How does AI translation work?

Better i18n translates with LLMs, and there is more than one path into them. They do not all carry the same context, which is the thing worth understanding before you judge the output.

The three paths #

Path	How you reach it	What the model is told
AI drawer	`⌘I` in the project	Your glossary and project context — the richest path
Cell suggestion	The AI action on a single value	The text, the source and target language, and your project instructions
Bulk translation	Translate many keys at once	Queued as a job and applied when it finishes
Your own agent	MCP, from your editor	`getTranslationContext` gives it the same glossary and instructions the drawer uses

The practical consequence: if terminology consistency matters, translate through the drawer or an MCP agent. A one-off cell suggestion is a good translation of that string, but it is not looking at your glossary — so a pinned term can come back as a synonym, and you would have no idea why.

What shapes the output #

Project instructions (Custom System Prompt) are inserted ahead of the platform's own rules. Formality, regional variant, register — see translation guidelines.
Glossary terms reach the drawer and MCP paths, matched by meaning rather than exact substring — see managing your glossary.
Three rules are always applied, whatever you write: return only the translation, keep the tone and style of the source, and preserve placeholders like {{variable}} and {count}.

That last one matters more than it sounds. A translation that drops {count} is not a slightly worse translation, it is a broken string, and it is the failure mode a human reviewer skims past.

Quality habits, in order of payoff #

Set the formality per language. No model can infer whether your German says Sie or du from a button label, and getting it wrong is the mistake every native speaker notices first.
Put the words you cannot afford to lose in the glossary, not in a prompt paragraph. Per-term entries are reviewable and reach machine-translation engines too.
Translate through the drawer for anything user-facing. Cell suggestions are for filling a gap, not for setting your product's voice.
Review what a customer reads closely — pricing, legal, empty states, error messages. AI is fast at the long tail; the short list at the top is where a human pass pays for itself.
Check the placeholders survived before publishing. The editor highlights them, so this is a glance, not an audit.

After translating #

Nothing is live until you publish. Each value keeps its own history with the origin of the change recorded — an AI-generated value says so — so a translation you are unsure about can always be traced rather than guessed at.