Why Public Defense Should Incorporate AI Carefully

In the United States, individuals facing criminal charges have a right to counsel. For those unable to afford private counsel, representation is usually provided by public defenders. In practice, this constitutional right to counsel meets the complexities of law, overwhelming

caseloads, and constrained resources. We argue that AI can support public defense if designed and deployed responsibly, and offices should start thinking about AI adoption.

In semi-structured interviews with 17 defenders, we assessed their AI perspectives. Common barriers to AI adoption include risks of hallucinations, failure of AI to appropriately apply legal reasoning or to capture state-specific nuances, and concerns around confidentiality (we confirmed these intuitions quantitatively in an exploratory pilot with a public defender office). On the other hand, defenders welcome use cases where AI instead provides information overviews, surfaces relevant briefs, makes sense of large volumes of evidence, and finds ‘needles in the haystack’.

The risks defenders fear—submitting briefs with AI-generated legal hallucinations in courts, producing inaccurate information, violating privilege—mostly apply to generative AI and chatbots. These risks have also been widely publicized in years of media coverage about legal hallucinations and lawyers being sanctioned as a result. That doesn’t mean that public defense should reject AI altogether. AI is an umbrella term, broadly describing generation, classification and retrieval methods. There exists AI which is directly responsive to actual defenders’ needs while minimizing risks. With the New Jersey Office of the Public Defender (NJOPD), we built the OPD Resource Library, an AI-powered search tool that makes the collective knowledge of the office available to all its employees within seconds. This tool is running securely within New Jersey state infrastructure.

Attorneys can ask any legal question, and customized AI models search through thousands of appellate briefs and hundreds of other internal documents to return the most relevant information. Importantly, these are all materials defenders already frequently consult and trust. The library mimics common office practices, where defenders already search for relevant briefs to kickstart their legal research. Having such documents handy can substantially streamline drafting of briefs. New Jersey public defender Jennifer Sellitti explains that defenders now have more time sitting across the table from clients—which is “what lawyers came to this work to do.”

This is one example of how AI can support public defense, and our research reveals other potential use cases: automatically identifying relevant spans in body-worn camera videos or flagging relevant snippets in thousands of pages entered into evidence. Rejecting AI adoption on the grounds of hallucinations and confidentiality concerns seems too simplistic and counterproductive. In the words of Sellitti: “When the existing technology couldn’t meet the demands of public defense, we didn’t lower our standards. We raised them.”

We recommend that offices start reflecting on their AI needs, and think about how to integrate AI responsibly in current workflows. Our collaboration with the NJOPD shows that it is possible to identify relevant use cases, design tools accordingly, iterate on the necessary AI research, and deploy them responsibly and securely. Collaborations like ours can serve as a blueprint, and in the best case result in deployed tools used by defenders, while advancing AI research itself.

We believe AI will have an impact on most knowledge work, including public defense. Under this premise, it seems rational for offices to start becoming more familiar with the technology. Dedicating some resources to AI now—so that offices are prepared for this future—very likely outweighs the costs. It does not have to be fully fleshed out collaborations or developing custom AI tools. A first step would be to assign an AI officer to start reflecting on their AI needs, internal AI policies and a long-term AI strategy. Having such in-house AI knowledge also facilitates procurement processes with vendors later—offices can guide these discussions, and decide on products given their actual needs.

Finally, responsible AI adoption is neither easy nor straightforward. In our work, defenders often emphasized that in the end, the goal of public defense is to provide the best possible outcomes for clients. And thus, “AI should be making [...] lawyers more human, not less,” Sellitti says. One of our interview participants argued that “clients have a right to counsel, not to machines.” We couldn’t be further away from advocating for AI to automate public defense, but encourage offices to start reflecting on how AI could responsibly be integrated in current practices.