Can Open-Source AI Models Be Trusted? Understanding Licensing, Risks, and Challenges

The rise of artificial intelligence (AI) has sparked an ongoing debate about the openness of AI models, their licensing structures, and associated risks. A recent discussion organized by the Software Freedom Law Centre (SFLC) shed light on the growing significance of open-source AI, the challenges it faces, and its role in the larger tech ecosystem.

What Defines an Open-Source AI Model?

The Open Source Initiative (OSI) outlines four primary criteria for an AI model to qualify as open-source:

Freedom to use the system for any purpose.

The ability to study and inspect the system's components.

Permission to modify the system for any purpose.

Sharing the system with or without modifications for any use.

Interestingly, despite its open-source claims, Meta's Llama 2 model falls short of OSI’s standards. Restrictions on training data access and commercial use by companies with over 700 monthly active users limit its openness. This controversy underscores the need for a universal definition of open-source AI, as Sunil Abraham from Meta India pointed out, suggesting that even 98% compliance might not suffice without consensus.

Challenges of Open-Source AI Development

Open-source AI models present developers with several hurdles:

Licensing Ambiguity: Determining the correct licensing conditions is critical. Chaitanya Chokkareddy, CTO of Ozonetel, highlighted how this uncertainty has delayed the release of his Telugu speech recognition AI model.

Copyright Issues: Companies fear litigation for using copyrighted training data. Mozilla’s Udbhav Tiwari warned that transparency in datasets could lead to lawsuits from authors and publishers.

High Costs: Developing open-source AI models is resource-intensive. Smita Gupta of the Open Justice for AI initiative shared her experience of building ""Aalap,"" a legal AI model, noting the absence of standard benchmarks and toolkits.

Risks of Open-Source AI Models

While open-source AI models democratize access, they also pose significant risks:

Content Moderation Issues: Open models can be misused, as safeguards can easily be removed. Tiwari highlighted that such models might propagate harmful content like CSAM if not regulated.

Data Privacy Concerns: Gupta emphasized the risk of personal identifiable information (PII) leaking through multiple layers of the open-source stack.

AI Hallucinations: Open-source doesn’t eliminate risks of misinformation. According to Abraham, AI's ""black box"" nature makes it challenging to trace hallucinations to specific features, even in open systems.

The Push for Open-Source AI in India

India’s open-source enthusiasts, like Chokkareddy, see an opportunity to challenge Big Tech's dominance by advocating for alternative AI development models. Unlike proprietary approaches by giants like OpenAI and Google, open-source models prioritize transparency and decentralization.

The Future of Open-Source AI

Despite its challenges, open-source AI holds immense potential to level the playing field in AI innovation. Regulatory frameworks, such as the EU’s AI Act, aim to strike a balance between openness and safety. However, experts agree that a standardized approach to licensing, ethical safeguards, and data transparency is essential to ensure the technology's responsible use.

As AI evolves, the debate between open-source and proprietary models will likely shape the future of innovation, ethics, and inclusivity in technology.