Open source artificial intelligence (AI) is gaining momentum following advancements like OpenAI and ChatGPT, but industry experts warn that confusion still surrounds what open source truly means in this context. While more AI models are being publicly released, many carry licenses that restrict use, undermining the very principles open source software was and continues to be built on.

In this episode of An Eye on AI, Steve Watt, Vice President in the Office of the CTO at Red Hat, discusses the evolving definition of open source AI, the challenges surrounding licensing and reproducibility, and why transparency, data ownership, and community collaboration are essential for building trustworthy AI systems.

Watt believes there is still significant confusion around what truly qualifies as open source AI. Red Hat’s minimum criteria are that a genuine open source AI model must include both openly licensed model weights and the open source software components required to run them. Watt points out a gap between industry claims and reality, saying, “A lot of models that purport to be open source don’t even meet that basic acceptance when it comes to the usage of their model weights.” Many of these models would not meet the standards outlined by bodies like the Open Source Initiative (OSI), particularly where licenses discriminate against specific groups or uses.

While AI software frameworks like PyTorch and inferencing tools are often truly open source, the same cannot be said for training data and pre-training pipelines. Watt highlights the New York Times lawsuit, where proprietary data was allegedly used without permission. This prominent example demonstrates the murky legal and ethical waters surrounding AI training datasets. Watt sees this area as the next frontier in defining what constitutes open source AI.

A key distinction between traditional open source software and AI lies in reproducibility. While software builds are typically deterministic (meaning they produce the same results given the same inputs), machine learning (ML) models can diverge even when trained on identical data with the same pipeline. Despite not being part of the OSI’s formal definition, reproducibility remains an implicit expectation in open source development. Watt explains how this complicates its application to AI.

On-premise solutions are becoming increasingly vital. Watt explains that organizations want to own the entire AI exchange internally to maintain full autonomy over their data and understand exactly where it’s going. This shift is being driven by factors such as compliance requirements, protection of trade secrets, and national regulations.

Major national investments in AI infrastructure, which Watt refers to as “hyper centers”, illustrate this trend toward managing AI systems in-house. These large-scale, nation-state-backed data centers are designed to pre-train models for specific outcomes and highlight the strategic value of retaining control over the full AI pipeline.

Watt believes open source AI is evolving similarly to open source software through experimentation, community involvement, and gradual convergence toward best practices. Transparency, shared innovation, and active collaboration will be key to shaping open source AI standards that are both widely adopted and truly open.

Guest: Steve Watt
Company: Red Hat
Show: An Eye on AI

This summary was written by Emily Nicholls.