Open vs Closed AI: The Risks of Free Models

Overview

The global debate over artificial intelligence governance has intensified, shifting from academic theory to urgent policy and commercial strategy. At the center of this discourse is a fundamental divergence regarding open versus closed AI models. This split is no longer just about technical preferences; it reflects deep disagreements on safety, economic power, and geopolitical stability.

Late last month, Anthropic released a significant report warning that GLM-5.3, a powerful new model developed by the Chinese company Z.ai, poses a major cyberattack risk. The core of Anthropic’s concern is that GLM-5.3 is freely available for anyone to download and modify. This incident highlights the growing tension between the democratization of AI technology and the need to prevent its misuse by bad actors.

The Commercial and Political Divide

Much of the current debate is playing out in public view, with substantial financial interests influencing the arguments on both sides. The positions taken by industry leaders often mirror their business models, creating a complex landscape where safety concerns and commercial incentives are deeply intertwined.

"I think a lot of the way you come down on the openness debate is influenced by how you come down on the AI debate." — David Gray Widder, Assistant Professor, University of Texas at Austin School of Information

On one side, Nvidia CEO Jensen Huang advocates enthusiastically for the United States to develop more open models. As a manufacturer of hardware used to run these models locally, Huang has a direct stake in a future where individuals and organizations purchase their own computing infrastructure rather than relying solely on cloud APIs.

Conversely, Anthropic’s CEO Dario Amodei advocates for a more cautious, restrictive approach. As the head of a company that provides powerful closed models via API, Amodei argues that unregulated open models could undermine safety standards and devalue the rigorous guardrails implemented by frontier labs.

Despite these commercial motivations, the disagreement is not purely mercenary. It stems from genuine philosophical differences about how AI should be governed and who should have access to its capabilities.

Understanding the Models

To evaluate the risks and benefits, it is essential to understand the technical distinction between open and closed systems. A closed model, such as those powering ChatGPT or Claude, is accessible only through websites, applications, or APIs. Users send prompts and receive responses without any visibility into the model’s internal architecture or training data.

In contrast, an open model allows users to download the model weights directly. These weights are essentially a long series of numbers encoding the model’s structure and learned knowledge. This transparency enables users to run the software on their own hardware, modify it, and even retrain it for specific tasks.

While closed models currently hold an edge in raw capability—dominating the top spots on the Artificial Analysis LLM leaderboard—the gap is narrowing rapidly. Open models are catching up in development speed, making them increasingly viable for sophisticated applications.

The Case for Openness

Proponents of open AI argue that restricting access concentrates too much power in the hands of a few large corporations and governments. The benefits of openness span research, business, and democratic equity.

• Academic and Technical Research: Researchers require open models to study inner workings, analyze "chain of thought" processes, and develop new safety mechanisms. Closed models limit researchers to observing inputs and outputs, hindering deep scientific inquiry.
• Economic Efficiency: Open models are often cheaper to deploy. Even when businesses rely on external providers to run the models, the underlying technology is less costly than proprietary API calls.
• Data Security: For companies handling sensitive or proprietary information, running a local open model ensures data never leaves their secure infrastructure, avoiding the privacy risks associated with sending data to third-party clouds.
• Democratization of Power: Open models challenge the tight control frontier labs have over AI usage. They ensure that foundational technology remains a public good rather than a private asset.

Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change, emphasizes that the question of who controls AI infrastructure is as critical as who controls information on the internet. Without open access, the most sophisticated hacking and defense models remain exclusively in the hands of US companies and government, potentially creating a geopolitical imbalance where other nations are defenseless against cyberattacks.

The Risks and Safety Concerns

The counter-argument centers on the potential for misuse. When powerful models are freely available, they can be easily adapted by malicious actors for cyberattacks, disinformation campaigns, or the creation of bioweapons. The GLM-5.3 controversy underscores this fear, as the model’s accessibility allows anyone to integrate it into offensive cyber tools without needing deep expertise in AI development.

Furthermore, there is a concern that open models may lack the rigorous safety alignment protocols implemented by major labs. Closed models often undergo extensive red-teaming and content filtering before release, whereas open models can be deployed with minimal modification, potentially releasing unfiltered or dangerous capabilities into the wild.

Hamish Hobbs, director of AI policy at the Centre for Long-Term Resilience, offers a nuanced perspective. He suggests that while open models pose risks, a world where individuals have loyal, secure, and safe AI systems is preferable to one where citizens must place blind trust in opaque government or corporate systems. The challenge lies in mitigating the risks of open distribution without falling back on total centralization.

Balancing Safety and Democratization

There is a spectrum of approaches between banning open models entirely and releasing them with no oversight. Governments and societies must navigate trade-offs between security and access. Complete prohibition stifles innovation and entrenches monopolies, while unrestricted release may accelerate harmful applications.

Future policy will likely involve a mix of watermarking, usage monitoring, and selective openness for models below certain capability thresholds. The goal is to preserve the benefits of distributed AI research and deployment while erecting enough barriers to prevent catastrophic misuse. As the technology evolves, the definition of "safe" and "useful" will continue to shift, requiring ongoing dialogue between technologists, policymakers, and the public.