of the term, and the developer Simon Willison told The Verge that the definition might make it easier to push back against companies that are open washing.[18][6] Alek Tarkowski and Paul Keller of the think tank Open Future wrote that the definition's requirements are weaker than mandatory release of training data but make cases of corporate openwashing easy to identify.[38]

The Model Openness Framework, set out in a paper by Matt White and six co-authors first posted in March 2024 and developed in the Generative AI Commons group of LF AI & Data, ranks a release in one of three classes by which components are published under open licences: Class III (open model), Class II (open tooling) and Class I (open science).[39][17][26] A release whose terms are too restrictive receives no class. Arnaud Le Hors presented the companion Model Openness Tool at Open Source Summit North America in 2026, in a talk LWN reported under the headline "MOT: a tool to fight openwashing in AI".[26] Le Hors noted that a licence written to cover a model and all its artifacts, OpenMDW, existed but was not yet in general use at the time of the talk in May 2026.[26] By 17 November 2026 the Hugging Face Hub listed about 500 model repositories under the licence's 1.1 tag.[50]

Governments have taken up the term as well. On 29 May 2026 the G7 digital and technology ministers approved a non-binding "Vision on AI Openness Opportunities and Shared Language", which names avoiding open washing among the purposes of a shared vocabulary and sets out four labels: Open Source AI with Open Data, Open Source AI, Open Weights AI and Weights Available AI.[51]

Openness indexes

The Opening up ChatGPT table has been superseded by the European Open Source AI Index, introduced on 24 February 2025. The index is part of Radboud University's Centre for Language and Speech Technology, was co-founded by Liesenfeld and Dingemanse, and uses the same 14 dimensions.[33][19][23] Anyone can submit a judgement with supporting evidence, such as official documentation, a preprint or source code, and a reviewer checks it before publication; as of November 2026 the website showed a database generated on 27 August 2026.[3][23] Its news posts apply the term directly to individual releases, as with Mistral 3, and they also record highly open releases such as the Swiss Apertus models.[30]

The "open-weight" label

A further response is to drop the contested phrase for releases that consist mainly of parameters. The OSI's explainer on open weights, published in January 2025, calls them a milestone that still stops short of open-source AI because training code and data transparency are missing, and notes that the licensing lawyer Heather Meeker published an Open Weights Definition in 2023.[40] OpenAI described its 2025 gpt-oss release as "open-weight models", and Google chose "open models" for Gemma.[41][34] OpenUK's July 2026 review of the three years since Llama 2 reported that industry language had shifted from "open source" to "open weight", and that the accusation of open washing against Meta had reached the mainstream press.[42] The label, its history and the licences involved are covered in the article on open weights and in the open-weight licence comparison.

Criticism of the term and counter-arguments

One objection is that the word presupposes an agreed meaning of "open". Introducing a 2014 issue of the Journal of Interactive Technology and Pedagogy, Suzanne Tamang and Gregory T. Donovan summarised a 2013 critique by Evgeny Morozov: talk of openwashing helps to question whether an open initiative is authentic but does little to say what openness should mean.[43] Watters reached a related view about her own definition. By 2014 she no longer thought the answer was a clearer definition of "open" or stricter adherence to a particular licence, and called instead for more transparency about the politics behind open projects.[11] Meta's position that there is no single definition of open-source AI is a version of the same objection.[6]

A second objection is that a strict definition is impractical. Ayah Bdeir, a Mozilla adviser who took part in the OSI process, told MIT Technology Review that requiring disclosure of all training sets would be difficult to enforce because of copyright and data ownership.[44] Nathan Lambert, writing in 2024 about his own views and the strategy of the Allen Institute for AI, said that a workable definition cannot require developers to commit potentially unlawful acts with data, that the OSI's text tried to sit between fully documented models and what developers "can grab off the shelf and use easily", and that Llama would remain popular whether or not it qualified.[45] Brock, speaking for herself and not for OpenUK, called it a mistake to define open-source AI at all, comparing AI to cloud and mobile as a category of products; she prefers to assess each component (software, data, models, weights) under existing definitions and licences.[46][7]

A third is that weights already deliver much of the practical benefit. OSI board chair Tracy Hinds acknowledged that some groups "believe that model parameters and architecture are enough to study and modify", a view the OSI did not adopt.[46] A 2024 position paper by Sayash Kapoor, Rishi Bommasani and 23 co-authors, among them Percy Liang, Arvind Narayanan and Maffulli, defined open foundation models simply as those with broadly available weights and found that they bring benefits, with some caveats, for innovation, competition, the distribution of decision-making power and transparency.[47] Zuckerberg's 2024 letter made its case for Llama in terms of what organisations can do once they hold the weights: train, fine-tune and distil their own models, avoid dependence on a closed vendor and keep sensitive data on their own infrastructure.[27]

The stricter definition is also criticised from the other side, as too weak or as ineffective. Luca Antiga of Lightning AI said that by not dealing with the licensing of training data the OSI was "leaving a gaping hole" for businesses deciding whether a model can be adopted.[18] Sam Johnston wrote that an ineffective and unimplementable definition risks open-source developers being crowded out by opaque "open-washed offerings from commercial vendors".[46] Bruce Perens said in 2025 that the problem a year after the AI definition was still openwashing and that most AI vendors had not accepted the OSI as an authority.[7] OpenUK's 2026 review said the definition had not gained support and described approaches that grade a system component by component, such as Radboud's and the Linux Foundation's, as likely the better route.[42]

Continue reading

More to read

Topics:AI EthicsAI Policy & RegulationOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6Michelle Thorne. "Openwashing". michellethorne.cc, 14 March 2009 (archived copy). web.archive.org/...openwashing
  2. ^Steven J. Vaughan-Nichols. "The open secret of open washing". The Register, 25 October 2024. theregister.com/...975125
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18Andreas Liesenfeld and Mark Dingemanse. "Rethinking open source generative AI: open-washing and the EU AI Act". Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24), pp. 1774-1787, June 2024. facctconference.org/...facct24-120.pdf
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7David Gray Widder, Meredith Whittaker and Sarah Myers West. "Why 'open' AI systems are actually closed, and why this matters". Nature 635, 827-833, 27 November 2024. nature.com/...s41586-024-08141-1
  5. ^1 ^2Sarah Kessler. "Some A.I. Companies Face a New Accusation: 'Open Washing'". The New York Times, 17 May 2024. nytimes.com/...what-is-openwashing-ai
  6. ^1 ^2 ^3 ^4 ^5Kylie Robison. "Open-source AI must reveal its training data, per new OSI definition". The Verge, 28 October 2024. theverge.com/...artificial-intelligence-meta-llama
  7. ^1 ^2 ^3Marco Fioretti. "Avoiding OSAID: Brock and Perens Reflect on a Year of Open Source AI Debate". FOSS Force, 30 September 2025. fossforce.com/...n-a-year-of-open-source-ai-debate
  8. ^1 ^2 ^3European Parliament and Council of the European Union. "Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)". Official Journal of the European Union, 12 July 2024. eur-lex.europa.eu/...eng
  9. ^1 ^2 ^3 ^4 ^5 ^6European Commission. "Guidelines on the scope of the obligations for providers of general-purpose AI models established by Regulation (EU) 2024/1689 (AI Act)", section 4. Published 18 July 2025; text C(2025) 7719 final dated 19 November 2025. digital-strategy.ec.europa.eu/...dels-under-ai-act
  10. ^Michelle Thorne. "About". michellethorne.cc (archived copy, 18 October 2014). web.archive.org/...about
  11. ^1 ^2 ^3Audrey Watters. "From 'Open' to Justice #OpenCon2014". Hack Education, 16 November 2014. hackeducation.com/...from-open-to-justice
  12. ^1 ^2Sophie Curtis. "Beware of cloud 'open-washing' says Red Hat". IT Business, 30 March 2012 (archived copy). web.archive.org/...17513
  13. ^Maximilian Heimstädt. "Openwashing: A decoupling perspective on organizational transparency". Technological Forecasting and Social Change 125, 77-86, December 2017. doi.org/...j.techfore.2017.03.037
  14. ^1 ^2 ^3 ^4 ^5Stefano Maffulli. "Meta's LLaMa license is not Open Source". Open Source Initiative, 20 July 2023. opensource.org/...ama-2-license-is-not-open-source
  15. ^1 ^2 ^3 ^4Andreas Liesenfeld, Alianda Lopez and Mark Dingemanse. "Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators". Proceedings of the 5th International Conference on Conversational User Interfaces (CUI '23), July 2023. arxiv.org/...2307.05532
  16. ^1 ^2David Gray Widder, Sarah West and Meredith Whittaker. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI". SSRN, August 2023. doi.org/...ssrn.4543807
  17. ^1 ^2 ^3LF AI & Data Foundation. "Introducing the Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency and Usability in AI". 17 April 2024. lfaidata.foundation/...parency-and-usability-in-ai
  18. ^1 ^2 ^3 ^4 ^5 ^6 ^7Kyle Wiggers. "We finally have an 'official' definition for open source AI". TechCrunch, 28 October 2024. techcrunch.com/...al-definition-for-open-source-ai
  19. ^1 ^2 ^3Mark Dingemanse and Andreas Liesenfeld. "Introducing the European Open Source AI Index". European Open Source AI Index, 24 February 2025. osai-index.eu/...introducing-eu-osai-index
  20. ^JJ Jasser. "'Open-Washing' Is Everywhere in AI. Four Criteria Cut Through It". Tech Policy Press, 25 July 2026. techpolicy.press/...i-four-criteria-cut-through-it
  21. ^1 ^2 ^3Jordan Maris. "Meta's LLaMa license is still not Open Source". Open Source Initiative, 18 February 2025. opensource.org/...license-is-still-not-open-source
  22. ^1 ^2Dick Blankvoort and Mark Dingemanse. "How Open is DeepSeek?". European Open Source AI Index, 25 January 2025. osai-index.eu/...deepseek
  23. ^1 ^2 ^3European Open Source AI Index. "About". Radboud University, accessed 17 November 2026. osai-index.eu/about
  24. ^1 ^2Elizabeth Gibney. "Not all 'open source' AI models are actually open: here's a ranking". Nature (news), 19 June 2024. nature.com/...d41586-024-02012-5
  25. ^1 ^2 ^3Open Source Initiative. "The Open Source AI Definition 1.0". Accessed 17 November 2026. opensource.org/...open-source-ai-definition
  26. ^1 ^2 ^3 ^4Joe Brockmeier. "MOT: a tool to fight openwashing in AI". LWN.net, 27 May 2026. lwn.net/...1073420
  27. ^1 ^2Mark Zuckerberg. "Open Source AI is the Path Forward". Meta Newsroom, 23 July 2024. about.fb.com/...open-source-ai-is-the-path-forward
  28. ^Meta. "Llama 3.2 Acceptable Use Policy". Accessed 17 November 2026. llama.com/...use-policy
  29. ^1 ^2Ross Kelly. "Meta faces new 'open washing' accusations with AI whitepaper". ITPro, 30 May 2025. itpro.com/...meta-open-source-ai-linux-foundation
  30. ^1 ^2Andreas Liesenfeld. "Mistral 3: A Heavyweight Model, Lightweight on Openness". European Open Source AI Index, 2 December 2025. osai-index.eu/...mistral-3-large
  31. ^Mistral AI. "Introducing Mistral 3". 2 December 2025. mistral.ai/...mistral-3
  32. ^DeepSeek. "DeepSeek-R1" (repository README). GitHub, accessed 17 November 2026. github.com/...DeepSeek-R1
  33. ^1 ^2Andreas Liesenfeld, Alianda Lopez and Mark Dingemanse. "Opening up ChatGPT: tracking openness of instruction-tuned LLMs" (archived leaderboard). Accessed 17 November 2026. opening-up-chatgpt.github.io
  34. ^1 ^2Google. "Building Open Models Responsibly in the Gemini Era". Google Open Source Blog, 21 February 2024. opensource.googleblog.com/...esponsibly-gemini-era
  35. ^1 ^2Open Source Initiative. "Open Source AI Definition: frequently asked questions". Accessed 17 November 2026. opensource.org/...faq
  36. ^European Commission. "Guidelines on obligations for General-Purpose AI providers" (questions and answers). Last updated 11 November 2025. digital-strategy.ec.europa.eu/...pose-ai-providers
  37. ^European Commission. "Guidelines for providers of general-purpose AI models". Shaping Europe's digital future, accessed 17 November 2026. digital-strategy.ec.europa.eu/...es-gpai-providers
  38. ^Alek Tarkowski and Paul Keller. "The Open Source AI Definition is a step forward in defining openness in AI". Open Future, 1 November 2024. openfuture.eu/...orward-in-defining-openness-in-ai
  39. ^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. arxiv.org/...2403.13784
  40. ^Open Source Initiative. "Open Weights: not quite what you've been told". 29 January 2025. opensource.org/...open-weights
  41. ^OpenAI. "gpt-oss" (repository README). GitHub, accessed 17 November 2026. github.com/...gpt-oss
  42. ^1 ^2OpenUK. "AI Openness: 3 Years on from Llama 2". 23 July 2026. openuk.uk/...aiopenness2026
  43. ^Suzanne Tamang and Gregory T. Donovan. "Introduction: Media and Methods for Opening Education". The Journal of Interactive Technology and Pedagogy, 9 June 2014. jitp.commons.gc.cuny.edu/...-for-opening-education
  44. ^Rhiannon Williams and James O'Donnell. "We finally have a definition for open-source AI". MIT Technology Review, 22 August 2024. technologyreview.com/...inition-for-open-source-ai
  45. ^Nathan Lambert. "On the current definition of open-source AI and the state of the data commons". Interconnects, 28 August 2024. interconnects.ai/...defining-open-source-ai
  46. ^1 ^2 ^3Richard Gall. "The Open Source AI Definition: What the Critics Say". The New Stack, 18 November 2024. thenewstack.io/...-definition-what-the-critics-say
  47. ^Sayash Kapoor, Rishi Bommasani and 23 others. "On the Societal Impact of Open Foundation Models". arXiv:2403.07918, 2024. arxiv.org/...2403.07918
  48. ^Moonshot AI. "Kimi K3" (model card). Hugging Face, 27 July 2026. huggingface.co/...Kimi-K3
  49. ^Nia Castelly, amanda casari and Olivier Lacombe. "Gemma 4: Expanding the Gemmaverse with Apache 2.0". Google Open Source Blog. opensource.googleblog.com/...averse-with-apache-20
  50. ^Hugging Face. "Models" (listing filtered by licence tag openmdw-1.1; count read on 17 November 2026). huggingface.co/models
  51. ^G7 Digital and Technology Ministers. "G7 Vision on AI Openness Opportunities and Shared Language". Paris, 29 May 2026. G7 Information Centre, University of Toronto. g7.utoronto.ca/...260529-vision

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 5,718 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check on November 17, 2026: 152 claims checked against 59 sources. Eleven corrections applied before publication, mostly attribution of named examples. A 2026 amendment to the EU AI Act was not read in the original.

Cite this page: AI Wiki. "Openwashing." aiwiki.ai, updated 17 Nov 2026, fact-checked 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/openwashing

Edit article