Not all ‘open source’ AI models are actually open: here’s a ranking

🌈 Abstract

The article discusses the implications of the European Union's Artificial Intelligence (AI) Act for research and AI models like ChatGPT. It highlights the lack of transparency and openness in the AI models developed by major tech companies, despite their claims of being "open source." The article also explores the legal and scientific importance of true open-source AI models.

🙋 Q&A

[01] What the EU's tough AI law means for research and ChatGPT

1. What are the key issues discussed in this section?

• Major tech companies like Meta and Microsoft are claiming their AI models are "open source" while failing to disclose important information about the underlying technology.
• The definition of "open source" for AI models is not yet agreed upon, but advocates say that true openness is crucial for making AI accountable.
• The EU's Artificial Intelligence Act will exempt open-source general-purpose AI models from extensive transparency requirements, which could incentivize companies to label their models as "open source."
• Researchers analyzed 40 large language models and found that many models claiming to be "open" or "open source" are actually just "open weight," meaning outside researchers can access and use the trained models but cannot inspect or customize them.

2. What are the key viewpoints and data presented in this section?

• Researchers Mark Dingemanse and Andreas Liesenfeld created a "league table" that identifies the most and least open AI models based on 14 parameters.
• They found that many models from major tech companies like Meta and Google DeepMind are just "open weight" and do not provide details about the data used to train the models.
• In contrast, models developed by smaller firms and research groups tended to be more open.
• The researchers argue that true openness is essential for reproducibility and scientific innovation in AI.

[02] Open-source AI chatbots are booming — what does this mean for researchers?

1. What are the key issues discussed in this section?

• The article highlights the lack of scientific papers and peer review for many of the AI models, with companies instead relying on "blog posts with cherry-picked examples" or "corporate preprints that are low on detail."
• The researchers argue that without access to the underlying data and model specifications, it is difficult to assess the true capabilities and limitations of these AI systems.
• The definition of "open source" in the context of the EU's AI Act will likely be a "single pressure point" targeted by corporate lobbies and large companies.

2. What are the key viewpoints and data presented in this section?

• The researchers emphasize the importance of openness and transparency for scientific reproducibility and innovation in AI.
• They argue that without access to the details of how these AI models are developed and trained, it is difficult to trust their capabilities and ensure they are not misused.
• The researchers hope their analysis will help fellow scientists avoid "falling into the same traps" when looking for AI models to use in their research and teaching.