Open Source vs Closed: The Regulatory Divide

The AI industry is divided along a fault line that has profound implications for regulation: open source versus closed source. On one side, Meta Platforms has released its LLaMA series of large language models as open-source, allowing anyone to download, modify, and deploy them. On the other, OpenAI, Anthropic, and Google keep their most capable models proprietary, accessible only through APIs with usage restrictions. This divide is not just a business strategy difference — it creates fundamentally different regulatory challenges that policymakers are only beginning to grapple with.

Meta's open-source strategy is inseparable from its regulatory strategy. By releasing LLaMA models freely, Meta argues that AI development should be open and decentralized, that transparency is best achieved through open access rather than regulation, and that proprietary models controlled by a few companies pose a greater risk to competition and safety than open models. The company has invested heavily in the AI Alliance, a consortium of over 50 organizations including IBM, Intel, and academic institutions, which advocates for open AI development. See our Meta company profile for full lobbying details.

OpenAI and Anthropic make the opposite case. They argue that the most capable AI models — those approaching or exceeding human-level performance in certain tasks — require careful, controlled deployment. Open-sourcing such models, they contend, would make it impossible to prevent misuse, from generating bioweapons instructions to creating sophisticated cyberattack tools. OpenAI's lobbying has focused on regulatory frameworks that distinguish between model capability levels, with stricter requirements for more powerful systems — a framework that would naturally favor their controlled-release approach.

The EU AI Act has brought this tension into sharp regulatory focus. The Act's requirements for "general-purpose AI models" — including technical documentation, copyright compliance, and risk assessment — apply differently depending on whether a model is open-source. Open-source models with "systemic risk" (roughly, models trained with more than 10^25 FLOPs) face nearly the same requirements as closed models, but smaller open-source models receive significant exemptions. This tiered approach has been praised by some as pragmatic and criticized by others as either too lenient on open source (safety advocates) or too burdensome (open-source advocates).

In the United States, the open-source divide has complicated legislative efforts. California's vetoed SB 1047 would have imposed safety requirements on models costing over $100 million to train — a threshold that effectively targeted both open and closed frontier models. Meta lobbied aggressively against the bill, arguing it would discourage open-source AI development in California. Governor Newsom cited similar concerns in his veto message. The debate exposed a real tension: how do you regulate a technology that, once open-sourced, cannot be recalled or restricted?

The competitive dynamics add another layer. Meta's open-source strategy is partly a competitive move against OpenAI and Google — by commoditizing the model layer, Meta aims to shift value to applications and platforms where it has advantages. OpenAI's closed approach preserves its ability to charge for model access. These business incentives shape each company's regulatory preferences in ways that don't always align with the public interest. The lobbying spend reflects this: Meta spent $25M+ opposing regulation that would burden open-source models, while OpenAI spent heavily supporting frameworks that would validate its controlled-release approach.

For policymakers, the open-source question is genuinely difficult. Open models enable innovation, academic research, and competition with tech giants. But they also make it harder to enforce safety standards, prevent misuse, and assign accountability when things go wrong. The regulatory frameworks being developed now will determine which approach prevails — and the lobbying dollars flowing on both sides ensure that this question will be settled by politics as much as by policy.