Chapter 1 Introduction
Lesson 1.3:Survival Glossary (in Plain Words)
AI comes with its own vocabulary. Here are the most common words you'll meet in this course, in vendor demos and in your IT team's emails, each with a plain definition, a concrete example and an illustration.
Skim them now, then come back whenever a word trips you up.
The model
LLM / model
A large language model (LLM) is an AI system trained on an enormous amount of text to guess which word comes next. Doing that guess over and over is how it writes answers. GPT, Claude, Gemini and Mistral are families of these models.
Example. Type 'The tenant shall give written…' and the model continues with 'notice', because that is what usually comes next in the text it was trained on. It didn't check any lease: it guessed, and here the guess happens to be right.
The tenant shall give written…
- notice 64%
- consent 21%
- reasons 9%
Token
The syllables of a model: text is read and written in small pieces, often a whole word, sometimes part of one. In English, 100 tokens is roughly 75 words.
Example. A 40-page lease is about 20,000 words, or roughly 27,000 tokens. Ask for a two-page summary and the model writes about 1,300 more. Both count toward the model's limit and your bill.
Indemnification clause
Indemnificationclause
2 words · 4 tokens
Prompt
Everything you send the model in one go: your question, your instructions and any document you attach. The model knows nothing else about your situation, so whatever you leave out, it guesses.
Example. 'Summarize this' gives a vague page. 'Summarize this lease in five points for a non-lawyer, focusing on the tenant's obligations' gives exactly what's needed, from the same document.
Context window
The model's working memory: the maximum amount of text it can take in at once, measured in tokens. Your instructions, your documents, the conversation so far and its own replies all fill it up. Once it's full, nothing more fits.
Example. Upload a 2,000-page data room and the tool will refuse it or only take in part of it: it doesn't fit in the context window. You have to pick the documents that matter, or use a tool that searches them for you.
- Instructions
- Your documents
- Conversation so far
Full? Nothing more fits.
Hallucination
When a model states something false or made up, such as a court case that never existed, in the same confident tone it uses for things that are true. It isn't lying; it is guessing, and it doesn't flag the difference.
Example. In Mata v. Avianca (2023), two New York lawyers filed a brief citing six cases ChatGPT had invented, complete with realistic-looking citations. The court fined them.
Harlow v. Meridian Holdings
758 F.3d 412 (2d Cir. 2014)
- Right format
- Real court and reporter
- No such case
Benchmark
A standard test used to score and compare models, a bit like an exam league table. Useful for a rough ranking, but a high score doesn't prove the model will do good work on your matters.
Example. A model can top a public leaderboard and still miss the break clauses in your leases. So we tested the shortlist on 20 of our own files before choosing.
- Model A 86%
- Model B 79%
- Model C 64%
Your contracts aren't the test.
The tools around it
Chatbot
An app that lets you talk to a model in a chat window, like ChatGPT, Claude or Le Chat. You type, it replies, and each reply takes the earlier messages into account. It answers; it doesn't act on your behalf.
Example. Ask whether a landlord can refuse to renew a lease, and the chatbot asks which country. Answer 'France', and every reply after that assumes French law without you repeating it.
Can the landlord refuse to renew?
It depends on the lease and the governing law. Which country?
France.
RAG
Retrieval-augmented generation. Before answering, the tool searches a chosen set of documents and hands the model only the relevant passages, so the answer rests on them and can cite them, rather than on memory alone. If the search misses a passage, so does the answer.
Example. Ask 'When can we terminate?' and the tool first finds §12 of the lease and the side letter in your files, then answers citing both, so you can check every point against the source.
- 1“When can we terminate?”
- 2 Lease §12Side letter
- 3Answer based on them, citing [1] [2]
Skill
A set of written instructions, saved once, that teaches an assistant how to do one task the way you want it done. The assistant picks it up on its own when the task comes up. Think of a recipe card, sometimes with templates or small scripts attached.
Example. Write your team's NDA review checklist into a skill once. From then on, whenever anyone asks the assistant to review an NDA, it follows that checklist: term, scope, unusual clauses, in your order.
nda-review/SKILL.md
## When to use
Reviewing an incoming NDA
## Steps
1. Check term and scope
2. Flag unusual clauses
MCP
Model Context Protocol: a standard plug that lets an assistant connect to outside tools and data, such as your cloud storage, a case-law database or an email account. One standard, so each tool doesn't need its own custom wiring.
Example. With MCP connections to your cloud storage and a case-law database, the assistant can search both from the same chat, with no copy-pasting between them.
Plugin
A bundle of several skills and MCP connections that you add to an assistant in one install. It is how one person's setup gets shared with a whole team.
Example. Install the contracts plugin and everyone on the team gets the same three skills and the connection to your document system, in a single step.
contracts-plugin
3 skills 1 MCP connection
Agent
A model set up to work toward a goal on its own, step by step: it decides what to do next, uses tools, checks the result and keeps going until the job is done or it gets stuck. Less like asking a question, more like delegating a task.
Example. Give it access to the data room and ask for a table of change-of-control clauses. It opens all 40 contracts one by one, pulls out the clauses, builds the table and hands it back for your review.
- Open the data room
- Read 40 contracts
- Extract key clauses
- Build the summary table
- Hand back for your review
Harness
Everything built around a model to turn it into a usable product: its standing instructions, the files and tools it can reach, and its safety checks. Same model, different harness, different results.
Example. Ask the same model the same question in a public chatbot and in a legal tool set up for your team: one gives a generic answer, the other a sourced one in your house style. The whole difference comes from the harness.
Harness
Model
API
An application programming interface: the way one program asks another for data or an action, with no human at a keyboard. Many services offer one: company registers, e-signature tools, case-law databases, AI models.
Example. When you add a counterparty, your contract tool sends its company number to the national company register's API and gets back the registered name, address and directors. Nobody visits the register's website or retypes anything.
GET /companies/123456789
200 OK
{ "name": "Acme SAS", "status": "active" }
The tech behind it
Open-source
Software whose source code is published for anyone to inspect, use and change. For AI models, the usual term is open-weight: you can download the trained model and run it on your own computers, even if how it was trained stays private.
Example. Llama and many Mistral models can be downloaded and run on your own servers. Claude and Gemini can only be used through their makers' apps and APIs.
Open: download it, inspect it, run it on your own servers
Closed: use it only through the vendor's app or API
Git
A system that keeps every version of a set of files and records who changed what, when and why. Any earlier version can be brought back. Like track changes, but for whole folders and forever.
Example. Someone asks why clause 8 of the template changed. Git shows who rewrote it, on which Monday and with what note, and Friday's draft comes back in one step.
- a1f3c9 Client edits Today
- 7be20d Clause 8 redone Mon
- 3c94e1 First draft Fri
Markdown
A simple way to mark up plain text with symbols: # for a heading, ** for bold, - for a bullet point. AI tools read and write it naturally, and skills are usually written in it.
Example. Those stray ## and ** that appear when you paste an assistant's answer into Word are Markdown marks that weren't converted into real formatting.
## Notice
**30 days**, in writing
- by registered mail
Notice
30 days, in writing
• by registered mail
Server
A computer somewhere else, usually in a data center, that does the work for you when you use an online tool. With an online AI tool, your text leaves your device and is processed on that server.
Example. A tool advertised as 'hosted in the EU' runs on servers located in Europe. That is often the first thing to check when your documents are covered by confidentiality or data-protection rules.
Wherever the provider runs it
Database
A very large table that can find, sort and update information in an instant: a filing cabinet that files itself. A register of contracts, counterparties and renewal dates, for example.
Example. Your contract register is a database, so your contract tool, your calendar and your assistant can all ask it 'which contracts renew in November?' and get the same answer in a second.
| Contract | Counterparty | Renews |
|---|---|---|
| Office lease | Dupont SA | 14 Nov |
| Supply agreement | Acme Ltd | 18 Nov |
| NDA | Leroy SARL | 2 Dec |
Cloud
Storage, computing power and software that you use over the internet, on someone else's servers instead of your own computer. 'In the cloud' does not mean nowhere: the data sits in a real data center, in a real country.
Example. Save a contract to OneDrive or Dropbox and it is in the cloud: you can open it from your laptop or your phone, and it survives a lost laptop. But a copy now lives on the provider's servers, under their security and their country's law.
Stored in a real data center
Cloud provider
A company that rents out computing power and storage in its own data centers, such as AWS, Microsoft Azure, Google Cloud or OVHcloud. Most online tools run on one of them, so where that provider stores your data matters.
Example. Your legal AI vendor runs on Microsoft Azure, which has data centers on several continents. Ask which region it uses: Paris keeps your data in the EU, Virginia sends it to the US.
AWSMicrosoft AzureGoogle CloudOVHcloud
Region:Paris or Virginia?
On-prem
Short for on-premises: software that runs on computers your organization owns and controls, in your own building, instead of in a provider's cloud. More control, more upkeep.
Example. An organization runs an open-weight model on its own servers, so confidential files never leave the building. The trade-off: its IT team now handles every update and security patch itself.
Your building
Your servers Your team
Updates and security: yours too.