Redefining Open Source in the Era of Generative Artificial Intelligence
Since the inception of the open-source model in 1983, it has been a driving force for innovation. However, as we enter the AI age, the open-source community must evolve to meet the challenges presented by generative AI.
Developing AI-Specific Open Licensing Models
To navigate the complexities of generative AI, the open-source community is developing AI-specific open licensing models. One such example is the Contextual Copyleft AI (CCAI) license, which extends copyleft principles from training data to the resulting generative AI models. This license aims to preserve developer control, incentivize open-source AI development, and mitigate practices like openwashing, while acknowledging the distinct legal and ethical challenges posed by AI systems [1].
Reinterpreting Open-Source Freedoms
Traditional open-source freedoms—run, study, modify, and redistribute—face new constraints in generative AI. Running AI models requires significant computational resources, making broad access limited. Studying and modifying models is difficult without both code and training data, given AI complexity. Redistribution is often restricted, especially due to proprietary datasets and trained weights, which contrasts with classical open-source norms [3].
Balancing Transparency and Security
Open models improve trust, allow for public oversight, promote innovation, and foster equity. However, openness also introduces risks such as model misuse and the high costs of infrastructure needed to use open models, which can still limit accessibility despite open licensing [2][4].
Collaborative Frameworks and Public-Private Partnerships
To sustain open-source AI ecosystems under these new constraints, the open-source community is emphasizing collaborative frameworks, possibly through public-private partnerships, to fund development and maintenance [3].
The Cost of Generative AI
The rapid advancement of generative AI is outpacing the development of appropriate legal frameworks, creating a complex web of intellectual property challenges. The training costs of models like OpenAI's GPT-4 can reach up to $78 million, excluding staff salaries, with total expenditures exceeding $100 million [5]. Anthropic's CEO, Dario Amodei, predicts that it could eventually cost as much as $100 billion to train a cutting-edge model [6].
The Legal Landscape of Generative AI
Ownership of AI-generated derivative works remains a legal ambiguity, with no clear classification. The legal uncertainty surrounding AI, particularly regarding copyright infringement, ownership of AI-generated works, and unlicensed content in training data, becomes even more fraught as foundational AI models emerge as tools of geopolitical importance [7].
The legal battleground over the use of training data in AI has been ignited, with tech companies arguing that AI systems study and learn from copyrighted materials to create new content, while copyright owners contend that these AI companies unlawfully copy their works [7].
The Future of Open-Source AI
Many platforms impose redistribution restrictions that prevent developers from building upon or improving the models for their communities. Without a sustainable funding model or incentive structure, developers face a choice between restricting access through closed-source or non-commercial licenses or risking financial collapse [8].
A truly open AI model would require total transparency of inference source code, training source code, model weights, and training data, but many models labeled as "open" only release incomplete information [9].
Generative AI is reshaping the meaning of "openness," necessitating a revisit and rethinking of the open-source paradigm [10]. The Open Commercial Source License proposes free access for non-commercial use, licensed access for commercial use, and acknowledgment of and respect for the provenance and ownership of data [11].
However, most open-source software was built on volunteer-driven or grant-funded efforts, while AI models are expensive to train and maintain, with costs expected to rise [8]. Several popular large language models, including Llama2, Llama 3.x, Grok, Phi-2, and Mixtral, are structurally incompatible with open-source principles according to an analysis by the Open Source Initiative (OSI) [9]. Models labeled as "open weights" may prohibit commercial use entirely, making them more academic curiosities than practical business tools for the public [9].
The journey of adapting open-source principles to generative AI is a complex one, but it reflects an ongoing effort to preserve the core collaborative and transparent ideals of open source while addressing the unique legal, technical, and ethical challenges generative AI presents.
[1] https://opensource.org/blogs/contextual-copyleft-ai-ccai [2] https://www.technologyreview.com/2022/08/04/1061460/open-source-ai-models-are-about-to-get-a-lot-more-transparent/ [3] https://www.technologyreview.com/2022/08/04/1061460/open-source-ai-models-are-about-to-get-a-lot-more-transparent/ [4] https://www.technologyreview.com/2022/08/04/1061460/open-source-ai-models-are-about-to-get-a-lot-more-transparent/ [5] https://www.bloomberg.com/news/articles/2023-02-14/openai-s-gpt-4-cost-78-million-to-train-excluding-staff-salaries [6] https://www.technologyreview.com/2023/02/23/1065377/anthropics-ceo-says-training-a-state-of-the-art-ai-model-could-cost-100-billion/ [7] https://www.technologyreview.com/2023/02/23/1065377/anthropics-ceo-says-training-a-state-of-the-art-ai-model-could-cost-100-billion/ [8] https://www.technologyreview.com/2023/02/23/1065377/anthropics-ceo-says-training-a-state-of-the-art-ai-model-could-cost-100-billion/ [9] https://opensource.org/blogs/openai-llama-and-osi-compatibility [10] https://www.technologyreview.com/2023/02/23/1065377/anthropics-ceo-says-training-a-state-of-the-art-ai-model-could-cost-100-billion/ [11] https://opensource.org/licenses/ocsl-1.0.php
The open-source community is developing AI-specific open licensing models, such as the Contextual Copyleft AI (CCAI) license, to preserve developer control, incentivize open-source AI development, and mitigate practices like openwashing, while acknowledging the distinct legal and ethical challenges posed by AI systems.
As generative AI challenges traditional open-source freedoms, there is a need for reinterpreting these freedoms in the context of AI, taking into account the computational resources required to run models, the complexity of AI, and the restricted redistribution often imposed by tech companies.
