Open models have real benefits. Do they outweigh the risks?

Open models have tangible advantages, but do those advantages outweigh the risks they entail? Late last month, Anthropic released a report highlighting that the Chinese company Z.ai's GLM-5.3 model may pose a significant cyberattack risk. This model is freely available for anyone to download and modify, prompting the latest debate in the ongoing discussion surrounding open vs. closed AI models.

There's a long-standing debate about whether frontier AI models should remain closed, like GLM-5.3, or remain open. The recent debate about open models is widely publicized, with commercial stakes involved. For instance, Nvidia CEO Jensen Huang supports the development of more open models in the US, while Anthropic's Dario Amodei advocates for a more cautious approach due to potential threats from cheap, potent models.

David Gray Widder, an assistant professor at the University of Texas at Austin School of Information, notes that one's stance on open model debate often depends on their stance on the AI debate. Safety advocates are worried about potential misuse of open models by malicious actors, while those focusing on practical AI potential argue that such power should not be confined to a few major corporations. They believe open models could help disseminate this technology more widely.

The choice to regulate or permit open models is a complex issue involving a balance between safety and democratization. Governments and societies must weigh the potential risks and benefits of open models carefully. Open models, also known as "open-weight models," provide users with the ability to directly access and modify the model's internal structure - encoded in a long series of numbers called model weights.

Currently, closed models like those powering ChatGPT and Claude are more powerful than open models. However, many see value in using open models over closed ones. Because open models are downloadable, users can run them on their own computers, or even retrain them for different purposes if they have the skills and resources. For academic researchers, working with open models is crucial as it allows them to observe the inner workings of the models, the 'chain of thought' used by the models, combine multiple open models to build new systems, compress them for better efficiency, and fine-tune them for new applications.

Businesses also see the benefits of open models. Even if they rely on external providers to run them, open models are often cheaper than closed models, according to a recent report from Mozilla. Additionally, if businesses want to process sensitive data or proprietary information using AI, running a local model on their hardware may be the best way to ensure security.

But most importantly, open models challenge the dominant control that frontier labs have over how most of us use AI. Elizabeth Seger, senior policy advisor at the Tony Blair Institute for Global Change, argues that the key argument for open models is preventing the concentration of power, ensuring this foundational technology becomes a public good.

She envisions different futures for AI, depending on the role of open models. One possibility is that frontier AI systems remain controlled by a select few, giving them increased control over the world. With open models, every individual could have a loyal-to-them AI system that is safe and secure, potentially changing the balance of power between countries.