Skip to Main Content

Wesleyan Researcher Asks If Policy Can Match the Speed of AI

The development of mainstream artificial intelligence has progressed at warp speed, going from basic chatbots to models capable of directing tens of thousands of agents to solve complex problems in just a few years. Sebastian Zimmeck, associate professor of computer science, discussed the need to align AI systems with our law and policy at a recent talk on campus.  

Zimmeck, an expert on data privacy, focused on two areas where AI must improve: privacy and security; and policy and law. “The systems that we have should not encourage illegal behavior,” he said. “They should, in their output, conform to the policies and laws that we have.” 

He called for better transparency to be built into AI models, including improving public knowledge of how the systems work, more open access to its internal operations, and how AI companies handle consumer data. He also mentioned a need to be able to enforce the parameters set by privacy laws, a difficulty policymakers face while trying to regulate the traditional internet. At the same time, he said it’s important to have useful systems that fulfill their intended purpose. 

“We need to design and implement the system in such a way that it behaves in the way we want,” Zimmeck said. “That may sound obvious, but it's actually not the case for many systems that we have.” 

Zimmeck suggested creating protocols for AI and a regulatory “outer bound” that must not be crossed, but not regulating AI to the point it stifles innovation. For example, an internet protocol is a universal set of rules that govern how data is sent and received on the internet, like the use of “http” or “https” to signal whether a website is secure or not. It’s possible to create these guardrails in AI models through system prompts, which are instructions that model creators, like OpenAI or Anthropic, can use to set parameters for the ways their models behave. Another way to build guidelines into AI models is through reinforcement learning, or training a model over time by giving it feedback.   

“We also need to consider how the different parts of the system and how different systems interact with each other, and that is often governed by protocols,” he said. 

He believes building these protocols into the systems is necessary since relying solely on regulations can lead to issues. Historically, the creation of digital systems laws tends to be slow enough that by the time they are enacted, the technology the law governs has progressed significantly, rendering the laws outdated. 

Zimmeck said the public underestimates the upside potential of AI, while also underrating the risks it poses. He said people have begun to form human-like relationships with AI models at times and there is potential for nation states, especially authoritarian ones, to dictate how models respond to questions about their regimes. “This kind of censorship is a very nuanced and subliminal way of the model interacting with people,” he said. “It’s suggesting a certain direction that is not true, or one that is aligned with goals of these states.”  

Currently, Zimmeck is working on a project to measure and improve the legal reasoning in AI systems with Wesleyan students and scholars from the University of South Carolina and Columbia University. He said companies and governments are increasingly using AI to make legal decisions. For example, the U.S. government wants to use AI to help decide veterans’ disability benefits claims. The research group is documenting the mistakes AI models are making, with the goal of creating guardrails to ensure the models do not provide incorrect information or legal reasoning.  

He closed the talk by taking audience questions on topics such as environmental concerns surrounding AI, how the larger citizenry can have input in these privately owned models, and previous examples of regulatory success that could apply to the effort to govern AI.