Why Generic AI Translation Mangles PDFs and EPUBs
A PDF is not text. It is a layout container that may or may not contain text. An EPUB is not a flat document; it is an archive of files with structure that has to be walked, not skimmed. Treating either as "just paste the content" is how AI translation produces output that looks plausible and contains nothing the source actually said.
What goes wrong
A scanned book is uploaded. The chat interface accepts it without complaint, the model produces fluent translated prose, and none of it matches the source, because the source was an image and the model invented something to translate. An EPUB with footnotes and chapter structure is flattened into one long blob, and references that depended on structure are lost.
Why generic AI translation fails here
The boundary between "this file has extractable text" and "this file is a picture of text" is invisible to a chat UI. The model is happy to receive anything, and the user has no way to know that nothing useful was actually parsed. The first signal of the problem is the final output, after the payment has been made.
How TranslationAI solves it
Document preflight runs before payment. Two extractors attempt to read the file in parallel; if neither finds usable text, the upload is rejected with a clear explanation of why. Image-only PDFs are refused at the door, not silently translated as gibberish. The user keeps their money and sees an honest rejection instead of a fluent mistranslation. EPUB structure is parsed as an archive, not flattened.
Further reading: why honest rejection beats fluent failure, how the same orchestration handles structured documents, why structural fidelity makes terminology enforceable.