Working paper
Who Keeps the Recording? What 53 Transcription and Meeting-Notes Services Say About Training, Retention and Access
Ankur Shrestha, Pepys, United States · 2026-11-13
Abstract
People upload interviews, meetings and calls to AI transcription and meeting-notes services, and the public documents of those services are the main source a user, an ethics office or a procurement team can consult about what happens to the audio. We analyzed the privacy policies, terms of service, data processing agreements, sub-processor lists and AI-data pages of 53 services: 35 consumer and prosumer transcription tools, 10 meeting-notes bots, 5 quote-based services and 3 meeting platforms. All 247 documents were taken from Internet Archive captures, not live pages; 233 (94.3%) were retrieved. Two coders worked independently from the saved text under a twelve-variable codebook fixed before collection, every quote was checked as an exact substring of the saved document, and disagreements were adjudicated against the text. Only 14 services (26.4%, 95% CI 16.4–39.6) state that customer content is not used to train AI models. Another 15 (28.3%) have at least one document that permits training by default, including 7 whose own documents contradict each other; for 18 (34.0%) the documents are silent or unclear. Of the 15, 8 describe no way to opt out. Audio retention is open-ended ("as long as necessary") at 22 services (41.5%) and a fixed period at 14 (26.4%). Eleven services (20.8%) describe routine human access to customer content. Compared with captures from October 2024, 23 of 41 comparable services are coded differently on training, and in 16 of those the 2024 policy had said nothing about it. In a companion survey of 130 universities, the tools institutions approve or bar most often do not line up with their own training terms. Inter-coder agreement ranged from κ = 0.69 to κ = 0.97. The documents give users and institutions an incomplete and sometimes contradictory account of whether a recording is used to train models.
53
services: 35 transcription tools, 10 meeting-notes bots, 5 quote-based services, 3 meeting platforms
14 of 53
services state that customer content is not used to train AI models (26.4%)
22
services describe audio retention as open-ended ("as long as necessary"); 14 give a fixed period
11
services describe routine human access to customer content
Do transcription services train on your recordings?
Often the documents do not say clearly. Only 14 of 53 services (26.4%) have documents stating that customer content is not used to train AI models. Another 15 (28.3%) have at least one document that permits training by default, including 7 whose own documents contradict each other, and for 18 (34.0%) the documents are silent or unclear. Of the 15, 8 describe no way to opt out.
Retention is rarely a number. Audio retention is open-ended at 22 services (41.5%) and a fixed period at 14 (26.4%). Eleven services (20.8%) describe routine human access to customer content, such as annotation or quality review.
The study reads documents, not systems: a service may do more or less than its documents say, and the documents were taken from Internet Archive captures not later than 3 October 2026, which can lag the live page. The paper counts what the documents say and does not rank or recommend services. Pepys is one of the 53 and was coded by the same procedure as the rest.
Where to find it
- SSRN preprint: View on SSRN
- DOI: 10.5281/zenodo.24204062 (Zenodo preprint) · all versions: 10.5281/zenodo.24203178
- Dataset DOI: 10.5281/zenodo.24203123 (CC BY 4.0)
Putting it into practice
- Companion survey: what 130 universities allow
- Plain-language summary: do transcription services train on your recordings
- Retention and access: how long services keep your audio
- Before you upload: an eight-question checklist
- Workflow for researchers: research interview transcription
Cite
@misc{shrestha2026who,
author = {Shrestha, Ankur},
title = {Who Keeps the Recording? What 53 Transcription and Meeting-Notes Services Say About Training, Retention and Access},
year = {2026},
howpublished = {SSRN preprint, abstract\_id 7701702},
doi = {10.5281/zenodo.24204062},
url = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7701702}
}RIS
TY - RPRT
AU - Shrestha, Ankur
TI - Who Keeps the Recording? What 53 Transcription and Meeting-Notes Services Say About Training, Retention and Access
PY - 2026
DA - 2026/11/13
PB - SSRN
ET - 1.1
UR - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7701702
DO - 10.5281/zenodo.24204062
N1 - Dataset: https://doi.org/10.5281/zenodo.24203123
KW - AI transcription
KW - privacy policies
KW - terms of service
KW - model training
KW - data retention
KW - meeting assistants
KW - document analysis
AB - People upload interviews, meetings and calls to AI transcription and meeting-notes services, and the public documents of those services are the main source a user, an ethics office or a procurement team can consult about what happens to the audio. We analyzed the privacy policies, terms of service, data processing agreements, sub-processor lists and AI-data pages of 53 services: 35 consumer and prosumer transcription tools, 10 meeting-notes bots, 5 quote-based services and 3 meeting platforms. All 247 documents were taken from Internet Archive captures, not live pages; 233 (94.3%) were retrieved. Two coders worked independently from the saved text under a twelve-variable codebook fixed before collection, every quote was checked as an exact substring of the saved document, and disagreements were adjudicated against the text. Only 14 services (26.4%, 95% CI 16.4–39.6) state that customer content is not used to train AI models. Another 15 (28.3%) have at least one document that permits training by default, including 7 whose own documents contradict each other; for 18 (34.0%) the documents are silent or unclear. Of the 15, 8 describe no way to opt out. Audio retention is open-ended ("as long as necessary") at 22 services (41.5%) and a fixed period at 14 (26.4%). Eleven services (20.8%) describe routine human access to customer content. Compared with captures from October 2024, 23 of 41 comparable services are coded differently on training, and in 16 of those the 2024 policy had said nothing about it. In a companion survey of 130 universities, the tools institutions approve or bar most often do not line up with their own training terms. Inter-coder agreement ranged from κ = 0.69 to κ = 0.97. The documents give users and institutions an incomplete and sometimes contradictory account of whether a recording is used to train models.
ER - Version history
- v1.1 (2026-11-13): Corrected PDF posted as a new Zenodo version; this is the file hosted on this page.
- v1.0 (2026-11-13): First posted.