Current Projects

Our Research

The Robertson Lab conducts empirical research on artificial intelligence in law and health care. Our current projects share a common thread: testing what AI systems actually do in legal writing, legal research, and litigation, rather than what they are assumed to do.

Claims about AI in law are everywhere. Evidence is scarcer. We build the datasets and run the experiments that turn open questions into answerable ones, and we publish what we find, whether it flatters the technology or not.

Empirical Study · Ongoing

Detecting AI-Generated Text in Legal and Academic Writing

Can AI-generated writing be reliably distinguished from human writing in high-stakes legal settings? As generative AI enters classrooms, law offices, and courtrooms, institutions increasingly rely on commercial AI-text detectors to answer that question, often without independent evidence of how well those detectors work on legal writing specifically.

This study builds a matched corpus of human-written and AI-generated law school exam answers and evaluates how well commercial detectors separate the two. Because a detector that works on one model’s writing may fail on another’s, the project examines accuracy across multiple AI models and writing conditions, including “humanized” AI text designed to evade detection. Later phases extend the analysis to court filings before and after the arrival of generative AI, asking how the composition of real-world legal writing is changing.

Why It Matters

Law schools set academic-integrity policy, bar examiners certify competence, and courts sanction attorneys based partly on assumptions about whether AI writing can be identified. If detectors are less reliable on legal writing than their marketing suggests, or reliable only under certain conditions, then those institutions need to know. And this validation measurement exercise sets up subsequent research identifying what legal filings are written by AI and how well that is working for the litigants and the system.

Empirical Study · Ongoing

Benchmarking AI Legal Research Tools

Legal AI platforms promise accurate, hallucination-free research, but how do they actually perform? Building on recent scholarship documenting hallucination in legal AI, this project evaluates commercial legal research tools and general large language models against a benchmark of legal questions, measuring accuracy, citation quality, and hallucination rates across platforms.

The study grades each platform’s answers under a common rubric, with attention to the questions where tools fail differently: retrieving the right authority but misstating it, citing real cases for propositions they do not support, or answering confidently where the law is unsettled. The design also examines how consistently human graders and AI-assisted grading agree, a methodological question that matters for anyone trying to evaluate these tools at scale.

Why It Matters

Lawyers are already relying on these tools, courts are already receiving their outputs, and law schools are already teaching with them. An independent, evidence-based picture of when legal AI can be trusted, and where it fails, helps the profession adopt the technology with open eyes.

Empirical Study · In Development

AI-Drafted Advocacy and Litigation Outcomes

Does AI-assisted drafting help or hurt litigants? Public debate has focused on AI’s failures in court, but even where nothing goes visibly wrong, do AI-drafted briefs persuade judges more, less, or about the same as human-written ones?

This project studies motions filed in federal court, combining AI-detection signals with case outcomes and controls to ask whether AI-drafted advocacy fares differently in real litigation. It is one of the first attempts to connect AI writing to real-world results rather than simulated evaluations.

Why It Matters

If AI drafting systematically changes outcomes in either direction, that reshapes access-to-justice arguments, professional-responsibility rules, and how courts should think about disclosure of AI assistance.

Empirical Study · Ongoing

The Robustness of AI Text Watermarking

Are you dealing with a human or a robot? The European Union’s AI Act (Article 50) mandates that providers of AI systems ensure machine-readable marking for synthetic content, including text. If this legal mandate is to be useful, and do more good than harm, then these watermarks must be robust against tampering.

Why It Matters

If watermarks are not robust, then the law could do more harm than good, by facilitating false negatives that people rely upon. We are testing that predicate empirically.

Empirical Study · Ongoing

AI Summaries of Scientific Evidence

Increasingly, both patients and clinicians rely on AI to advise and make health care decisions. But how well does AI summarize the scientific evidence? How does it weigh secondary discussions of the literature, versus official abstracts of scientific papers, versus the full body texts of scientific papers, which are often buried behind paywalls?

Why It Matters

The scientific literature is not easy to synthesize, and errors may cause patients to stop taking their meds. For example, a scientific abstract may conclude that a drug does not work, when the body of the paper admits that the study was underpowered to detect important benefits on the key outcome. Our preliminary research suggests that AI over-relies on the qualitative claims of the abstract, rather than scrutinizing the confidence intervals or reading the disclaimers buried in the full text.

Legal & Empirical Analysis · Ongoing

Liability at the Human-AI Interface

While “human-in-the-loop” is often invoked as a simple solution to every AI problem, humans are sometimes the problem. Sometimes we actually need AI to say, “I am sorry, Dave, I cannot do that.” The proper allocation of responsibility is shifting quickly as AI becomes more capable, while humans can be distracted, forgetful, or drunk. So just as we need humans to override AI when it is behaving dangerously (for example, a medical system proposing the wrong diagnosis), so too we need the AI to override humans when they are failing (for example, when a driver is speeding egregiously, the AI arguably needs to comply with the law for the safety of others).

Why It Matters

Human safety and welfare require the right allocation of responsibility between AIs and humans. When a physician uses an AI diagnostic tool or a driver uses an advanced driver-assistance system, and harm occurs, juries must navigate the allocation of fault between the human and the corporate developer of the AI system.

Empirical Study · In Development

An Ambient AI Agent for Consumer-Directed Health Care

American health policy has spent four decades exposing households to the price of their own care, on the theory that patients who face costs will shop, decline low-value services, and discipline providers. While patients now have the incentives, do they have the information and capacity to act on them? Can AI help a patient know, decide, and act? The needs are many: a physician recommends a treatment made by a company that pays them a consulting fee; a drug formulary changes with an appeal clock already running; a hospital charity-care policy could save thousands for a cash-strapped patient; bankruptcy can wipe away debts.

This project develops and tests whether an AI agent can supply what “consumer-directed health care” policy has always assumed. The intervention is an ambient patient agent: software holding standing, revocable authorization over a household’s clinical and financial data streams, which watches those streams, alerts the household when a decision is live, works the decision through to a recommendation, and executes only on instruction.

Why It Matters

In the United States, the safety net is delivered largely through private-law instruments: premium tax credits routed through private carriers, the employer-coverage tax exclusion, statutory external review of denials, the charity-care obligations of nonprofit hospitals, and bankruptcy discharge. A household that never appeals a denial or never claims charity care forgoes a public benefit as surely as one that never files for SNAP. Whether agentic AI can close that claiming gap, for whom, and at what cost per household has not been tested.

Empirical Study · Ongoing

AI Agents as Negotiators

Long ago, AI systems beat humans at games of strategy, like chess and Go, and AIs can produce endless amounts of persuasive text. But how well do they negotiate? Can they create value by jointly discovering solutions (growing the pie), or just haggle (splitting the pie)? How do they navigate the disclosure dilemma? We are training models and running thousands of negotiations to see how they do.

Why It Matters

AIs may help resolve legal disputes. And once we understand how AIs compare to human negotiators, AI simulations can be used as models for hypothetical negotiations, which are sometimes used in legal doctrine. For example, in patent cases, damages are keyed to a license that would have been negotiated at the time of the infringement.

Work With Us

Interested in our work or in contributing to a project? We welcome collaborators, including students, from a wide range of fields.

Contact the Lab

This is the research lab website of Professor Christopher Robertson, Boston University School of Law. Boston University Website Disclaimer