Introduction to legal research using AI

What AI does well

Language models are effective at translating a lay description of a problem into legal issues, at suggesting the vocabulary a court would use, and at proposing which statutes and areas of law are likely to be engaged. They are good at orientation in an unfamiliar field and at producing a structured first draft of a research plan.

Where a model is connected to a real document library — as LEX ZED's assistant is — it can also locate provisions and judgments and summarise long documents you have already decided to read. That is a genuine saving of time.

What AI does badly

A language model predicts plausible text. Plausible citations are therefore a structural failure mode, not an occasional bug: a fabricated case name, a real case name attached to the wrong proposition, or a section number that does not exist. Models are also weak on currency — they will happily discuss a repealed provision — and on jurisdictional discipline, drifting into English, Kenyan or South African authority where Zambian authority was asked for.

The correct posture is that an AI answer is a hypothesis about where the law is, never a statement of what the law is.

A workflow that works

Use the assistant to frame issues and identify candidate authorities. Then open each authority in full and read it. Then verify currency against the legislation. Then write your own analysis from the primary materials, using the AI output only as a checklist of things you might have missed.

Never copy an AI-produced passage containing a citation into a document without having opened the cited authority. This single rule prevents nearly every reported instance of AI-related professional embarrassment.

How LEX ZED is designed around this

LEX ZED's assistant is grounded in a library of Zambian judgments and legislation and is instructed to refuse rather than speculate when a judgment's full text is not available to it. It shows its sources and links to the authoritative document where one exists. It is built to shorten the search, not to replace the reading.