When Justice Can't Keep Up: Inside Public Defenders' Race Against the AI Revolution

A constitutional promise meets crushing caseloads—and an AI industry that's solving the wrong problems

Part I: The Impossible Equation

P5 remembers the exact moment she realized something had to change. Sitting in her cluttered office at 11 PM, she faced a stack of evidence that would take three weeks to review properly. Her client's trial started in five days.

"I had over a hundred thousand records," she recalls, her voice tight with the memory. "And there were so many duplicates. I was drowning."

She's one of America's public defenders—the attorneys who represent people too poor to hire lawyers. For more than 40 years, she's fought to uphold a constitutional promise made in 1963, when the Supreme Court declared in Gideon v. Wainwright that everyone deserves legal counsel. Justice Hugo Black wrote then that "lawyers in criminal courts are necessities, not luxuries."

Sixty years later, P5 manages about 20 major cases simultaneously. In Louisiana, some of her colleagues juggle 1,000 clients. In Utah, public defenders handle 525 misdemeanor cases each. The math is brutal: just a few hours per client when justice demands so much more.

"It's expensive," P5 explains, describing why her office can't simply buy its way out of the crisis. "I work for a public agency, so every kind of new technology needs to be properly tested out and vetted, and the funding has to be approved. So it's a little bit tricky that way."

Enter artificial intelligence. Tech companies promise that AI could help close this "justice gap"—transcribing recordings, summarizing documents, accelerating research. There's just one problem: almost no one has actually asked public defenders what they need.

Until now.

Part II: Listening to the Front Lines

From March to August 2025, a team of Princeton researchers did something revolutionary: they sat down with 14 legal practitioners and simply listened.

The interviews stretched across federal and state offices, from current public defenders handling active caseloads to law professors who'd spent decades in the trenches. The practitioners represented 2 to 40+ years of experience. Together, their insights painted a picture radically different from Silicon Valley's assumptions about what lawyers need.

The research team started with a hypothesis: public defenders would primarily use AI for legal research and writing. After all, that's where the $2 billion in legal tech investment has flowed. That's what headlines focus on when they cover AI in law.

They were wrong.

"We initially thought research and writing were the main areas," admits one researcher. "But early interviews showed us that was only part of their work. Participants kept identifying evidence review as more suitable for AI support."

The revelation forced the team to redesign their entire approach. They adopted what they call "abductive coding"—letting the data guide the theory rather than imposing predetermined categories. Over 84,908 words of transcripts, patterns emerged that challenged everything the legal tech industry thought it knew.

Part III: The Four Walls

Despite all the promise of AI tools, 71% of the practitioners interviewed don't regularly use them for work. Why? Four barriers stood between them and the technology that supposedly could save them.

Wall #1: The Confidentiality Crisis

"Without 100% confidentiality, I would never put my client's info into AI," one defender stated flatly.

It's not paranoia—it's professional duty. Public defenders handle the most intimate data imaginable: jail calls where clients confess fears to family members, medical records revealing mental health struggles, confidential informant identities that could mean life or death.

The problem: most AI tools send data to external servers. That data might be retained for training future models. Worse, it could fall outside attorney-client privilege—meaning prosecutors could subpoena it.

P3 and P4 explained that protective orders explicitly ban AI use in many cases. Even for simple transcription. "Prosecutors provide unredacted sensitive discovery—bank records, warrants identifying confidential informants—only under agreements that this material will not be shared outside the office or submitted to AI systems," they said.

P7 was blunt about what they needed: "I would like a closed universe. I would like to be able to give as much information as possible to the tool to be effective, and confidentiality is what limits me from doing that."

Not everyone saw confidentiality as insurmountable. P2 pointed out an inconsistency: "When people say they're worried about confidentiality, that is a perfectly valid reason to be skeptical, but then they'll use Gmail or Google Workspace or Dropbox to host files and scroll to the bottom and click agree."

Still, for most defenders, the risk wasn't theoretical. It was professional suicide waiting to happen.

Wall #2: Tools That Miss the Point

More than half the participants described AI outputs as fundamentally unreliable for actual legal work. The tools could return surface-level matches, but they missed what mattered.

P1, a research attorney with over 40 years of experience, shared a telling example: "I was asking about the standard for ordering mistrial based on post-trial conduct, and it gave me a case in which there had been a mistrial in the guy's first trial, but this decision was about a subsequent trial, so the issue of mistrial wasn't part of what the case was about."

The AI found the word. It missed the meaning.

P5 elaborated: "It feels superficial to me. It feels as if the analysis is made on a very literal-minded level, and it doesn't seem to really be able to read into those facts what the significance is of them, not what the definition is of the words, but what those facts mean in terms of the weight of the evidence."

Then there were the hallucinations. Not just obvious fake citations—those were almost easy to catch. It was the subtle misleading summaries, the incorrect interpretations, that posed the real danger.

P13 described their experience with premium legal AI: "They'll produce text that sounds reasonable, but they often cite to cases that do not stand for the principles that they say the case says. If you ask them a question about if a person has an insanity defense, will Missouri recognize temporary insanity, and it might give an answer that says the courts in Missouri are split on whether or not it will be recognized, but this court says it will if these conditions are met, and then they cite to a case. But when you read the case, the case does not say what the AI assistant said it says."

Independent research bears this out: even "hallucination-free" legal AI tools produce incorrect information 17-34% of the time. As of September 2025, courts have documented 246 incidents of fabricated citations—involving 135 attorneys, 103 self-represented litigants, and 3 judges.

One California judge expressed what many felt: "Fabricated citations and erroneous statements of law have required this court to spend excessive time," harming taxpayers and people "many of whom wait years for a resolution."

Wall #3: Bureaucracy and Confusion

For public defenders employed by government offices, using AI isn't a personal choice—it's governed by institutional policy.

P1's office took a hardline approach: "Our offices said, stay away from it outside of Lexus or Westlaw." Only enterprise-level systems approved. Everything else forbidden.

Other offices had no formal rules at all. Attorneys navigated based on vague judicial warnings and unhelpful training programs.

P2 mentioned judicial orders: "I think some judges have issued rulings or standing orders that requires attorneys to disclose if they've used any Generative AI." But the orders were "vague and unenforceable because no one can really agree on a definition of what Generative AI means—such as whether it would include a simple auto-correct or spelling mistakes."

P7 attended continuing legal education programs on AI and found them useless: many trainings "provide potential pitfalls, not applicable best practices."

Private attorneys who took public defense cases part-time had more freedom—but also less support when things went wrong.

Wall #4: The Price Tag

P2, a court-appointed private attorney, put it simply: "I don't have a Westlaw subscription just because I can't afford it, and I'm not part of an office. The only legal database I have access to is the free one that comes with the Washington State Bar, and that's Fastcase."

Even basic legal research tools were luxuries. AI subscriptions? Impossible.

P14 explained the gap: public defenders "scrounge for investigation on their own budget"—purchasing automated synopses of body-worn camera footage that prosecutors have access to "falls far short" of what's needed.

The cruel irony: the technology that could help defenders most is priced for the law firms that need it least.

Part IV: What Defenders Actually Need

Through iterative conversations, the researchers identified five categories of public defense work—each with different opportunities and risks for AI assistance.

🟢 Evidence Investigation: "If only we could just send it all out for transcription"

Ten of fourteen participants saw evidence investigation as the highest-impact opportunity for AI. This is where defenders are genuinely drowning.

"If only we could just send it all out for transcription, that would be a good start because I read a lot faster than I can hear," P5 said.

Modern criminal cases generate overwhelming digital evidence. Body camera footage can mean 12 hours of video when the actual crime took 4 minutes. Jail calls pile up by the hundreds. Document dumps reach 100,000+ pages.

The recordings are "informationally sparse," as one participant put it, "but it's difficult to know a priori which portion to watch or listen to." Police reports might point to key moments, but they "might not be comprehensive either and may give a biased representation."

Defenders identified four specific AI applications they desperately wanted:

Transcription (10/14 participants): Convert audio and video to searchable text. Enable keyword searching. "I read a lot faster than I can hear."

Summarization (10/14): Generate document and recording summaries for triage. Help determine what's worth detailed review before investing hours.

Span extraction (6/14): Find key moments in videos. "Can you give me the timestamp of when he reads the Miranda rights to the defendant?" P6 asked. Or as P2 described: "identify an area within a video—let's say warehouse door opens, and you don't want to go through six hours of video to find out when exactly that happens."

De-duplication (5/14): Identify and remove duplicate records in massive datasets. "I have a case right now with over a hundred thousand records. And there are so many duplicates," one defender shared.

But here's what defenders emphasized: "It's not as if the program is doing the actual work, but it's framing things so that [we] know what point to fast forward to, and that just saved [us] two hours."

Human judgment still drives the investigation. AI just makes it feasible.

🟡 Legal Research & Writing: "95% of the cases are going to be bad and 5% good"

Eight participants saw value in AI for generating summaries of legal information or providing starting points for new topics.

P1, the experienced research attorney, described the core challenge: "95% of the cases are going to be bad and 5% good, and winnowing is one of the most difficult tasks in criminal law research."

P10 imagined AI's potential: "You can say, how does every State treat guilty except for insanity defenses, and it would go survey all the States. That could be really persuasive where you could say, our state is the outlier."

Some used AI for quick orientation. P2 reported asking ChatGPT to "generate quizzes to test [their] knowledge and retention." P4 highlighted AI tools as "a launching point for additional cases."

A few experimented with document editing. P11 found it "beneficial for editing writing: running your motion or your brief into ChatGPT, and say, edit this for clarity and concision—you're gonna get an effective use."

But defenders emphasized this is a smaller part of their job than many assume. They usually rely on office templates rather than drafting from scratch.

And here's what was unanimous: all 12 participants stressed they would verify every citation, every legal claim. The verification burden means AI provides acceleration, not automation.

🟡 Client Communication: "AI cannot substitute having a strong, healthy connection"

Three participants use or would use AI to reword messages for clarity or translate legalese for clients with different education levels.

P7 uses AI tools to draft letters and emails: "I do not speak in legalese or convoluted ways, and ChatGPT helps keep me in check so that my parlance is a little bit more tailored to my audience."

P8 described a specific use case: "My clients have not gone that far in school. And so, one thing that you could use AI for is you could write a letter and then ask it, please, make edits to the letter to bring it to a 6th grade educational level."

Others imagined AI creating visual aids—flowcharts of the legal process, decision trees clients could reference after meetings.

But here's what AI cannot do: "AI cannot substitute having a strong, healthy connection with our client and making our client feel like we care about them," P11 emphasized.

Building trust requires reading nonverbal cues. P10 explained: "The clients often give either explicit or more subtle feedback, and the students learn from that. They learned that there was a really awkward pause right there, the client reacted kind of harshly to that, and so they get a sense of maybe don't say that thing again, or maybe word it in a more careful way."

Law professors worried that over-reliance on AI for client communication could prevent new defenders from developing these essential skills.

🔴 Courtroom Representation: "That's where the lawyer comes in"

Most participants believed courtroom work requires near-perfect reliability since there's no time to verify mid-argument. It demands emotional resonance with juries. It requires responding to unexpected developments.

P14 saw potential in preparation: "Your cross-examination in court isn't spontaneously conceived. If you're a good lawyer, you are preparing for that in advance, and so that's one other category that you could consider."

Some imagined AI helping with modular tasks—summarizing witness testimony, drafting opening statements, quickly pulling references.

But for oral advocacy itself? P8 was clear: "That's where the lawyer comes in."

Attorneys make emotional arguments and tell stories that resonate with human juries and judges. Some participants thought AI simply cannot do this.

P5 added a practical concern: courtroom tools would need near-perfect reliability "since attorneys cannot stop mid-argument to verify an AI's output."

And in some jurisdictions, the question is moot—electronic devices are prohibited in courtrooms entirely.

🔴 Defense Strategy: "I think I have more experience than ChatGPT does"

A few participants imagined AI helping with brainstorming—generating strategy suggestions, laying out pros and cons of different options, identifying weaknesses in the prosecution's case.

But strategy involves client preferences, accumulated wisdom about local judges and prosecutors, and confidential information that defenders don't want in any AI system.

"I think I have more experience than ChatGPT does in representing clients," P14 stated flatly.

P14 also explained how experience shapes strategy in ways AI can't replicate: misdemeanor clients in Washington, D.C. have almost no incentive to enter plea bargains "since the immediate penalties are minor and a guilty plea creates a lasting record, unless multiple convictions accumulate."

That kind of jurisdictional wisdom comes from years of practice, not training data.

P11 raised the confidentiality concern: "Defense strategy is one of the most privileged pieces of information, and I wouldn't want to just be putting it all in ChatGPT."

Even participants who could imagine AI helping with brainstorming emphasized that strategic judgment is inseparable from attorneys' lived experience.

Part V: The Three Non-Negotiables

Through all the conversations, defenders identified clear requirements for responsible AI adoption:

1. Mandatory Human Verification

Every AI output must be reviewed. "Anything generated or modified by AI goes through a check conducted by humans," P11 insisted.

When using AI for research, P2 described their process: use tools to "kickstart research" and find the "vocabulary necessary to do further research"—but always check citations and references.

For evidence investigation, defenders said they'd use AI for a first pass but "still watch the recordings themselves to verify the accuracy of the output and to catch more things that AI tools might have missed."

P11 imagined future regulation: "There needs to be a regulation in place to ensure that anything generated or modified by AI goes through a check conducted by humans because people who are working under time constraints would be tempted to take AI outputs at face value."

2. Avoid Over-Reliance

Participants in teaching positions expressed deep concerns about over-reliance preventing skill development.

P12 noted that "judges and prosecutors, being humans, may not act like what the court opinions make it seem like they do. If AI tools are modeled after what court opinions indicate would happen in a courtroom, it's going to give students a false sense of what they are actually going to experience."

An essential component of public defense is "the ability to think on their feet and respond to unexpected situations in the moment." Since AI tools cannot be used in courtrooms, defenders must rely on their own abilities.

P11 worried: "Are we going to create a whole generation of lawyers and students and just people who have lost the ability to think creatively on their own? AI makes mistakes, and it's important to make sure that our critical thinking stays sharp."

3. Preserve Human Relationships

Some aspects of public defense rely on human relationships and contextual wisdom that AI cannot replicate.

Communication with clients depends not only on words but also on non-verbal cues—tone, posture, hesitation. These skills can only be learned from engaging directly with clients and reflecting on interactions.

Courtroom advocacy is deeply tied to experiential knowledge. Public defenders learn the preferences, temperaments, and informal habits of specific judges and prosecutors.

P12 put it bluntly: "It'd be really hard for AI to help lawyers figure out what the best approach is in a given case because everything is also so personality specific."

Even if AI could one day observe nonverbal cues and generate contextually appropriate responses, P14 emphasized that "Clients have a right to counsel, not to machines."

The time a defender spends with a client is itself a sign of care and commitment. This relational value extends beyond the case: "A good public defender does a lot of social work—getting their client into the drug program, getting their client mental health services, making sure their client gets enrolled in school, or does their community service to complete their diversion programs."

These efforts signal to clients that their defenders genuinely care about them as people, not just case numbers.

Part VI: The Two Paths Forward

The research reveals a critical juncture. Without intervention, the United States is heading toward a two-tiered justice system where the promise of AI deepens rather than closes the justice gap.

Path 1: Vendor Dependence (The Dangerous Default)

Here's what worries researchers most: vendor domination.

Legal tech AI has attracted $2 billion in startup investment. Large law firms are partnering with vendors, accessing tools that public defenders can't afford. Startups consolidate contracts with established firms to obtain domain-specific training data.

The core repositories of U.S. law remain behind costly paywalls. Public initiatives like Free Law Project and the Caselaw Access Project have made progress, but they're insufficient against the proprietary data moats vendors are building.

The risk: a two-tiered system gets layered with a two-tiered AI capability gap.

Well-resourced prosecutors and private attorneys: cutting-edge AI tools for evidence analysis, research acceleration, case strategy.

Public defenders: falling further behind, drowning in digital evidence they can't process, unable to afford the subscriptions that could help.

The justice gap doesn't close. It widens.

And there's a more fundamental problem with vendor promises. Even when vendors claim "closed universe" confidentiality, if these services rely on proprietary back-end models hosted by corporations like Microsoft or OpenAI, "it is unclear whether these systems can be considered a closed universe insulated from outside access."

Inputs and outputs handled by third-party vendors could fall outside privileged information, subject to mandatory disclosure. There are no public evaluation benchmarks allowing public defender offices to audit whether vendors live up to their performance and confidentiality promises.

Path 2: Open-Source Equity (The Promising Alternative)

But there's another possibility—one that could genuinely close the justice gap rather than widening it.

Open-source models solve the confidentiality crisis. Running AI locally, on-device, means outputs become attorney work product, protected by privilege. Data never leaves the public defender's office. External servers aren't involved.

The cost barrier disappears. Models like GPT-OSS-20B (20 billion parameters) or vision-language models like Qwen2.5-VL can now run on common laptops. There's an upfront hardware investment, but it eliminates recurring subscription fees and API costs.

The performance gap is narrowing. Open models are approaching proprietary quality. Model optimization is accelerating. And critically, open models allow technical adjustments—like citation retrieval systems—that make outputs more trustworthy.

Domain-specific fine-tuning becomes possible. Public defenders don't need the most advanced version of LLMs. They need practical, accessible tools that support high-volume, rule-based tasks like document review, triage, and structured evidence summarization.

Research by Perron et al. shows that local LLMs achieve near-human accuracy in classifying and extracting problems in child welfare investigations. The technical capability exists. It just needs to be directed toward public defense.

But building this alternative requires three pillars:

1. Prioritize Open-Source Model Development

• Focus research on evidence investigation tasks: transcription, summarization, span extraction

• Develop domain-specific datasets spanning legal texts, client communications, and body-worn camera footage

• Create public evaluation benchmarks to assess not just text accuracy but also confidentiality, bias, and utility

• Build systems that can audit both open- and closed-source tools, including commercial vendors

2. Build with Defenders, Not for Them

• Participatory design that incorporates defender feedback through reinforcement learning

• Adaptive systems that dynamically integrate corrections from attorneys

• Personalization to local legal contexts and jurisdictional quirks

• Regular testing with actual public defender workflows, not lab scenarios

3. Invest in Organizational Capacity

• States should consolidate resources to provide technical support for public defender offices

• Hire data/IT specialists who understand both AI systems and legal confidentiality requirements

• Provide GPU hardware for local model deployment

• Create knowledge-sharing networks across offices to document lessons learned

This isn't just a technical challenge. It's a policy choice about what kind of justice system America wants to build.

Part VII: Beyond the Hype

The Princeton research challenges several common narratives about AI and law:

Myth: Lawyers using AI are lazy or incompetent.

Reality: Defenders want AI for time-consuming evidence review that currently makes adequate representation impossible. The goal is spending more time on strategy and clients, not less. They're asking for tools to process 12 hours of video footage, not to avoid the hard work of building a defense.

Myth: Every AI output must be manually verified, making AI pointless.

Reality: Verification burden varies dramatically by task. Checking citations takes hours and is manageable. Verifying whether AI correctly flagged events in 200 hours of video would defeat the purpose—that's the manual work AI is supposed to help with. The question isn't "human or machine" but how to design systems where human verification is proportionate to risk.

Myth: Legal AI is mainly about research and writing.

Reality: Defenders spend significant time distilling massive records into insights—video analysis, document classification, span extraction. That's where AI could meaningfully help. But that's not where the $2 billion has flowed. The tech industry is solving the wrong problems for this population.

Myth: AI will replace public defenders.

Reality: Defenders identified clear boundaries where AI cannot go: courtroom advocacy rooted in emotional intelligence, defense strategy requiring accumulated jurisdictional wisdom, client relationships built on trust and nonverbal communication. These rely on human judgment and relational value that machines cannot replicate.

Part VIII: What Happens Next

The researchers propose a forward-looking agenda with concrete steps:

For Researchers and Developers:

• Shift focus from legal writing to evidence investigation tasks

• Build public evaluation benchmarks for transcription accuracy on African American English, video analysis miss rates, bias in dialect-influenced outputs

• Partner with public defender offices for participatory design—build with defenders, not for them

• Prioritize open-source models that can run locally

For Policymakers and Funders:

• Direct $5-10M demonstration grants to defender-specific AI development, not commercial subscriptions

• Increase base funding for public defender offices to achieve parity with prosecutor tool access

• Support state-level technical support consolidation

• Mandate vendor transparency: public performance disclosure, confidentiality audits, price caps for government contracts

For Public Defender Offices:

• Pilot open-source transcription tools as highest-impact, lowest-risk starting point

• Establish clear AI use policies with mandatory verification protocols

• Invest in technical capacity: hire data/IT specialists, provide GPU hardware

• Share results across offices to accelerate learning

For Legal Professionals:

• Share insights about what tools would actually help your work

• Demand transparency and evaluation benchmarks from vendors

• Advocate for confidentiality-preserving, locally-deployable solutions

• Participate in research studies that center practitioner voices

Part IX: The Stakes

At 11 PM in her cluttered office, P5 finally found what she was looking for in that stack of 100,000 records. A single document that changed everything about her client's case.

She almost missed it. She easily could have, buried in the duplicates and the volume.

That's what keeps her up at night—not the cases where she found the needle in the haystack, but the ones where she didn't. The clients she couldn't adequately represent because there simply wasn't enough time.

"Clients have a right to counsel, not to machines," P14 said. It's become the rallying cry of this research.

AI should support that right, not substitute for it.

The promise isn't replacing attorneys with algorithms. It's giving overworked defenders time to do what only humans can: build trust with clients, develop winning strategies, read the courtroom, navigate the personalities of judges and prosecutors, make the emotional arguments that resonate with juries.

The promise is making the constitutional guarantee of Gideon v. Wainwright real, 60 years after Justice Hugo Black declared that lawyers are "necessities, not luxuries."

But that only happens if we build the right tools, with the right safeguards, addressing the right problems.

That requires listening to defenders themselves—not Silicon Valley's assumptions about what lawyers need.

The technology exists. The need is urgent. The path forward is clear.

The only question is: which path will we choose?

The Bottom Line

On one path: vendor dependence deepens the justice gap, creating a two-tiered system where only the well-resourced benefit from AI while public defenders fall further behind.

On the other: open-source innovation, participatory design, and strategic investment create tools that genuinely serve justice—locally deployed, confidentiality-preserving, and accessible to those who need them most.

P5 is still in that office, still facing impossible caseloads, still fighting for clients who can't afford anyone else.

But now, for the first time, there's a roadmap forward that actually addresses what she needs.

The question is whether we'll follow it.

This story is based on "How Can AI Augment Access to Justice? Public Defenders' Perspectives on AI Adoption" by Inyoung Cheong, Patty Liu, Dominik Stammbach, and Peter Henderson, Princeton University (2025). The research involved semi-structured interviews with 14 legal practitioners across the United States, combining qualitative analysis with literature synthesis to identify opportunities and safeguards for responsible AI adoption in public defense.