When Meta released Llama to the research community with relatively permissive licensing, it sparked a debate that shows no sign of resolving. The open-source AI argument has two poles, and both are compelling.
Proponents argue that open models democratise AI capability, enable academic research, allow independent safety auditing, and prevent a dangerous concentration of power among a small number of closed-model labs. The history of the internet itself is offered as evidence: open protocols enabled an ecosystem that closed proprietary networks never could have produced.
Critics, including a growing number of AI safety researchers, argue that this analogy breaks down at a critical point: open-source software generally cannot be used to cause catastrophic harm at scale. The concern is that sufficiently capable open-source AI models could be fine-tuned for dangerous purposes — from disinformation at scale to assisting in the synthesis of dangerous substances.
The regulatory landscape is fractured. The EU AI Act takes a risk-based approach but has significant carve-outs for open-source models that critics say effectively create a loophole for frontier systems. No international governance body with meaningful authority currently exists.