■ HIGH RISK ■ Legal
The core of this job — finding and summarizing relevant law — is precisely what large language models do best, and legal-tech vendors know it. What survives is verification, strategy, and knowing when the AI is confidently citing a case that doesn't exist.
“AI reads case law at the speed of light. You read it at the speed of caffeine.”
Our AI replacement risk score — how we score jobs
Legal researchers live inside Westlaw and Lexis: translating an attorney's half-formed question into search strings, reading dozens of opinions to find the three that matter, checking that each case is still good law, tracing how a doctrine split across circuits, and packaging it all into a memo the attorney will skim in four minutes. It is skilled, meticulous work — and it is almost entirely reading, retrieval, and synthesis of text.
That's the bullseye for generative AI. The major research platforms have already embedded AI assistants that take a natural-language question and return a synthesized answer with citations, collapsing hours of Boolean searching into a prompt. Citator checks are automated. First-draft research memos come out of the machine in minutes. Firms that once staffed research through junior associates and dedicated researchers are discovering that one person plus an AI produces what five used to — and clients, who always resented paying for research hours, are pushing firms in exactly that direction.
The residue is quality control and judgment, and it's not trivial. AI legal tools still hallucinate authority — courts have sanctioned lawyers for filing briefs with invented citations, which made verification a survival skill rather than a formality. Someone must confirm every cite, catch the subtly-wrong characterization of a holding, and understand the client's strategic posture well enough to know which accurate answer is the useful one. Novel questions with thin precedent, legislative-history archaeology, and cross-jurisdictional nuance resist the pattern-matching. Our 74 reflects a role collapsing into a smaller, more senior verification-and-strategy function.
Automatability: our editorial assessment of current and near-term AI capability
This one isn't waiting for 2030 — AI research assistants are already deployed inside the major platforms and firms are actively rethinking how many research hours they bill. Expect steady contraction of dedicated research roles through the decade, with the surviving positions redefined around verification, complex matters, and supervising the tools. The billable research memo as a junior rite of passage is ending now.
The retrieval-and-summarize core, largely yes — that's already happening inside Westlaw and Lexis. But entirely, no: courts have sanctioned lawyers over AI-invented citations, so every AI research product now depends on humans who verify output and frame questions strategically. The job shrinks and moves up the judgment ladder rather than vanishing.
As a standalone career, it's risky — the hours clients will pay for pure research are collapsing. As a skill inside a broader role, it's still essential. The durable path pairs research expertise with verification, knowledge management, or a substantive specialty where you interpret law rather than just locate it.
Good enough to be useful, wrong enough to be dangerous. They excel at surfacing relevant authority fast, but they can mischaracterize holdings and occasionally cite cases that don't exist. Treat output as a first draft from a fast, overconfident junior: verify every citation and read the key cases yourself before anything is filed.
Learn the AI tools your firm licenses better than the attorneys do, and document where they fail. Volunteer to run the verification workflow. Build expertise in research the models handle badly — legislative history, agency materials, unpublished decisions. Being the human who makes AI research trustworthy is the defensible position.