From Westlaw to GPT: A Brief History of Legal Technology

TL;DR

• The lag is the pattern. Legal tech has followed general technology on a 10–20 year delay for half a century — Westlaw arrived two decades after the first legal information retrieval experiments, e-discovery platforms a decade after businesses went digital.
• The economics drove adoption; the courts get the credit. Zubulake, Da Silva Moore, and Mata v. Avianca are the datable markers, but firms were already moving — the rulings ratified and accelerated a shift that changing economics had already made inevitable. The durable trigger is when not adopting costs more than adopting.
• The cycle is compressing from decades to months. The database revolution took ~20 years from research to product. E-discovery took ~10. Cloud practice management took ~5. LLM-powered legal tools took ~2. The next wave won’t wait for a six-month vendor evaluation.
• Today’s AI debates are yesterday’s e-discovery debates. Data security, accuracy, vendor lock-in, and professional responsibility — the legal profession has worked through every one of these concerns before, with the same initial resistance and the same eventual accommodation.

In 1975, the year West Publishing launched Westlaw, you could also buy an Altair 8800 — the first commercially successful personal computer. In 2004, when courts were grappling with whether companies had a duty to preserve electronic documents, Google went public. In November 2022, when OpenAI released ChatGPT, most law firms were still debating whether to allow Microsoft Teams.

Legal technology doesn’t just follow general technology. It follows it on a remarkably consistent delay — and that delay has been shrinking with each wave. Understanding the pattern is the best tool available for evaluating what’s happening now and what’s coming next.

The Research Era (1950s–1970s)

#

The idea that computers could help lawyers find law predates the personal computer by two decades. Starting in 1959, John Horty at the University of Pittsburgh Health Law Center began developing one of the first legal information retrieval systems; by the early 1960s it could search Pennsylvania statutes by keyword — something that required hours of manual index work. By the mid-1960s, the Ohio State Bar Association had partnered with Data Corporation to build a full-text search system for Ohio case law, the project that would eventually become Lexis.

These were research projects, not products. The hardware filled rooms. The databases covered single states. The query languages required technical training no practicing lawyer had. But the core insight — that computers could search legal text faster and more completely than humans browsing indexes — was sound. It would take another fifteen to twenty years before that insight became something a lawyer could actually use.

The Database Revolution (1973–1995)

#

Lexis launched commercially in 1973. Westlaw followed in 1975. For the first time, lawyers could search case law electronically instead of pulling reporters off shelves and flipping through West’s printed digest system.

The resistance was immediate and predictable. Senior partners argued that electronic search was unreliable — you couldn’t trust a computer to find what a trained associate with a digest system would find. Cost was a constant objection: Lexis terminals were expensive, per-search charges added up, and firms couldn’t easily pass the costs to clients who didn’t understand what they were paying for. Training was a burden. Many lawyers simply refused to learn the systems, relying on associates or librarians to run searches for them.

The objections sound familiar because they’re the same ones firms raise about every new technology, including AI. But the profession adapted. By the late 1980s, Westlaw and Lexis had become essential infrastructure. Law schools taught electronic research. Associates who couldn’t run a competent search were at a disadvantage. The shift from physical law libraries to electronic databases didn’t happen because firms decided it was a good idea. It happened because the alternative — maintaining comprehensive physical libraries and the staff to manage them — became economically irrational once the technology was reliable enough.

The database revolution also established the vendor lock-in model that still defines legal tech. Westlaw and Lexis built proprietary editorial enhancements (West’s Key Number System, Lexis’s Shepard’s Citations) that made switching costly. Fifty years later, both platforms still dominate legal research, and the switching costs are higher than ever. When Thomson Reuters dropped 16% after Anthropic released a legal plugin in February 2026, the market was pricing in the possibility that the lock-in might finally break. It hasn’t — yet.

Research to adoption: ~20 years. Horty’s 1959 experiments to Westlaw’s 1975 commercial launch.

E-Discovery and the Compliance Era (2000–2015)

#

By the late 1990s, businesses ran on email, and litigation was drowning in it. A single lawsuit could involve millions of electronic documents. The legal system’s rules for document production — written for a world of paper files and filing cabinets — had no framework for preserving, collecting, reviewing, and producing electronically stored information (ESI).

The inflection point was Zubulake v. UBS Warburg (S.D.N.Y. 2003–2004). Judge Shira Scheindlin issued a series of opinions establishing that parties had a duty to preserve relevant electronic documents once litigation was reasonably anticipated, and that failure to do so could result in adverse inference instructions and sanctions. Zubulake didn’t create new law — it applied existing preservation obligations to electronic records. But it pushed every large organization to build systems for managing ESI, and it catalyzed the e-discovery industry that took shape over the following years.

The Federal Rules of Civil Procedure amendments of 2006 codified ESI obligations: meet-and-confer requirements for electronic discovery, rules for the form of production, and a safe harbor for routine document destruction. The 2015 amendments tightened sanctions for spoliation and introduced proportionality as a constraint on discovery scope.

Between Zubulake and the 2015 amendments, an entire ecosystem emerged. Relativity (platform launched 2006, market-leading by the late 2010s) became the standard platform for document review. Everlaw and DISCO offered cloud-based alternatives. Predictive coding — using machine learning to prioritize documents for review — went from controversial to court-approved in Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), where Judge Andrew Peck held that technology-assisted review was an acceptable alternative to manual review. The profession’s initial reaction to predictive coding — suspicion, resistance, demands for proof that it worked as well as human reviewers — previewed exactly how it would respond to LLMs a decade later.

The pattern is usually told as courts forcing adoption, but that overstates their role. Firms were already moving — discovery at email scale couldn’t be done by hand at any price a client would pay, so the economics had turned against manual review well before Zubulake. What the courts added was a floor: once failing to manage ESI competently became a sanctionable offense, opting out stopped being an option. The rulings get the attention because they’re dramatic and datable, but the economics were the engine.

Research to adoption: ~10 years. ESI problems emerged in the late 1990s; the 2006 Federal Rules amendments made e-discovery compliance mandatory.

Cloud and SaaS (2010–2020)

#

While large firms were building e-discovery infrastructure, a parallel wave was reshaping how smaller firms operated. Cloud-based practice management platforms — Clio (2008), MyCase (2010), PracticePanther (2012) — moved calendaring, billing, document management, and client communication to the browser. A solo practitioner with a laptop and a Clio subscription could run a practice that would have required a secretary, a filing system, and an office a generation earlier.

The shift was as much business model as technology. Software moved from a licensed product you installed and maintained to a subscription you rented — billed monthly, updated continuously, hosted by the vendor. That stripped out the fixed costs of running a firm’s systems: no servers, no IT staff, no perpetual-license fees. LegalZoom (2001) and Rocket Lawyer (2008) took the same delivery model direct to consumers with standardized documents, which the organized bar fought as unauthorized practice of law before largely accommodating.

No court forced this wave — it was pure economics. Firms that moved to the cloud ran leaner than the ones that didn’t, and the cost advantage compounded until staying on-premises stopped making sense.

Research to adoption: ~5 years. Salesforce proved the SaaS model in the early 2000s; legal-specific cloud platforms were mainstream by 2010–2012.

The LLM Moment (2022–Present)

#

OpenAI released ChatGPT on November 30, 2022. Within two months it had 100 million users — the fastest a consumer application had ever reached that mark, against roughly nine months for TikTok and two and a half years for Instagram. Within a year, every major legal tech vendor had either integrated a LLM into their product or announced plans to do so.

The speed was unprecedented. Harvey — founded in late 2022, built on OpenAI’s models — secured an exclusive partnership with Allen & Overy by February 2023 and raised over $200 million by early 2025. Thomson Reuters launched CoCounsel (originally built by Casetext, which Thomson Reuters acquired for $650 million in 2023) as an AI assistant integrated with Westlaw. Spellbook, EvenUp, and dozens of other startups raised hundreds of millions in venture funding for legal-specific AI applications.

The profession’s Zubulake moment arrived in June 2023, when a federal judge in Mata v. Avianca sanctioned two attorneys who submitted a brief containing six fabricated case citations generated by ChatGPT. The attorneys hadn’t verified the citations. The cases didn’t exist. The court imposed $5,000 in sanctions and required the attorneys to notify the judges whose names appeared on the fake opinions. Mata didn’t create new law any more than Zubulake did — it applied existing competence obligations to a new technology. But it forced every lawyer using AI to confront a specific failure mode: LLMs hallucinate, and the consequences of unverified AI output fall on the attorney, not the tool.

The regulatory response followed quickly. ABA Formal Opinion 512 (July 2024) established that lawyers using AI must understand how the technology works, must supervise its output, and must protect client confidentiality when using third-party AI services. Several state bars issued their own guidance. In February 2026, United States v. Heppner held that a defendant’s conversations with consumer-tier Claude were not privileged — extending the regulatory reckoning from competence to confidentiality. (The privilege and work-product questions AI raises are mapped in Privilege, Work Product, and AI: A 2026 Doctrinal Map.)

The current wave is also the first where the technology itself is writing the next generation of tools. Vibe coding — describing what you want in natural language and letting an AI generate the software — means lawyers and legal ops professionals can build custom tools without engineering departments. A quarter of Y Combinator’s Winter 2025 startups had codebases written almost entirely by AI. The AI Use Spectrum captures the range: from individual lawyers prompting ChatGPT (Level 1) through vibe-coded scripts (Level 3) to enterprise platforms (Level 5). The levels that are hardest to govern — 1 through 3 — are where most actual adoption is happening.

This wave also breaks the market structure of the earlier ones. The database, e-discovery, and cloud platforms were built by legal specialists — West, Mead, kCura, Clio. The LLM layer is built by general-purpose frontier labs that sell into every industry, and they have begun reaching past their legal-tech customers to the firms — and the clients — directly.

That February 2026 sell-off carried a meaning beyond lock-in. Anthropic, a model supplier, had shipped a legal plugin for Claude, and the market read it as the supplier turning into the competitor.

Both leading labs are now staffing the bet. Anthropic expanded that plugin in May 2026 into a branded Claude for Legal line, with Mark Pike — an associate general counsel — serving as its product lead for the legal industry. OpenAI hired Ironclad co-founder Jason Boehmig in June 2026 to build a legal vertical, having long supplied the models under tools like Harvey.

The lawyer-into-tech crossover isn’t new. Brad Smith — now Microsoft’s president — joined Covington & Burling in 1986 and is remembered as the first attorney there to make a personal computer on his desk a condition of taking the job; he deleted the firm’s word processor and installed Microsoft Word in its place. He went on to run Covington’s software practice before leaving for Microsoft in 1993, where he became general counsel and then president. The difference now is scale: the crossover used to be one tech-minded lawyer betting on a software company; today the software companies are hiring the lawyers and building the legal products themselves.

The labs also reach end users with no legal product at all — the lawyer in Mata and the defendant in Heppner were using consumer ChatGPT and Claude directly. The pressure comes from both ends, and the legal-tech middle is what’s exposed.

[Medium confidence] Legal is the vertical the frontier labs target next, because coding showed them the playbook. (Anthropic has already shipped a direct legal product, and coding is the task the labs have most thoroughly productized.)

Code is what LLMs do best: training data is abundant and outputs are cheaply verifiable — a compiler or a test suite confirms in seconds whether an answer works. That tight feedback loop pushed coding quickly into agentic tools and an ecosystem of MCP connectors (the Model Context Protocol, which has since become a cross-vendor standard) that let models act inside real systems.

[Low confidence] Legal AI may rhyme with that arc — multi-step agents, connectors into research databases and document management — but how much of the playbook actually transfers is genuinely unknown. Code has clearer ground truth than law usually does: a failing test is obvious, while a subtly wrong indemnification clause or a hallucinated citation is not, and a silent error in legal work lands on a client and an attorney’s license, not a CI pipeline. Those differences could slow the transfer, reshape it, or matter less than they look. The coding analogy is the best available map, not the territory.

Research to adoption: ~2 years. LLMs became commercially capable with GPT-3 in 2020, and legal-specific AI tools arrived within months of ChatGPT’s November 2022 launch.

What the Pattern Means Now

#

Every technology wave in legal history has followed the same arc: resistance on grounds of reliability, cost, and professional responsibility → shifting economics that make the old way untenable, with courts weighing in along the way — sometimes warning, sometimes sanctioning, always stretching existing doctrine to cover the new technology → grudging adoption → eventual normalization. The court rulings are the visible markers and they draw outsized attention, but they tend to ratify and channel a shift the economics already set in motion. Zubulake didn’t invent the duty to preserve and Mata didn’t invent the duty of competence; in each wave the doctrine expanded to fit — ESI first, then AI — rather than courts forcing a profession that was standing still. Observers over-credit the rulings partly because a licensed, self-regulating profession is assumed not to change on its own, so the shift gets attributed to an external shock rather than to the quieter, steadier pressure of cost. The arguments against Westlaw in 1975 (too expensive, can’t trust the results, lawyers shouldn’t need computers) are structurally identical to the arguments against AI in 2024 (too expensive at scale, can’t trust the output, lawyers shouldn’t rely on machines for judgment).

The difference is the clock speed. The database revolution gave firms twenty years to adapt. E-discovery gave them ten. Cloud tools gave them five. LLM-powered legal tools gave them two — and that window has already closed. Firms that are still running six-month evaluations of whether to adopt AI are operating on an e-discovery-era timeline in an LLM-era cycle. The Thomson Reuters 2025 Future of Professionals Report found that individual professionals were adopting AI faster than their organizations — the same pattern that preceded every prior wave, except now the gap between individual adoption and institutional response is closing in months rather than years.

The firms that navigated each prior wave successfully weren’t the ones that adopted first. They were the ones that understood what the technology actually did, matched it to the right problems, and built governance around reality rather than aspiration. That’s as true for LLMs as it was for Westlaw.

Further Reading

#

• Attention Is All You Need. Vaswani et al. (2017), the Transformer paper that enabled modern LLMs.
• The Illustrated Transformer. Jay Alammar’s visual walkthrough.
• Zubulake v. UBS Warburg. Wikipedia summary of the e-discovery landmark.
• Da Silva Moore v. Publicis Groupe. The predictive coding approval decision.
• Mata v. Avianca. The AI Hallucination sanctions case.
• ABA Formal Opinion 512. ABA guidance on lawyers’ duties when using AI.
• United States v. Heppner. Harvard Law Review analysis of the AI privilege ruling.
• Privilege, Work Product, and AI: A 2026 Doctrinal Map. How privilege and work-product doctrine is adapting to AI use.
• Clio Legal Trends Report. Annual data on how small and mid-size firms operate.
• Thomson Reuters 2025 Future of Professionals Report. Data on professional AI adoption rates.
• LegalTech: SaaSpocalypse Now. The market reaction to Anthropic’s legal plugin.

This post is part of the History of Legal Tech series on LegalRealist AI. It is intended for informational and educational purposes only and does not constitute legal advice. AI capabilities, features, and regulatory developments described here reflect publicly available information as of the publication date and are subject to rapid change. Laws and ethics rules governing AI use in legal practice vary by jurisdiction.