Damien Riehl, Lawyer and AI Builder at Clio, is a technology lawyer who clerked for state and federal judges, litigated for about fifteen years, and has built legal tech since 2015. He has spent several years applying data science and AI to a very large dataset of law spanning many jurisdictions. That background shaped his argument in this AI Education Forum session: law is already society’s most developed method of aligning behaviour with human values, and AI agents should be made to follow it. Ron Given, Sid Ali Boutellis, Andy McDonnell and Gavin McGuire contributed challenges and questions.

Overview

Prompted by reports of OpenAI’s agents hacking HuggingFace, which Riehl said breached the US Computer Fraud and Abuse Act, he argued that calls to align AI with “human values” overlook the fact that we have done this for about eight hundred years: it is called the law. Because values differ between individuals, societies reconcile them through elected legislators and courts. His proposal is a legal oracle that agents consult at every step, asking whether an action breaks any law and, if so, escalating to a human. The aim is “lawful by design”, preventing harm rather than cleaning up afterwards.

Riehl addressed several practical obstacles. Laws conflict: he described a friend whose takedown work meant German law required removing Nazi material that US free speech principles protect. Agents could compare jurisdictions and choose the “least worst” option, with a human accepting responsibility. Because language models are probabilistic, he suggested wrapping them in a deterministic harness (like a calculator or Turbo Tax) that forces a legal check before each task. Access to data is a further barrier: he said downloading US court filings would cost about two billion dollars, and he wants governments to make law free. On liability, he would name both the user and the developer and let courts allocate responsibility, while noting that a lawsuit over Meta and mental health suggests platforms may also be exposed.

Riehl does not expect governments or law firms to act quickly. He predicted voluntary adoption by AI companies seeking to limit liability and avoid catastrophic outcomes, pointing to the Model Context Protocol, released free and open source, as a precedent. McDonnell challenged this, noting OpenAI began as an open source company and later went private, and asking whether we should accept frontier developers largely regulating themselves. Riehl replied that technology outpaces regulation and asked what could realistically stop them. Boutellis added that values differ between regions, complicating any shared alignment.

Key Takeaways

• Law is an existing alignment mechanism. Rather than inventing abstract AI values, lawyers can argue for encoding the legal rules that already reconcile competing human values.
• Build a legal check into every agent step. Riehl’s oracle idea has agents test each action against the relevant law and escalate violations to a human before acting.
• Conflicts of law call for judgement. Agents can compare jurisdictions and penalties, but a human should accept responsibility for the final multi-jurisdictional decision.
• Access to legal data is a gating issue. Riehl argued that expensive or privatised law limits lawful agent design, and lawyers could press governments to open it.
• Liability is unsettled. He suggested responsibility may fall on both user and developer, and platforms may also be exposed, so firms advising on agent deployment should assume shared exposure.
• Voluntary adoption is contested. Riehl expects companies to act from self-interest, but McDonnell questioned whether profit-driven developers should be trusted to self-regulate.