"AI translation" is no longer a single, uniform thing — results differ significantly depending on how a service works internally. These criteria help you judge what you're actually getting, regardless of provider.
1. Consistency across the whole book
The most important, most easily checkable point: does a secondary character have the same name in chapter 2 as in chapter 20? Is a recurring technical term or place name always translated the same way? That sounds trivial but is the most common weak spot in long texts — especially with tools that process text in independent chunks without shared memory.
Technical background: models with a limited context window "forget" earlier chapters as they keep working. Solutions that maintain a cumulative glossary (names, technical terms, recurring expressions) throughout the entire translation process and feed it into each new section avoid this problem structurally.
2. How wordplay and cultural references are handled
An automatic translation that simply smooths over an untranslatable pun, with no indication, isn't a bad translation in a technical sense — but it takes away your ability to decide how to handle it. Better solutions flag such passages explicitly instead of silently "resolving" them.
3. Tense and register consistency
Does the narrative unexpectedly shift from past tense to present? Does the form of address switch mid-dialogue? These are errors that barely register in individual sentences but become noticeable across a whole book — and a single-pass translation doesn't automatically catch them itself.
4. Is there an independent quality check?
A second pass, run independently from the first translation attempt, that explicitly looks for omissions, meaning errors, and the inconsistencies mentioned above, is a strong quality signal. Without this step, you're trusting that the first attempt was already correct.
5. Do you stay in control, or is it a black box?
Can you choose between multiple variants at uncertain passages (literal, semantic, creative), or do you just get a finished result with no insight into decisions made? For literary text, where tone and style matter, the ability to adjust yourself is often more important than a single technically "perfect" result.
A quick test to try yourself
To test a tool before trusting it with a whole book: run a paragraph containing wordplay, one with a recurring proper name (that also appears later in the text), and one with dialogue in shifting register. The three weak spots mentioned above usually show up already in a short test passage, before you invest time or money in a full manuscript.