This paper explores the concept of openness in artificial intelligence (AI), including relevant terminology and how different degrees of openness can exist. It explains why the term 'open source' – a term rooted in software – does not fully capture the complexities specific to AI. There is ongoing debate about the risks, benefits and trade-offs of making AI models or their components publicly available, particularly regarding increasingly advanced AI foundation models that exhibit general purpose capabilities. 

This paper analyses current trends in open-weight foundation models illustrating both their potential benefits and associated risks. It seeks to support policy discussions on how to balance the openness of generative AI foundation models with responsible governance.

Key findings

• Releasing foundation model weights has many benefits, such as allowing external evaluation, speeding up innovation, and spreading control over a potentially transformative technology.
• Releasing foundation model weights also presents potential for malicious use and unintended consequences, such as cyberattacks, sexual abuse, and violation of intellectual property and privacy rights.
• Because of the significant potential risks and benefits, foundation models need careful consideration when they are shared and used. 

Report structure

• Section 1 – defines key terms and scope, delving into the different levels of openness in AI
• Section 2 – analyses current trends in open-weight models using experimental data from the OECD
• Section 3 – illustrates the potential benefits and risks of open-weight models and presents the concept of marginality
• Section 4 – concludes.