Generative AI tools such as ChatGPT, Claude, Gemini, Copilot and legal products built on them (Harvey, CoCounsel, Lexis+ AI and others) are built on large language models (LLMs). An LLM has read a huge amount of text and learned to predict the next words. Ask it a question and it writes the most likely-sounding answer.
That one fact explains both the promise and the risk for legal work.
What it is good at
- Reading and summarizing long documents fast: contracts, deposition transcripts, board papers, regulations.
- Finding and comparing clauses, issues and patterns across many documents.
- First drafts of emails, memos, policies and contract mark-ups that a lawyer then edits.
- Explaining a topic in plain words, or switching tone and language.
What it is not
- Not a database of the law. A general chatbot doesn't look up cases; it writes text that looks like a citation. Legal research tools connect the model to real sources (this is called grounding or retrieval), which helps but doesn't make it perfect.
- Not a lawyer. It has no duty of care, no professional judgment and no privilege. You stay responsible for anything you rely on.
- Not private by default. What you type goes to a provider. Whether it is stored, reviewed or used for training depends on the product and the contract.
Three words you will hear
- Prompt: the instruction and material you give the tool.
- Hallucination: a confident answer that is false, such as an invented case.
- Enterprise vs consumer: business versions usually come with contract terms on confidentiality and no training; free consumer versions often don't.
The right mental model: a very fast, very well-read junior who never says "I don't know", and whose work you must check before it leaves the building.