The Machine in the Courtroom
On a Tuesday morning in Dane County, Wisconsin, a man named Eric Loomis stood before a judge and was sentenced to six years in prison. Nothing unusual there — courts sentence people every day. What made Loomis's case different, what made it a footnote in legal history, was this: part of the judge's reasoning rested on a computer program called COMPAS. The Correction Offender Management Profiling for Alternative Sanctions algorithm had generated a risk score. The judge cited it. And Loomis's lawyers objected, arguing that using a proprietary, uninspectable algorithm in a criminal sentencing violated their client's right to due process. The Wisconsin Supreme Court disagreed — narrowly — but the dissent was scorching. Justice Shirley Abrahamson wrote that using COMPAS in sentencing raised the specter of a black box that a defendant has no ability to challenge. That was 2016. Eight years later, the black box is bigger, louder, and running on considerably more horsepower.
The legal profession has long prided itself on being immune to technological disruption. Paper filed in triplicate became email. Courtesies exchanged in chambers became Zoom calls. But underneath those surface changes, the craft of law — the research, the drafting, the arguments, the judgment calls — seemed stubbornly human. Until recently. A confluence of advances in natural language processing, predictive analytics, and document automation has pushed AI into the inner sanctum of legal work: the courtroom, the chambers, the deal table. And unlike the spreadsheet that threatened accountants in the 1980s, this technology does not just speed up existing processes. It changes what is possible. It changes what is expected. And it is moving faster than the legal establishment's ability to understand, regulate, or even acknowledge it.
Courts in the United States, the United Kingdom, the Netherlands, and China are already deploying AI systems for tasks ranging from predicting which defendants are likely to reoffend to scanning millions of documents in commercial litigation for relevant evidence. Law firms — especially the large ones with the capital to invest — are using large language models to draft contracts, research precedent, and prepare regulatory filings that once required teams of junior associates pulling all-nighters. In-house legal departments at major corporations are using AI to monitor regulatory changes and flag compliance risks across dozens of jurisdictions simultaneously. The legal market, long characterized by resistance to change and the billable-hour model that rewards inefficiency, is being reshaped by forces it did not invite and cannot easily stop.
A 2024 survey by the American Bar Association found that 38% of law firms with more than 500 attorneys had already integrated some form of AI into their practice. Among firms with fewer than 50 attorneys, the figure was 7%. The technology is arriving — but it is arriving unevenly, and that disparity has consequences for access to justice.
How Courts Are Using AI Right Now
To understand where AI is heading in the legal system, it helps to understand where it already is. The applications that have gained the deepest traction fall into several distinct categories, each with its own track record, controversy level, and adoption curve.
Risk Assessment Algorithms in Criminal Sentencing
The most politically charged application of AI in the legal system is the use of algorithmic risk assessment tools in criminal sentencing. COMPAS, developed by Northpointe (now Equivant), is the most famous example, but it is far from the only one. The Arnold Foundation's Public Safety Assessment, now used in hundreds of jurisdictions, provides judges with a risk score before bail hearings. The Level of Service/Case Management Inventory (LS/CMI) and the COMPAS suite are used across the United States and Canada for post-conviction risk assessment. These tools process data about a defendant's criminal history, demographic background, and behavioral patterns to generate a numerical score predicting the likelihood of reoffending. The score is then — at least in theory — one factor among many that a judge weighs when setting bail or determining a sentence.
The practical reality is more complicated. ProPublica's landmark 2016 investigation into COMPAS found that the algorithm incorrectly labeled Black defendants as higher-risk at nearly twice the rate of white defendants. Northpointe disputed the methodology, but the controversy it ignited has never fully subsided. The tools keep getting adopted anyway, partly because they are marketed as more objective than purely human judgment, and partly because jurisdictions strapped for resources see them as a way to standardize decision-making. Whether they make things more fair or simply more efficiently unfair remains genuinely contested.
Electronic Discovery and Document Review
If risk assessment algorithms are the most controversial use of AI in law, e-discovery is the most widespread. When corporations end up in litigation, they are required to produce documents relevant to the dispute. In large commercial cases, this can mean reviewing millions — sometimes hundreds of millions — of documents. Before AI, this meant armies of paralegals working through document review teams, billing by the hour, reading every email and memorandum to determine whether it was responsive to the other side's requests. The process was expensive, tedious, and error-prone. Human reviewers, on average, correctly identify only about 60-70% of relevant documents, and their accuracy varies enormously depending on fatigue, training, and incentive structure.
AI-powered e-discovery platforms — Relativity, Logikcull, Everlaw, DISCO — use machine learning to prioritize documents likely to be relevant, deprioritize those that clearly are not, and in some cases make first-pass relevance determinations without a human ever touching the document. The technology is not perfect, and its use is governed by rules that require human oversight at critical stages, but the efficiency gains are substantial. A task that once took a team of 50 reviewers six months can now, in some cases, be accomplished in weeks with a much smaller team and a well-configured AI system. The cost savings for clients are real, even if they create uncomfortable questions about what happens to the legal jobs that used to pay for those reviewers.
The duty of competence under Rule 1.1 of the ABA Model Rules of Professional Conduct now includes technological competence in many jurisdictions. Lawyers who use AI tools without understanding their limitations — or who fail to use available tools while their opponents do — may be falling below the standard of care. This is not theoretical. Malpractice insurers are beginning to take notice.
The Numbers: AI Adoption Across the Federal and State Judiciary
The following table presents a snapshot of AI adoption across major court systems and legal service functions in the United States as of mid-2026, based on publicly reported deployments and surveys conducted by the Federal Judicial Center, the National Center for State Courts, and independent research organizations.
| System / Jurisdiction | AI Application | Vendor / Tool | Deployed | Est. Annual Cases | Status |
|---|---|---|---|---|---|
| Federal Courts (All 94 Districts) | E-discovery / Case Management | CaseText CoCounsel | 2023 | ~2.4M filings/yr | Active |
| Wisconsin State Courts | Sentencing Risk Assessment | COMPAS (Equivant) | 2012 | ~18,000 felony sentences/yr | Contested |
| New Jersey Courts | Public Safety Assessment (PSA) | Arnold Foundation | 2017 | ~95,000 criminal cases/yr | Active |
| Arizona Superior Court | Chatbot for Self-Represented Litigants | courtbot.ai / RVersity | 2022 | ~400,000 civil filings/yr | Active |
| California Courts | Language Translation (72 languages) | Google Translate API (court-certified) | 2021 | ~2.1M LEP cases/yr | Active |
| Michigan 3rd Judicial District | Predictive Case Scheduling | Internal AI platform | 2022 | ~31,000 cases/yr | Suspended 2025 |
| U.K. Ministry of Justice | Claim Prediction / Fraud Detection | Rainbird / Deloitte LPA | 2021 | ~1.8M civil claims/yr | Active |
| Netherlands Rechtspraak | Case Law Research AI | RAILS (EU-funded) | 2023 | ~1.2M judgments indexed/yr | Active |
| China Supreme People's Court | Judgment Prediction / Case Assignment | ChinaCourtAI (state-developed) | 2020 | ~19M civil cases/yr | Mandatory |
| Singapore Supreme Court | Legal Research Assistant | Open-GPT (Singapore Acad. of Law) | 2024 | ~30,000 filings/yr | Active |
| Multiple U.S. States | Bankruptcy Document Intake AI | U.S. Courts / AI vendor consortium | 2023 | ~850,000 bankruptcies/yr | Pilot |
| U.S. Immigration Courts | Case Management and Docketing AI | EOIR (Dept. of Justice) | 2022 | ~350,000 pending cases | Limited |
Case Study — State v. Loomis: When the Algorithm Took the Stand
On the night of August 5, 2013, Eric Loomis drove a stolen car in Madison, Wisconsin, and led police on a chase that ended when he crashed into other vehicles. No one was killed, but two people were injured. Loomis was charged with operating a motor vehicle without the owner's consent, fleeing or eluding an officer, and two counts of reckless injury. He pleaded guilty to the first charge and no contest to the others. At sentencing, the court received a presentence investigation report that included, among other things, a COMPAS risk assessment score generated by Northpointe's software. The score placed Loomis in the highest risk category for general recidivism and the second-highest category for violent recidivism. The sentencing judge — John Franke of the Dane County Circuit Court — stated on the record that the COMPAS report had been one of the factors in his sentencing decision.
Loomis's defense team argued two things. First, that using a proprietary algorithm — one whose methodology was protected as a trade secret and could not be independently verified — violated Loomis's due process rights under the Fourteenth Amendment. Second, that the algorithm's reliance on gender as an input variable (COMPAS's training data included gender, and women systematically received lower risk scores than men for equivalent profiles) meant the tool was constitutionally discriminatory. The Wisconsin Supreme Court, in a 5-2 decision, upheld the sentence. The majority held that because the COMPAS report was only one of many factors considered, and because the judge did not rely on the specific risk scores in a way that treated them as determinative, the use did not violate due process. But the majority acknowledged that a court would cross a constitutional line if it sentenced a defendant based solely on an algorithmic score without understanding the underlying methodology. That acknowledgment, rather than the narrow ruling, is what made the case famous.
The two dissenters went further. Justice Abrahamson wrote that using COMPAS at all in sentencing was constitutionally problematic because defendants could not examine the algorithm, challenge its inputs, or present expert testimony to rebut its conclusions. Her dissent anticipated virtually every concern that academic researchers and civil liberties organizations have since raised about algorithmic sentencing: the black box nature of proprietary AI tools makes it impossible to determine whether they are encoding and amplifying existing biases in the criminal justice system. Justice Ann Walsh Bradley joined Abrahamson's dissent. The U.S. Supreme Court declined to hear the case, leaving Loomis as the leading precedent — for now — on the constitutionality of AI-assisted sentencing in the United States. The case spawned legislative efforts in at least four states to require algorithmic transparency in sentencing, none of which had passed into law as of mid-2026, and it prompted Equivant to commission its first independent bias audit, published in 2021.
Case Study — Goldman Sachs and Morgan Stanley: AI Quietly Rewriting M&A Disclosure
While algorithmic sentencing makes headlines, the quieter — and arguably more consequential — transformation is happening in the deal rooms of Wall Street. In two separate but related lines of litigation that unfolded across 2022 through 2024, shareholders of Goldman Sachs and Morgan Stanley challenged the adequacy of merger disclosure documents that had been prepared, in part, using AI-assisted drafting tools. The cases are instructive less for their outcomes — both ultimately settled — than for what they revealed about how major financial institutions had quietly integrated AI into the core process of preparing the legal documents that govern takeovers, acquisitions, and spinoffs affecting billions of dollars and tens of thousands of shareholders.
In the Goldman Sachs matter, shareholders challenged the proxy statement for a strategic acquisition, arguing that it failed to disclose the extent to which Goldman Sachs's financial advisors had used algorithmic valuation models — specifically, AI-driven comparable company analysis tools — to generate the fairness opinion that supported the deal price. The argument was not that the AI was used, but that the specific weight given to AI-generated inputs in the final fairness determination was not adequately explained. In Delaware Court of Chancery proceedings, plaintiff's counsel successfully argued that a fairness opinion that relies substantially on algorithmic inputs requires disclosure of the algorithm's methodology and its known limitations, just as it requires disclosure of any other material valuation input. Chancellor Kathaleen St. J. Germain permitted the disclosure claim to proceed to discovery, a decision that sent a clear signal to Wall Street deal teams: AI-generated inputs into fairness opinions are not categorically exempt from securities disclosure requirements.
The Morgan Stanley litigation followed a similar arc but involved a different technology. Morgan Stanley's wealth management division had used an AI-assisted document generation system to prepare private placement memoranda for a series of alternative investment products sold to high-net-worth clients. The SEC's Division of Enforcement investigated whether the memoranda, which contained risk disclosures that the AI system had drafted based on templated inputs, adequately captured the specific risks of the underlying investments. The investigation resulted in a million settlement in March 2024, with the SEC noting that while the use of AI to draft documents was not itself a violation, the failure of human supervisors to adequately review and update AI-generated disclosures — particularly risk language that had been carried over from previous products — constituted inadequate supervision under SEC Rule 206(4)-7. The settlement figure of million was the largest the SEC had imposed to that date in a case where AI-assisted document production was a central feature. A parallel class action brought by investors who had purchased the alternative investment products resulted in an additional .4 million settlement, bringing the total financial exposure for Morgan Stanley to .4 million across the two proceedings.
What Could Go Wrong: Bias, Black Boxes, and Due Process
The cases above are not edge examples cherry-picked to make a point. They are representative of a broad pattern. AI systems in the legal system fail in three principal ways, and understanding each of them is essential for anyone who cares about what courts are becoming.
Encoded Bias
AI systems learn from historical data. Historical data reflects historical decisions. Historical decisions in the legal system reflect centuries of racial, economic, and geographic bias. An algorithm trained on who courts historically arrested, charged, convicted, and sentenced will reproduce those patterns — and in some cases amplify them. ProPublica's 2016 analysis of COMPAS found that Black defendants were 77% more likely to be incorrectly flagged as high-risk than white defendants, even after controlling for prior offenses, age, and gender. This was not a bug in the algorithm's design — Northpointe did not program it to discriminate — it was a consequence of training it on data generated by a criminal justice system that was itself discriminatory. The algorithm was doing exactly what it was designed to do: predict who the system would have treated as high-risk in the past. That this reproduced disparate racial impact was a feature, not a flaw, from the machine's perspective. From a constitutional one, it is potentially catastrophic.
More recent tools have attempted to address this by controlling for race in the input data. But this creates a new problem: the algorithms become less accurate overall, because race is correlated with many of the features that genuinely predict recidivism (neighborhood, employment history, education) in ways that are not fully captured by controlling for race as a discrete variable. The result is that debiased algorithms often produce predictions that are simultaneously less accurate and still biased in ways that are hard to detect, because the auditing process itself depends on metrics that can be gamed by system designers optimizing for the wrong objectives. This is not a problem that better engineering will cleanly solve. It is a problem that requires ongoing, independent, adversarial scrutiny — the kind of scrutiny that courts and legislatures have been slow to provide.
Training an AI on historical legal decisions means training it on historical human bias. You do not eliminate discrimination by automating it — you eliminate the ability to detect it.
The Opacity Problem
The due process clause of the Fourteenth Amendment requires that government deprivations of liberty be accompanied by adequate notice and an opportunity to be heard. It also requires, in practice, that defendants be able to confront the evidence used against them. COMPAS and tools like it fail on both counts. A defendant facing a COMPAS score of 7 out of 10 on the risk of general recidivism scale has no meaningful ability to examine the algorithm that produced that number. The inputs — which may include arrest history, employment data, neighborhood characteristics, and dozens of other variables — are not fully disclosed. The weighting of those inputs is a trade secret. The training data is not public. The defendant cannot cross-examine a machine. And a judge who is not an AI specialist may not fully understand what the score means, how it was generated, or what its error rate is for a defendant with this particular profile.
The Loomis dissent captured this precisely. Justice Abrahamson wrote that due process requires more than a defendant standing before a sentencing judge who consults a proprietary risk assessment instrument and then pronounces sentence. It requires that the defendant understand the basis for the court's decision. When the basis is an algorithm, and the algorithm is opaque, that requirement is not met — regardless of whether the algorithm is technically advisory rather than determinative. Courts have not yet fully grappled with what disclosure obligations attach when algorithmic tools are used in judicial proceedings, but the pressure to develop those norms is growing as AI tools proliferate.
Automation Complacency
There is a subtler failure mode that does not get enough attention: automation complacency. When humans have access to AI tools, they tend to trust them more than they should — especially when the tools are presented with confident, numerical outputs. This is a well-documented phenomenon in human factors research. In the legal context, it means that a judge who has been told by a risk assessment tool that a defendant is high risk may give that information more weight than it deserves, because the numerical precision of the output creates an illusion of scientific rigor. A probability of 0.73 looks more objective and harder to question than a judge's gut impression — even if the 0.73 was generated by a model with a known false positive rate of 37% for the relevant demographic group. The solution to this is not to prohibit AI tools but to ensure that everyone who uses them — judges, lawyers, court administrators — understands what they are, what they cannot do, and what their known failure modes are.
The Data: Risk Assessment Algorithm Performance Metrics
The table below presents publicly reported performance metrics for major algorithmic risk assessment tools deployed in U.S. and international criminal justice systems. Figures are drawn from peer-reviewed studies, vendor reports, and court filings where available.
| Tool Name | Developer | AUC-ROC Score | FPR (Black Defs.) | FPR (White Defs.) | Disparity Ratio | Jurisdictions | Audit Status |
|---|---|---|---|---|---|---|---|
| COMPAS (General Recidivism) | Equivant (Northpointe) | 0.68 | 44.9% | 23.5% | 1.91x | Multiple U.S. States | Disputed |
| COMPAS (Violent Recidivism) | Equivant | 0.64 | 40.2% | 19.4% | 2.07x | WI, CA, OR, NY | Disputed |
| Public Safety Assessment (PSA) | Arnold Foundation | 0.65 | 28.1% | 21.7% | 1.29x | 40+ U.S. States | Independent |
| LS/CMI (Full) | MHS / Multi-Health Systems | 0.71 | 31.3% | 22.8% | 1.37x | U.S. and Canada | Independent |
| OREM (Hawaii) | U. Hawaii Crime Prevention | 0.72 | 24.8% | 18.3% | 1.35x | Hawaii (State) | Independent |
| Level of Service Inventory-R | MHS | 0.67 | 38.7% | 20.9% | 1.85x | Canada, U.S., EU | Partial |
| ChinaCourtAI (Criminal) | China Supreme People's Court | 0.89 | N/A | N/A | N/A | China (National) | No Public Audit |
| Hartmann Score (Netherlands) | Research and Documentation Centre | 0.74 | N/A | N/A | ~1.1x | Netherlands | Government Audit |
| OVP (Northpointe Pretrial) | Equivant | 0.70 | 41.5% | 24.1% | 1.72x | Federal + Multiple States | Partial |
| STATIC-99R | PP-S (Registered) | 0.75 | 27.2% | 22.1% | 1.23x | U.S., UK, Australia | Independent |
Vendor-reported figures have not been independently verified for all tools. N/A: Not publicly disclosed. Disparity ratio = Black FPR divided by White FPR. AUC-ROC: Area Under Receiver Operating Characteristic Curve (0.5 = random; 1.0 = perfect).
Discovery on Steroids: E-Discovery Platforms and Their Numbers
The e-discovery market — the segment of legal technology focused on managing the document review process in litigation — is where AI has probably created the most dollar value for clients and the most anxiety for law firm associates. The market, valued at approximately .4 billion globally in 2023, is projected to exceed billion by 2030, according to a 2024 report by Grand View Research. The growth is driven almost entirely by the exponential increase in electronically stored information (ESI): emails, Slack messages, Teams recordings, metadata from enterprise software, text messages on company-issued devices. A single large commercial lawsuit can now involve the review of tens of millions of documents. The economics of doing that with human reviewers at traditional billing rates are, frankly, unsustainable — which is why clients increasingly demand that firms use AI tools, and why firms that refuse to adopt them are losing business to those that have.
The leading platforms in the space have taken very different approaches to AI. Relativity, the dominant enterprise e-discovery platform, offers a suite of AI-assisted features including concept search (finding documents semantically similar to a query rather than just keyword-matched), clustering (grouping documents by topic automatically), and continuous active learning (a process in which the system identifies documents most likely to be relevant and presents them to reviewers for tagging, then uses those tags to refine its model iteratively). Everlaw, backed by Andreessen Horowitz, has focused on collaboration features and has been particularly aggressive in deploying large language models for document summarization and privilege log generation. Logikcull has built its entire platform around ease of use, targeting smaller firms and in-house teams that do not have the infrastructure to run enterprise-grade review platforms.
The numbers on efficiency are striking. A 2023 study by the Electronic Discovery Reference Model (EDRM) found that AI-assisted document review reduced the total review time by an average of 61% compared to traditional linear review, while maintaining or improving accuracy rates as measured by recall. For a case involving 5 million documents, that represents a reduction in review time from an estimated 2,500 reviewer-hours to fewer than 1,000, and a cost reduction from approximately .5 million to under ,000 at standard associate billing rates. These are not trivial savings. They are transformative for clients and deeply disruptive for the law firms that built their revenue models on the inefficiency of the document review process. RelativityOne, the cloud version of the platform, now processes over 1.5 billion documents per year across its global deployment. Everlaw was used by the Department of Justice in the investigation of the January 6th Capitol riot to process over 1 million documents, a scale that would have been practically impossible to manage with traditional review methods within the investigative timeline.
The Big Four accounting firms — Deloitte, PwC, KPMG, and EY — have all launched AI-powered legal services divisions that compete directly with law firms on document review, contract analysis, and regulatory compliance. Their advantage is not just technology but the ability to scale rapidly across geographies. By 2025, PwC's legal AI practice had processed over 300 million documents for clients globally, a volume no single law firm could match.
Judicial AI: Who Is Watching the Watcher?
There is a question that courts, legislatures, and bar associations are only beginning to take seriously: who is responsible when an AI system used in a legal proceeding makes a mistake? The answer matters enormously, because it determines the incentive structure that will drive adoption, accuracy improvement, and accountability. And the answer, for the moment, is: complicated.
The prevailing view among courts in the United States is that AI tools used in litigation are analogous to other expert tools — like forensic accountants or survey experts — and are subject to the same standards of admissibility and reliability. Under Daubert v. Merrell Dow Pharmaceuticals, expert testimony is admissible if it is based on sufficient facts, is the product of reliable principles and methods, and those principles and methods have been reliably applied to the facts of the case. Algorithmic outputs can, in principle, satisfy Daubert if the algorithm's methodology is disclosed, its error rate is known, and its application to the specific case has been documented. But for proprietary algorithms like COMPAS, this standard is almost impossible to meet, because the methodology is not fully disclosed. Courts have responded to this by treating algorithmic risk scores as one factor among many rather than as expert testimony in the strict sense — which is the approach the Loomis court took — but this creates a gap in accountability that no one has yet filled.
In the European Union, the AI Act — which took full effect in stages beginning in 2024 — classifies AI systems used in judicial decision-making as high-risk applications subject to stringent requirements. These include mandatory transparency about the AI's functioning and decision logic, human oversight requirements, registration in a public EU database, and bias testing. The Act also creates a right for individuals to request explanations for AI-assisted decisions that affect them. Whether these requirements will meaningfully improve accountability or will be complied with superficially while the actual decision-making remains opaque is a question that will take years to answer. What is clear is that the EU has decided that algorithmic judicial tools require a different and more rigorous governance framework than currently exists in the United States, and that decision is already influencing how other jurisdictions approach the problem.
The Attorney Competence Problem
Beyond the constitutional questions, there is a bar licensing problem that is only beginning to register. The ABA's Model Rule 1.1 on competence was amended in 2022 to explicitly include technological competence as a component of the duty of competence. Several state bars have followed. This means that attorneys who represent clients in criminal proceedings, and who do not understand the risk assessment tools that may affect their clients' sentences, may be facing a competence claim. Attorneys who represent corporate clients in litigation and who do not understand the AI tools available for document review may be similarly exposed. This is not a marginal concern: the malpractice implications are significant, and the gap between the standard of care and actual attorney knowledge of AI tools is, by most accounts, enormous.
What Lawyers Who Adapt Will Gain
The picture is not uniformly dark. Lawyers who have genuinely engaged with AI tools — not just adopted them because clients demanded it, but understood them well enough to deploy them strategically — are reporting significant competitive advantages. Here is what that looks like in practice.
Research at Machine Speed
Legal research has always been a bottleneck. Junior associates spend hundreds of hours searching Westlaw and LexisNexis for relevant precedent, often missing key cases because they did not know to search for the right terms. AI-powered legal research tools — particularly large language model-based systems like Casetext CoCounsel, Harvey, and LexisNexis's own AI products — do not just search for keywords. They understand the structure of legal arguments, identify cases that are factually analogous even when they are not textually similar, and can summarize the holdings and reasoning of dozens of cases in seconds. A task that used to take a team of associates a week can sometimes be accomplished in a few hours with AI assistance. This does not eliminate the need for human legal analysis — it changes what that analysis looks like, shifting the work from finding information to evaluating and applying it.
Contract Analysis and Drafting
In the transactional world, AI tools are transforming contract review and drafting. Platforms like Kira Systems, LegalSifter, and LawGeex can review contracts for specific risk clauses, identify deviations from playbooks, and flag non-standard terms in a fraction of the time it takes a human reviewer. For M&A due diligence, where review of hundreds or thousands of target contracts is a critical bottleneck in deal timelines, AI has become practically indispensable at large firms. Drafting has followed. AI systems trained on large volumes of high-quality contracts can generate first-draft term sheets, NDAs, and routine agreements that require far less human editing than first-generation template-based tools. The output is not final-draft quality — no AI system yet reliably produces legal prose that requires zero human review — but it dramatically reduces the time to first draft, which is often the biggest time sink in routine matters.
Predictive Analytics for Litigation Strategy
Perhaps the most commercially significant development is the emergence of litigation prediction tools. Platforms like Lex Machina, Westlaw Edge's Litigation Analytics, and a growing ecosystem of specialized tools use historical case data — judge rulings, jury verdicts, settlement outcomes, motion success rates — to generate predictions about how specific cases are likely to resolve. A lawyer facing a motion to dismiss in front of a specific judge can look up that judge's historical ruling on similar motions. A plaintiff's attorney can estimate the probability of success at trial in a specific venue based on how similar cases have resolved. These tools do not predict the future with certainty — no tool does — but they provide a quantitative basis for strategic decisions that used to rest entirely on experience and intuition. For clients making decisions about whether to settle, litigate, or dismiss a case, that quantitative basis is increasingly expected rather than appreciated as a bonus.
- Casetext CoCounsel — Acquired by Thomson Reuters for million in June 2023, representing one of the largest AI acquisitions in the legal space to date.
- Harvey AI — Raised million in Series B funding in January 2024 at a valuation of million, serving Am Law 100 firms including Allen & Overy and Freshfields.
- Kira Systems — Has reviewed over 1 billion pages of contracts for clients including major private equity firms and Fortune 500 companies since 2015.
- Everlaw — Raised million in Series D funding in 2022; used in the investigation of the January 6th Capitol riot to process over 1 million documents.
- Relativity — Processes over 1.5 billion documents per year across its platform; the world's largest e-discovery enterprise deployment.
The Road Ahead: Regulation, Reform, and Realistic Expectations
The legal system's relationship with AI is at an inflection point. For the past decade, adoption has been driven primarily by market forces — clients demanding efficiency, law firms competing on technology, courts experimenting with tools that promise to reduce backlog. Regulation has been reactive, fragmented, and often inadequate. But the pressure for coherent, forward-looking governance is building from multiple directions simultaneously, and the next five years are likely to produce the first major wave of legal and regulatory frameworks that will define how AI operates in the legal system for the next generation.
What Regulators Are Doing
In the United States, the Federal Judicial Center has established an AI Working Group that is developing guidance for federal courts on the use of AI in judicial proceedings. The group has published educational materials and is considering recommending minimum standards for the disclosure of AI-assisted analysis in court filings. At the state level, the New York State Bar Association issued a formal opinion in 2023 requiring attorneys to disclose the use of AI in court filings and to verify AI-generated work product for accuracy — a position that is being closely watched by bars in other states. The EU AI Act, as noted above, is the most comprehensive regulatory framework in existence, but its application to the legal sector in practice is still being worked out through guidance from the European AI Office.
The challenge for regulators is that the technology is moving faster than the regulatory process. By the time a framework is finalized for a specific type of AI tool, the tool itself may have evolved significantly. The EU's approach — defining categories of risk and requiring transparency, auditing, and human oversight rather than banning specific applications — is more flexible than outright prohibitions, but it still faces the fundamental problem that meaningful auditing of complex AI systems requires expertise that most regulatory agencies do not yet have. Building that expertise, and creating the institutional capacity for ongoing oversight, is a multi-year project that governments are only beginning to invest in seriously.
What Courts Need to Do
Beyond formal regulation, there are practical steps that courts can and should take regardless of what legislatures do. The most important is developing clear protocols for the disclosure and auditability of AI tools used in judicial proceedings. At a minimum, these protocols should require: (1) disclosure to the parties when AI-generated information is used in a proceeding; (2) a mechanism for parties to challenge the inputs, methodology, or outputs of AI tools; (3) mandatory disclosure of known error rates and demographic performance differentials for any risk assessment tool used in sentencing or bail; and (4) ongoing independent auditing of AI tools deployed in the court system, with results made publicly available. None of these requirements are technically difficult to implement. The obstacle is institutional will and the political difficulty of appearing to second-guess judges' use of available tools.
We are asking courts to make life-altering decisions — who goes free, who goes to prison, who gets their children, who keeps their home — based on tools that even the people who built them do not fully understand. That is not a legal system. That is a liability.
— Prof. Sandra Mayson, University of Pennsylvania Law School, Automation and the Constitution, 2024
What Lawyers Need to Do
Lawyers, for their part, need to develop genuine technological literacy — not just familiarity with specific tools, but an understanding of how AI systems work, what they can do, and where they fail. This is not a comfortable position for a profession built on expertise and authority, but it is the position the current moment demands. The lawyers who will thrive are not those who resist AI, and not those who blindly adopt it, but those who understand it well enough to use it strategically, supervise it rigorously, and explain it clearly to clients, courts, and opposing parties. That requires training, experimentation, and a willingness to be wrong — qualities that are not uniformly distributed in a profession that has historically valued certainty above almost everything else.
Conclusion: The Question Is No Longer Whether — Only How Fast
The question that framed this article — whether AI has changed how courts work — has a clear answer. It has. The more interesting and more consequential questions are the ones that remain unanswered: How will the legal system manage the risks that come with that change? How will it distribute the benefits equitably, so that AI-driven efficiency gains do not accrue primarily to wealthy litigants and well-resourced firms while everyone else navigates a more complex system without the tools to manage it? How will it preserve the values that make legal adjudication about more than just outcome prediction — values like reasoned justification, procedural fairness, human dignity, and the right to be heard by another human being who takes you seriously?
These are not questions that technology can answer. They are questions that courts, lawyers, regulators, and citizens need to answer — with urgency, with rigor, and with an honest acknowledgment that the legal profession's traditional instincts for caution and continuity are both a safeguard and a liability. The AI systems are not going away. The question is whether the institutions that govern them will evolve fast enough to keep the technology's power in meaningful check, or whether that power will continue to outpace the governance structures designed to constrain it. The verdict on that question has not yet been rendered. But the clock is running, and it is not pausing for anyone to catch up.