- Hallucination
- An output of a generative model that is fluent and wrong, such as a citation to a case that does not exist.
- Automation bias
- The tendency to accept an automated output without the scrutiny you would give the same work from a person.
- Shadow AI
- AI tools used for work without the organisation's knowledge or approval.
- Retrieval-augmented generation
- A technique in which a model answers from documents retrieved for the question, not from its training alone.
- Agentic AI
- AI that plans and carries out steps towards a goal with limited human input.
- Legal-specific model
- A model built or tuned for legal work, usually on legal sources, as opposed to a general-purpose model.
- Context window
- The amount of text a model can take into account at once.
- Prompt library
- A shared set of tested prompts for recurring tasks.
- Guardrail
- A technical or procedural control that limits what an AI tool can do or produce.
- Model tiering
- Sorting AI tools by risk, so that each tier carries its own approval and checks.
- AI inventory
- The register of AI tools in use, with their owner, purpose and risk.
- Technology-assisted review
- The use of machine learning to rank or classify documents in e-discovery.
- Elusion testing
- Sampling the documents a review set aside, to estimate how many relevant ones were missed.
- Electronically stored information
- Information in electronic form that may have to be disclosed in litigation.
- Privilege log
- The list of documents withheld on grounds of privilege, with the reason for each.
- Unauthorised practice of law
- Providing legal services without authorisation, a risk when AI tools advise clients directly.
- Legal engineer
- A professional who designs and builds the technology and processes for legal work.
- Data sovereignty
- The principle that data is subject to the laws of the country where it is stored or processed.
