Page 40 - The Annual AI Governance Report 2025 Steering the Future of AI

The Annual AI Governance Report 2025: Steering the Future of AI

training datasets or code. A wide range of actors—including major AI labs, open science
172
consortia, and civil society organisations—are involved in shaping norms and practices around
openness.

Defining Practices and Actors in Open(ish) AI. Current open source practices vary widely across Theme 6: AI Safety
labs and projects. Meta's LLaMA and Mistral models, while referred to as “open,” are released
173
with restrictive licences. ‘Truly’ open models, such as those by EleutherAI, Hugging Face’s
174
BigScience project, or StabilityAI, publish code, weights, and sometimes training data. In
175
China, Alibaba Cloud’s Qwen family , DeepSeek-LM , and Baidu’s forthcoming open-weight
177
176
ERNIE models all release or pledge to release weights (and, in Qwen’s case, code) on public
178
repositories. India’s AI4Bharat programme at IIT Madras follows a similar ethos, publishing
multilingual Indic models such as IndicTrans 2 under permissive licences to support low-
resource languages and local innovation. In contrast, some models from Google DeepMind
179
or OpenAI are fully closed. Initiatives such as the AI Alliance (led by IBM and Meta) promote
“open innovation,” but with few enforceable standards. This fragmented landscape reflects
ongoing negotiations between openness, safety, and commercial interest. The common thread
is that outside researchers can, in some capacity, inspect, fine-tune, and deploy the models
locally rather than calling a remote API, a shift advocates say restores scientific reproducibility
and lowers entry costs. 180

Benefits and Challenges—Accountability vs. Proliferation Risk. Proponents argue that open
source models enhance auditability, reproducibility, and global participation, allowing
actors—including those in Global Majority countries—to experiment with and build on frontier
capabilities. Critics, including some government agencies and labs, argue that open weights
increase proliferation risks for misuse, including disinformation or bioterrorism. Tensions
181
between openness and control have led to calls for differentiated governance: e.g., restricting
capabilities above certain thresholds, while keeping smaller models fully open.

Adoption and Advocacy. Open source AI is increasingly used for capacity building and
technological sovereignty. For instance, South Africa’s Masakhane project and Latin America’s
182
Cohere For AI have developed multilingual language models based on open source
183

172 Seger, E., Dreksler, N., Moulange, R., Dardaman, E., Schuett, J., Wei, K., Winter, C., Arnold, M., Héigeartaigh,
S. Ó., Korinek, A., Anderljung, M., Bucknall, B., Chan, A., Stafford, E., Koessler, L., Ovadya, A., Garfinkel, B.,
Bluemke, E., Aird, M., Gupta, A. (2023, September 29). Open-Sourcing Highly Capable Foundation Models:
An evaluation of risks, benefits, and alternative methods for pursuing open-source objectives. arXiv.org.
173 Meta’s LLaMa license is still not Open Source. (2025, February 18). Open Source Initiative.
174 Wiggers, K. (2025, March 19). ‘Open’ AI model licenses often carry concerning restrictions. TechCrunch.
175 Wiggers, K. (2025b, June 6). EleutherAI releases massive AI training dataset of licensed and open domain
text. TechCrunch.
176 QwenLM. (n.d.). GitHub - QwenLM/Qwen: The official repo of Qwen (通义千问) chat & pretrained large
language model proposed by Alibaba Cloud. GitHub.
177 DeepSeek’s release of an open-weight frontier AI model. (April 2025). IISS.
178 Williams, K. (2025, June 30). China’s biggest public AI drop since DeepSeek, Baidu’s open source Ernie, is
about to hit the market. CNBC.
179 Gala, J., Chitale, P. A., Ak, R., Gumma, V., Doddapaneni, S., Kumar, A., Nawale, J., Sujatha, A., Puduppully, R.,
Raghavan, V., Kumar, P., Khapra, M. M., Dabre, R., & Kunchukuttan, A. (2023, May 25). IndicTrans2: Towards
High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages. arXiv.org.
180 White, M., Haddad, I., Osborne, C., Liu, X. Y., Abdelmonsef, A., Varghese, S., & Hors, A. L. (2024, March 20).
The Model Openness Framework: Promoting completeness and openness for reproducibility, transparency,
and usability in artificial intelligence. arXiv.org.
181 Harris, D. E. (2023, December 6). How to regulate unsecured “Open-Source” AI: No exemptions. Tech Policy
Press.
182 Masakhane. Masakhane: A grassroots NLP community for Africa, by Africans. Retrieved June 19, 2025.
183 Cohere Labs. Aya. Retrieved June 19, 2025.

31