For most of the profession’s history, legal research began and ended with a human reader: a lawyer found the relevant cases and statutes, read them closely and – in the case of a litigator – wrote for a judge who would, in turn, read the result. That loop has now changed because large language models (LLMs) read, synthesise, and reason over legal text alongside lawyers, summarising judgments, surfacing patterns across thousands of documents, and drafting analysis in seconds.
This shift is valuable, but it also changes where lawyers add value. As machines take on more of the reading and finding, the lawyer’s role increasingly becomes one of judging – deciding what to trust, what to verify, and how to write for an audience that now includes both people and machines. Legal teams that want the benefits of AI-driven research without eroding the quality of their work should focus on three practical steps to adapt to this new reality.
Step 1: Make adjustments for a new “dual-reader” reality
Now that machines can read, legal documents serve two very different audiences. Writing for human readers is built around narrative and context, with judges expecting advocates to tell a story, establish background, and gradually build towards their analysis – a convention that often places the substance in the middle of a lengthy document. LLMs, however, tend to overweight the beginning and end of a document while underweighting the middle, which is precisely where the legal analysis sits.
Abandoning narrative to suit the machine is no answer, since it would alienate judges and undermine an advocate’s credibility – and judges themselves write for the people their decisions affect while knowing those decisions will become case law that machines reason over.
Adapting to this dual-reader system also means accepting that LLMs are far from flawless readers. The word “negligence”, for example, carries one meaning in a criminal context, another in a civil context, and no particular legal meaning in everyday use. An LLM trained on broad internet data may conflate all three, producing a polished synthesis that knits together concepts which do not belong together. That is not a hallucination in the usual sense, but it does not reflect how the law operates, which is why lawyers need the judgment to spot when something looks wrong.
Pattern recognition carries a similar risk. While LLMs can find patterns no human researcher could, those patterns may be irrelevant, may be built on conflated information, or may fail to account for how human intervention changes outcomes in the data – and deciding whether they matter remains a human judgment.
On a practical level, applying heading styles consistently creates a machine-readable hierarchy that allows an LLM to navigate straight to the analysis rather than leaving the lawyer to hope it finds the most germane material. Additionally, defining more terms supplies the clarity machines cannot infer from context. The plain language movement has long favoured fewer defined terms, but specifying, for instance, that “negligence” refers throughout a document to the civil standard gives machine readers exactly the disambiguation they need.
These near-term steps will buy some time, but in the long term, what’s needed is more system-wide classification, ontologies, taxonomies, enrichment and shared legal standards that are shared across the legal ecosystem.
SALI and NOSlegal provide good starting points. Firms that build their internal taxonomies from this type of shared baseline create systems that can map to one another, creating a more unified approach as the profession moves increasingly into a machine-read and machine-assisted world.
The broader challenge is to give AI tools a reliable place to search and a trusted source of context. This is where a semantic context layer becomes important. It does not define legal concepts itself; instead, it provides additional background and connections so that when LLMs reason over legal writing, they focus on the right material.
With a context layer in place, the resulting work product is more likely to rely on the right material, make sense to human readers, and remain usable by machines.
Step 2: Build the judgment muscle
In this AI-driven era, lawyers now need well-developed judgment that tells them when to dig deeper and when a tool is likely to be reliable.
Part of that judgment comes from understanding the technology: because generative AI relies on statistical relationships between words in its training data, a recent decision may barely feature in that training data. A landmark ruling discussed for a decade is a different proposition, and knowing what to expect in each case is nearly half the battle. The other part comes from practice-area expertise, and the “negligence” problem from earlier illustrates it well, since an experienced lawyer should recognise when AI output is grammatically sound and confidently expressed, yet legally wrong.
Firms can strengthen this judgment muscle by using an “engagement matrix” that places low-risk uses of curated AI tools – such as spellchecking an email – in a safe zone while flagging high-risk uses of uncurated tools – such as drafting a transaction document with a free chatbot – as requiring far greater scrutiny.
Lawyers should also keep AI-generated material separate from their working document and, whenever they import something, add a comment explaining why they trust it – something like “I know this is correct because I cross-referenced it against two primary sources”. A lawyer who cannot supply the “because” does not yet understand the material well enough to rely on it.
Step 3: Incorporate accountability by design
Accountability has traditionally been bolted onto the end of the process, with lawyers doing the work, auditing it, and recording the audit – but that structure no longer holds up when AI produces the first draft.
Consider a litigator drafting the factual background of a submission, where the narrative states that a payment of £3,510 arrived on a particular date. In the paper era, the risk of a figure being fabricated by an automated drafting tool would not have arisen, yet once AI drafts the sentence, the amount, the sender, and the date all need confirming against the source, ideally with a link to it. The same principle applies to legal conclusions, so even if an AI tool correctly finds that a cure period has expired and enforcement remedies are available, confirming that analysis remains a professional obligation.
Accountability by design also means that verification happens as part of the work and is recorded where the next person can see it. When an associate checks a fact and initials it, the supervising lawyer can trust that point and move on. Without such a structure, AI creates what has been called the verification-value paradox, in which several lawyers re-check every line for invented authorities until the promised efficiency disappears.
Good handoffs matter just as much, whether work passes from associate to partner or from lawyer to AI agent. Lawyers often ask colleagues or AI tools to “just do X” without explaining what needs to happen first, whereas a proper handoff sets out the goal, the sources needed, and how to resolve questions that arise.
This is the essential difference between delegating and assigning. Delegating means handing a task off with enough context to do it well and a clear picture of what “done well” looks like, whereas assigning means handing it off and hoping for the best. If a lawyer cannot judge whether a task was done well, they have assigned it rather than delegated it.
Judgment is the real value-add
AI-driven research is not making professional judgment less important – if anything, it is placing judgment at the centre of the lawyer’s value. As machines take on more of the reading, finding, and summarising, the lawyer’s contribution increasingly lies in writing for both audiences, testing what the machine produces, and being able to stand confidently behind the finished work. In the end, that judgment is what earns a client’s trust.
.jpg&w=3840&q=75)
Copyright & permissions
