How AI Translation Works in Lectio
When you press "AI Translate" on a segment, Lectio does not simply send your source text to a language model and hope for the best. It assembles a rich context from four layers of information, each designed to make the AI's suggestion more accurate, more consistent, and more useful to you as a professional translator.
The Four-Layer Context System#
1. Global Context
Before the AI sees a single word of the segment you are translating, it receives the translation brief -- a summary of your document's content, tone, themes, and intended audience generated when you choose to prepare the document. It also receives the source and target languages and any translation rules you have attached to the project.
Translation rules are natural-language instructions you write to guide the AI's behavior. For example: "Use formal register throughout," or "Transliterate all proper names rather than translating them." These rules apply to every segment in the project.
This global layer ensures the AI understands what kind of document it is working on before it starts translating. A training manual and a devotional book call for different vocabulary and tone, and the AI adapts accordingly.
2. Glossary Context
The AI receives every glossary term that appears in the current segment. Terms are divided into two categories:
- Confirmed terms -- the AI is instructed to use these exactly as specified. If your glossary says "covenant" translates to a specific word in your target language, the AI will use that word.
- Candidate terms -- suggested translations that the AI treats as strong preferences but not absolute requirements.
This layer is where organizational terminology decisions take effect. A team that has invested time building a glossary for theological vocabulary will see that investment pay off in every AI suggestion. The AI will not invent a new rendering for "justification" or "sanctification" when your glossary already specifies one.
3. Translation Memory Context
If you have a translation memory (TM) attached to the project, Lectio searches it for sentences similar to the current segment. Matches of 70% similarity or more are sent to the AI as reference translations.
The AI uses these TM matches as examples of how your organization has translated similar content before. This is especially valuable for recurring phrases, standard introductions, or content shared across a series of related documents. The AI does not copy TM matches verbatim (unless the match is exact), but it draws on them to maintain consistency with your past work.
4. Segment Context
Finally, the AI receives the surrounding segments -- the paragraphs immediately before and after the one you are translating. This local context helps the AI understand how the current segment fits into the flow of the document.
Without segment context, a sentence like "He did" is ambiguous. With the preceding paragraph visible, the AI knows who "he" refers to and what "did" means. This dramatically reduces mistranslation of pronouns, connectives, and context-dependent vocabulary.
How These Layers Work Together#
Each layer addresses a different dimension of translation quality:
- Global context sets the right tone and register
- Glossary context enforces term-level consistency
- TM context maintains sentence-level consistency with past work
- Segment context ensures the suggestion reads naturally in its immediate surroundings
A glossary with "atonement" mapped correctly means the AI will not produce a different rendering each time it encounters the word. A TM filled with completed chapters means the AI has examples of your team's preferred style. Translation rules let you encode decisions that apply universally -- "always use inclusive language" or "prefer dynamic equivalence over formal equivalence."
The AI as a Tool, Not an Author#
Lectio's AI is a tool under your control. It suggests; you decide. Every AI-generated translation is placed in the editor as a draft for your review. You can accept it, modify it, or discard it entirely. The AI never publishes, confirms, or finalizes anything on its own.
This is by design. Professional translation requires human judgment -- especially for theologically sensitive content where word choice carries weight. The AI accelerates your work; it does not replace your expertise.
Why Suggestions May Vary#
AI models are probabilistic, which means the same input can produce slightly different output on different runs. If you delete an AI suggestion and request a new one, it may not be identical to the first. This is normal behavior, not a bug.
The variation is usually minor -- a different word order, a synonym, a slightly different phrasing. If you need a specific rendering, add it to your glossary or write a translation rule. These give the AI deterministic guidance that reduces variation in the areas that matter most to you.
Credit Consumption#
Each AI action uses credits based on the amount of text the AI processes (both input and output tokens). Longer segments consume more credits than shorter ones. Features that send more context -- like AI Translate, which includes glossary, TM, and surrounding segments -- use more credits per request than simpler features like Explain.
Document preparation is also credit-based, but it never begins merely because a file was uploaded. Lectio shows an estimated range first. For split documents, you can prepare every part, the first part, or leave all parts for later. When the workflow includes Bible Verse Review, finding Scripture quotations is part of that preparation step.
For a full breakdown, see Understanding Credits.
Copyediting AI drafts#
After generating a new draft, Lectio normally asks GPT-6 Luna to polish its wording using relevant translation memory and project guidance. Human-selected terminology is protected in its existing draft form, including permitted inflection. Saved sentence translations and approved Bible insertions are restored exactly after copyediting. A response that loses, duplicates, or alters a protection token is discarded, keeping the initial draft.
Copyediting uses normal AI credits and adds processing time. Team owners can turn it on or off in Teams → their team → Translation → AI copyediting. Administrators can also change the same setting in Operations Console → AI & Content → Team access. Exact translation-memory reuse and saved whole-sentence translations bypass this step. Copyediting does not substitute for your review of the final translation.