MMOOZZGG.com

AI for lawyers

Legal AI glossary: AI terms for lawyers from neural network to anonymization

6 min read5 sources

This legal AI glossary explains AI terms for lawyers in short: what each word means and why it matters when you work with client documents. The numbers show why you need them: a Stanford RegLab and HAI study (May 2024) found that legal AI research tools were wrong in more than 17% of answers, and a lawyer needs to understand where an error comes from. There are 40 terms in three groups: general, data and confidentiality, and models.

On this page · 3 sections

General terms

Benchmark: a set of test tasks for comparing models. A high test score does not mean the model will handle your firm’s documents as well.

Hallucination: a confident but invented answer from a model, such as a case, a rule or a quote that does not exist. In the study that Stanford RegLab and HAI cite, general chatbots were wrong on 58–82% of legal questions.

Generative AI: AI that creates new text, images or code, and does not only sort and classify data.

Prompt: a task for a model. A good prompt has a role, a task, the source data, the answer format and a rule for checking.

AI agent: a program on a language model that does not only answer but acts in systems: it creates a request in the CRM, books a consultation, prepares a document from a template.

AI assistant: AI that gives a lawyer hints and prepares drafts, while a person performs the action.

Artificial intelligence: technologies that solve tasks that used to need a person: understanding text, search, classification, translation.

Pilot metric: the number used to check AI’s result on a process, such as time to first reply or the share of checklist items marked the same way as by a lawyer.

Neural network: a model of connected mathematical nodes that learns from examples. Lawyers often use the word for any AI tool.

Pilot: a test of AI on one process, over a short period and with a metric written into the contract.

Citation check: checking each rule and court ruling in an AI answer against the original source before the document goes to a client or a court.

System prompt: standing rules for the model that the user does not see: what to answer, what is forbidden, when to pass a question to a lawyer.

Chatbot: a program for chatting. Unlike an AI agent, it usually only answers and does nothing in the firm’s systems.

Human in the loop: a rule under which a person, not the model, confirms a decision, a signature and the sending of a document.

Data and confidentiality terms

Attorney-client privilege: under article 8 of Russia’s Federal Law on Advocacy, it covers any information related to legal help given to a client; only the client can release the attorney from it.

Prompt injection: a hidden instruction in an email or document that tries to make an AI agent break its rules, for example send a client list. Protection: the agent does not follow commands from incoming messages and passes such a request to a person.

Action log: a record of each action of an AI agent in the system where it works. The log shows who changed what and when.

Closed network: the model, the data and the log work inside the firm’s network with no connection to the outside.

Trade secret: a confidentiality regime that the owner of information sets up. Its terms are worth moving into the contract with the AI vendor.

Data localization: recording and storing personal data of Russian citizens in databases inside Russia, part 5 of article 18 of 152-FZ, Russia’s personal data law.

Depersonalization: under clause 9 of article 3 of 152-FZ, actions after which, without extra information, you cannot tell which person the personal data belongs to.

Personal data operator: the party that organizes processing and sets its purposes (clause 2 of article 3 of 152-FZ). For client documents, the firm is the operator, and the AI contractor processes them on its behalf.

Personal data: under clause 1 of article 3 of 152-FZ, any information related to a directly or indirectly identified or identifiable individual.

Processing assignment: a contract under which the operator passes processing to another party, part 3 of article 6 of 152-FZ. The operator answers to the data subject for that party’s actions, part 5 of article 6.

Role-based access: an AI agent’s rights are limited to its process. It has a separate account, with no database export and no admin rights.

Cross-border transfer: sending personal data abroad, including to a foreign AI service. Since 1 March 2023, the operator must notify Roskomnadzor, the Russian data protection regulator, before it, part 3 of article 12 of 152-FZ.

Leak through AI: information reaches the service vendor or third parties because an employee pasted it into a prompt.

Model terms

API: access to a vendor’s model over the internet. The prompt text goes to the vendor’s servers, so it matters where they are and whether there is a data processing agreement.

Vector search: search by meaning, not by words. Document fragments become numeric vectors, and the system finds passages close in meaning even when the wording differs.

Fine-tuning: extra training of a model on your own examples. For a law firm, retrieval-augmented generation is usually enough, and fine-tuning is needed for the style and format of documents.

Quantization: compressing a model, for example to 4 bits per number, so it fits into the memory of your own server. Quality drops slightly.

Context window: how much text a model sees in one request. A long case may not fit whole, and then it is split into parts.

Local model: a model on the firm’s server or in a cloud in Russia, with no text sent to an outside vendor.

Open model: a model with published weights that you can download and run on your own hardware.

Model parameters: the numbers inside a model, set during training. Their count affects quality and the memory a server needs.

Retrieval-augmented generation, RAG: the model first finds fragments in the firm’s document base and then answers from them with a link to the source. Such an answer is easier to check.

Reasoning model: a model that goes through a chain of intermediate steps before it answers. It helps with complex analysis, but it answers more slowly and costs more.

Temperature: a setting for the randomness of an answer. For legal tasks it is set low, so the model answers the same question the same way.

Token: a part of a word that a model uses to count text volume and the price of a request. Russian text usually takes more tokens than English text of the same length.

Language model, LLM: a neural network trained on large amounts of text to predict what comes next. Chatbots, AI assistants and AI agents run on it.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Stanford HAI, 23.05.2024 (hai.stanford.edu)
  2. Great Russian Encyclopedia (bigenc.ru)
  3. Brace Law Firm (brace-lf.com)
  4. ConsultantPlus, 152-FZ “On Personal Data”, article 3, 27 July 2006 (consultant.ru)
  5. ConsultantPlus, 152-FZ “On Personal Data”, article 6, 27 July 2006 (consultant.ru)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us