Which AI Notetakers Train AI on Your Meeting Data
“We don’t train on your data” reads like a yes/no field, and it is not. Across the vendors tracked here the same claim arrives in four different shapes, plus a fifth case where nothing is published at all, and the difference between them decides whether a buyer is covered.
So we record it in five states: a plain no, a no that covers Enterprise plans only, training on anonymised data with an opt-out, training only where a customer opts in, and not published. Each state below carries the vendor’s own sentence, quoted verbatim, with the page it came from and the date we read it.
The interesting information is almost never in the sentence itself. It is in what the sentence is scoped to - a subprocessor’s API, one plan, or data received through somebody else’s platform. The last section of this page sets out the three narrowing patterns we keep finding. Where a vendor’s pages disagree with each other on any subject, those are catalogued separately on the vendor claim tracker.
13 vendors, last verified 2026-09-09.
| Vendor | Trains on your meeting data | Verified |
|---|---|---|
| Fireflies.ai | No | 2026-09-09 |
| Jamie | No | 2026-09-09 |
| MeetGeek | No | 2026-09-07 |
| Noota | No | 2026-08-26 |
| Sembly | No, for Enterprise plans only | 2026-09-02 |
| Fathom | Yes, on anonymised data, with opt-out | 2026-09-09 |
| Granola | Yes, on anonymised data, with opt-out | 2026-09-09 |
| Otter.ai | Yes, on anonymised data, with opt-out | 2026-09-09 |
| Read AI | Opt-in only | 2026-09-09 |
| Avoma | Not published by the vendor | 2026-09-01 |
| Krisp | Not published by the vendor | 2026-09-09 |
| Notta | Not published by the vendor | 2026-09-01 |
| Tactiq | Not published by the vendor | 2026-09-09 |
No
The vendor states plainly that it does not use customer meeting content to train models, with no plan, product surface or subprocessor carve-out inside the sentence. This is the only state where a buyer needs no further reading.
“We do not use personal information for AI model training and we contractually prohibit our vendors from using this information for their own model training.”
2026-09-09
No
“Your data is never used for training of models, not by us or any of the model providers.”
2026-09-09
No
“Zero data training”
2026-09-07
No
“Your data is not use to train any of our models. We do not sell your data. We do not share your data.”
2026-08-26
No, for Enterprise plans only
A no-training commitment exists, but it attaches to one plan. Buyers on every other plan are outside it, including plans sold at a published price.
No, for Enterprise plans only
“does not use Enterprise Plan customer meeting content, including audio, video, or transcripts, for AI model training”
2026-09-02
Yes, on anonymised data, with opt-out
The vendor does train on customer content, states that the content is anonymised first, and offers a way out. The default is participation.
Yes, on anonymised data, with opt-out
“Fathom uses de-identified customer data to improve the accuracy of our proprietary AI models in order to improve our service for all users.”
2026-09-09
Yes, on anonymised data, with opt-out
“Granola trains on your anonymized data so we can keep making Granola better. You can opt out of this in your Settings.”
2026-09-09
Yes, on anonymised data, with opt-out
“Improve and monitor the Services, including training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information).”
2026-09-09
Opt-in only
Training happens only where a customer has actively turned it on. The default is non-participation.
Opt-in only
“Contributing to our model is completely optional, and opt-out by default.”
2026-09-09
Not published by the vendor
The vendor publishes no statement about whether meeting content trains its models. Some of these vendors publish a sentence that looks like an answer but is scoped to something else - a subprocessor, or an API surface. We record silence as silence rather than reading it either way.
Not published by the vendor
“Data obtained through Google Workspace APIs is used exclusively to provide and enhance the services requested by our users. We do not use this data to develop, improve, or train generalized AI or machine learning models.”
2026-09-01
Not published by the vendor
“no information obtained through such third-party apps or services (e.g. Google Workspace APIs) is used to develop, improve, or train generalized AI and/ML models (this is scoped to third-party data only and does not cover Krisp's own use of meeting content).”
2026-09-09
Not published by the vendor
“Google Workspace APIs are not used to develop, improve, or train generalized/non-personalized AI and/or ML models.”
2026-09-01
Not published by the vendor
“No. Azure OpenAI will not use data submitted via their API to train or improve their models. (Tactiq publishes no statement about its own training use - this sentence covers its sub-processor only.)”
2026-09-09
Three ways the sentence gets narrowed
A no-training sentence can be true and still not cover the buyer reading it. These are the three narrowings we keep finding, in order of how hard they are to notice.
1. Scoped to the Google Workspace APIs
The vendor’s only no-training sentence is scoped to data received through the Google Workspace APIs, which Google’s own API terms require them to say. It covers the calendar and Drive data Google hands them; it does not cover the meeting recordings and transcripts the product itself creates.
It is the narrowest promise available, and in a privacy policy it reads exactly like a general one.
Not published by the vendor
“Google Workspace APIs are not used to develop, improve, or train generalized/non-personalized AI and/or ML models.”
2026-09-01
No
“Zero data training”
2026-09-07
Not published by the vendor
“Data obtained through Google Workspace APIs is used exclusively to provide and enhance the services requested by our users. We do not use this data to develop, improve, or train generalized AI or machine learning models.”
2026-09-01
2. Scoped to a subprocessor
The sentence is about what a model provider does with data sent to its API, not about what the vendor does with the transcripts it stores. A confident No that answers the question on somebody else’s behalf.
Not published by the vendor
“No. Azure OpenAI will not use data submitted via their API to train or improve their models. (Tactiq publishes no statement about its own training use - this sentence covers its sub-processor only.)”
2026-09-09
3. Scoped to one plan
The exclusion is real and clearly written, but attaches to one plan. Every plan sold at a published price sits outside it.
Worth saying plainly: writing the limit into the sentence itself is more transparent than either pattern above, even though the commitment it makes is narrower.
No, for Enterprise plans only
“does not use Enterprise Plan customer meeting content, including audio, video, or transcripts, for AI model training”
2026-09-02
What we do not do here
- We record what a vendor publishes on its own domain, not what it may or may not do privately.
- A vendor with no published statement is marked not published. It is not assumed innocent and it is not assumed guilty.
- Where a vendor’s own pages disagree, we take the weaker claim.
- A certificate or policy that exists but does not appear on the vendor’s own site is not recorded here as held.