As artificial intelligence systems become increasingly capable of acting on their own, Bennett Borden is focused on a question that goes far beyond how lawyers use AI: Who should write the rules that govern the machines?
And since lawyers have traditionally interpreted rules governing human conduct, what role should they play in designing rules for machines?
Borden, a lawyer and data scientist and founder and CEO of Clarion AI Partners, is developing AI systems in which principles governing an agent’s behavior are incorporated into the system’s design. He describes the approach as writing a “constitution” for AI, which includes defining a system’s role, setting boundaries and establishing principles for how it should act.
The concept of “constitutional AI” has roots in Anthropic’s approach to AI development. “You literally bound its behavior by giving it a job, like a prime directive,” Borden said.
Those rules can include “thou shalt” and “thou shalt not” principles addressing issues such as accuracy, privacy or how an AI agent interacts with a client. More sophisticated systems can incorporate scoring mechanisms to evaluate how well an agent follows those principles.
From Data Science to AI Constitutions
Borden’s interest in “constitutional AI” grew out of a career spanning law, technology and data.
While studying history, anthropology, sociology and economics at university in the United Kingdom, he became interested in technological revolutions and the forces that cause societies to change. In 1992, he was recruited by the CIA to work at a new data analytics operation. He recalled thinking, “That sounds really cool, but I’m not sure you have the right guy.”
It turned out that the agency was interested in understanding what could be learned from the enormous amounts of digital information people were beginning to generate through the internet, email and cellphones. Borden spent almost a decade working in data science before attending Georgetown Law and earning a graduate degree in data analytics from New York University.
He sees an inherent connection between the fields of law and data science. “In the practice of law, we are the ultimate dealers in information,” he said. “No matter what a lawyer does, we take in information, we apply it to the law, we add our legal acumen to it, and then we produce an information-based product.”
His career subsequently focused on that intersection. Before founding Clarion, Borden worked at large firms including Gibson Dunn and DLA Piper, where he advised clients on automated decision-making systems and represented companies including OpenAI, Anthropic, Google, Amazon and Microsoft. He also participated in early AI policy discussions, including in White House meetings, congressional briefings and with the National Institute of Standards and Technology.
Yet his thinking about the future of law changed when he was asked at DLA Piper to study how generative AI could affect legal work. Borden analyzed 10 years of time entries from approximately 14,000 attorneys across practice areas. He said he found that lawyers, regardless of specialty, repeatedly perform a relatively small number of tasks.
His conclusion was that at least 80% of what lawyers do could be better performed by AI, he said. “Big Law, as it is currently established, is not going to survive this,” he said.
Rather than spend the next decade trying to adapt the traditional model, Borden decided to build something different. At age 56, he left what he described as “the pinnacle of his career” to launch Clarion.
From Rules on Paper to Rules in Code
For Borden, governance cannot simply be a policy sitting in a binder or on a company intranet. The principles an organization establishes need to be reflected in how an AI system actually operates.
An organization might establish that an AI system’s purpose is to interact with customers about a particular product while giving it principles such as accuracy, compassion and empathy.
The challenge comes when those principles collide, such as accuracy conflicting with speed or privacy conflicting with safety. Although competing considerations are familiar to lawyers, AI can force organizations to make those conflicts explicit before a system begins acting on their behalf. “You have to build it with all the controls in place,” Borden said.
Systems can also be built in which one AI agent performs a task while another monitors its behavior. The second system can identify problems and send the first agent back to try again, making the agent’s behavior observable and accountable.
For similar reasons, Borden creates systems that include an evidentiary layer that offers compliance “telemetry,” allowing an organization to document what an AI system did and why. The distinction becomes important as AI systems move beyond generating text or retrieving information and begin performing tasks on behalf of organizations.
An organization may be able to articulate policies requiring accuracy, privacy or fairness. But when an AI system acts with a degree of autonomy, the organization also needs mechanisms for testing whether it is following those policies and documenting what happened when it did not, he said.
Who Writes the Rules?
AI autonomy is where Borden returns to the idea of “constitutional AI.”
If legal and ethical requirements for AI can be translated into technical controls, lawyers may become involved in the design of systems before they are deployed, helping determine their purposes, boundaries and mechanisms for accountability.
That does not necessarily mean lawyers need to become programmers. “The cool thing about this technology is, you don’t have to know how to actually do any of that,” Borden said. “You can just ask it.”
He continued, “This is a magical glowing orb of power that’s being handed to the average person. That will literally teach you how to use it for any purpose you can imagine.”
But the greater the power of the technology, he said, the greater the need for those principles to be built into the systems themselves.
The future role of the lawyer, in other words, may not be limited to interpreting the rules. It may increasingly involve helping write the rules that machines follow.