Why a checklist
Two things are true at once. In a survey of 130 universities, institutions give researchers different answers, and 40 had no guidance we could find. And in a read of 53 services' public documents, only 14 clearly say customer content is not used to train AI models. The sources are in the guidance paper and the terms paper.
Nobody is going to hand you a verdict. These eight questions let you reach your own, and write it into your protocol.
1. What does my ethics office expect for cloud processing?
Start here, not with the vendor. Only about a quarter of institutions with guidance say researchers must disclose AI transcription in the ethics protocol, and many pages are written by IT rather than ethics offices. A missing rule is a question to ask, not a permission.
2. Does the service say it does not train on customer content?
Look for a plain statement. Fourteen of 53 services have one without a contradicting document. For 25 the documents do not give a clear answer, and for 15 at least one document permits training by default.
3. Is there an opt-out, and who can use it?
Of the 15 services that permit default training, 8 describe no opt-out, 4 require a request to the company, 3 have a self-serve setting and 1 reserves it for enterprise customers. Check whether business-tier terms differ from yours: 15 services state different terms for business customers.
4. How long is audio kept, and how long does deletion take?
Look for a number. 'As long as necessary' is the most common answer at 22 services, and only 14 give a fixed period. Separately, 29 state a deletion time after you delete content, most often 30 days. More in how long services keep your audio.
5. Can people at the company or its contractors hear it?
Eleven services describe routine human access, 24 describe access only under conditions, 3 say never, and 10 say nothing. If your consent form promised that only the research team would hear the audio, this line decides whether it still holds.
6. Which sub-processors and AI providers touch the file?
Twenty-five services publish a named sub-processor list and 22 name the AI or speech providers. Documents of 19 say those providers may not train on customer content, while 28 say nothing about it. A named list is something you can put in a data management plan.
7. Where is data stored and processed?
Forty-four services state it. If your ethics board or data protection officer cares about location, a service that does not state it has not answered the question.
8. Do the documents agree with each other, and when did I read them?
Seven services publish documents that contradict each other about training. Compare the privacy policy, terms and security page, and write down the date. Terms move: against October 2024 captures, 23 of 41 comparable services are coded differently on training, and in 16 of those the 2024 policy had said nothing about it. The study counts what documents say, not what a service does, so treat a clean read as evidence, not a guarantee.
Put the answers in the protocol
Name the tool, the date you read its terms, the retention and deletion terms, and where the audio is processed. Our guides on qualitative research transcription, confidential transcription and anonymizing a transcript cover the steps around it, and the researcher workflow shows the end-to-end path. Pepys sells a cloud AI transcription service and is one of the 53 services in the terms study.

