A good glossary is the difference between a smooth assignment and a stalled one. AI can draft it in minutes, if you treat every line as a suggestion until you've checked it.
Most interpreters know the feeling: the briefing arrives the night before, the subject is a corner of a field you've never touched, and there are forty terms you need to own by morning. AI tools are very good at the first draft of that list. They're also confident when they're wrong. The workable approach is to let AI draft and organize, and to let you decide. Here is a five-step workflow that saves real prep time without trusting a language model with your credibility.
General-purpose assistants like ChatGPT and Claude are strong at the mechanical parts of glossary work: extracting candidate terms from a long document, grouping them by sub-topic, proposing equivalents in a target language, and drafting a one-line definition so you remember what each term means in context.
They are weak at exactly the part that matters most in the booth or on the call: knowing which equivalent is the one this field, in this country, actually uses. A model will happily produce a fluent, plausible, wrong term, especially for newer concepts, regional variants, and institution-specific vocabulary.
Paste in the agenda, the slide deck, the contract, the case summary, or the program notes, whatever the client has actually shared. A glossary built from the real document is far more useful than one built from a generic prompt like "medical terms for cardiology".
Request a table: source term, target-language equivalent, a one-line definition, register or usage notes, and a confidence flag. Ask the model to mark anything it isn't sure of, then treat the unmarked terms with suspicion too.
Check each term in a specialist dictionary, an official terminology database, published standards, or a parallel text in the target language from the same field. When the stakes are high, ask the client or a subject expert. This is the step that can't be skipped.
Split the list into terms you must have instantly (names, numbers, acronyms, the five concepts the meeting is about) and terms you only need to recognize. Memorize the first group, keep the second as a reference.
A glossary on a laptop in another window helps less than one built into your tools. Load verified names and terms into the dictionary of whatever captioning or CAI tool you use during the assignment.
The difference between a useless list and a useful one is usually the prompt. Be specific about the language pair, the country, the setting, and the format you want back. A pattern that works:
| ChatGPT / Claude | Spreadsheet or term base | Live captioning tool | |
|---|---|---|---|
| Best at | Drafting and organizing quickly | Storing your verified terms long-term | Showing the right term live, during the session |
| Biggest risk | Confident wrong terms | Gets stale if not maintained | Only as good as the terms you load |
| When you use it | Night before the assignment | Before and after every job | During the assignment |
| Verification | Always required | You are the source of truth | Check unusual terms in the transcript |
The glossary you've verified is only useful if the names and terms show up correctly when the speaker says them. Unicaption is a real-time captioning and translation copilot for interpreters, and its custom dictionary is built for this step.
It's a copilot, not a replacement: the live text is a reference for you, and the interpreting is still yours. See how to improve live caption accuracy and custom dictionaries and terminology accuracy for how the dictionary works in practice.
The best glossaries grow. After each job, move the terms that actually came up into your permanent term base, delete the ones you guessed wrong, and note the ones the client corrected. That feedback loop is worth more than any model: over a year it becomes a resource tuned to your languages, your clients, and your fields.
If you're still putting your pre-assignment routine together, the interpreter prep checklist covers the rest of the night-before work, and the best AI tools for interpreters compares what's available for each stage.
They can draft one quickly: extract terms from your assignment documents, propose target-language equivalents, and add short definitions. But they can also produce fluent terms that don't exist or aren't used in your country, so every term must be verified against a specialist dictionary, an official terminology source, or the client before you rely on it. Unicaption's custom dictionary then loads the verified list into your live captions.
Look it up in a specialist dictionary or official terminology database for your field, find it in a parallel text written by native professionals in that field, or ask the client or a subject expert. If you can't confirm the term, don't use it as if it were fact. Treat AI output as a candidate list, not a reference.
Only if your NDA, the client, and the tool's data policy allow it. Many assignments forbid sharing material with third-party tools. When in doubt, build the glossary from public information or your own notes. Unicaption is built for live sessions: it is end-to-end encrypted, never stores audio, deletes transcripts at the end of the session, and never uses content to train models.
It's a list of names, products, acronyms, and specialized terms you give the tool so it spells them correctly when speakers say them. In Unicaption you add your terms before the session, and they appear correctly in the live captions alongside built-in medical, legal, and government terminology.
Custom dictionary, live translation, nothing stored. 30 free minutes every week.
Start Free Trial →