The Dangers of Relying on AI Transcription in Translation Projects
Erica Huttner
I recently worked on a translation project that was a great example of the pitfalls of AI. I was hired by an agency to translate transcripts of interviews that had been conducted in Spanish. When I initially skimmed the documents they had received from their client, it seemed like a straightforward task. However, once I got started, it became clear that the transcription had been generated by AI, which the agency had not been told.
How did I know it was transcribed by AI? The transcripts were a mess, riddled with the types of errors a human would not make. For example:
When the interviewers introduced themselves, their common female names (which were referenced elsewhere in the document) were changed to somewhat similar male names (think Jane > James or Anna > Andrew). Default male bias, perhaps?
At one point a university with an unusual name was referenced. The AI simply invented a word based on the pronunciation (which produced no search results on Google), but I was able to correct it with an educated guess based on context.
Every time uno or dos was said, the transcription cited the numbers “1” or “2”. However, in most instances the speakers were using terms like cada uno or los dos.
Worst of all, text was repeatedly attributed to the wrong speaker. I didn’t have access to the original audio recording, but the issues were obvious. The interviewer would ask a question and the interviewee would start to answer, but then the transcript would switch randomly between speakers, as if the interviewer had the power to read the interviewee’s mind about their personal experiences.
In the end, this translation took much longer than it should have because the source text was generated by AI. Since it was for an agency that I have a great relationship with and the transcription issues were not their fault, I took the time to find and correct as many errors as possible, as well as sending them a list of the issues I had discovered so that they could pass it along to their client. I doubt I’ll ever hear what comes of it, but my hope is that the client might rethink their use of AI given the serious inaccuracies that were introduced into their transcripts.