Three of the eight AI notetakers I checked train on your meeting content by default. Four say they do not. One says it does not, with a qualification worth knowing before you repeat its headline.
I checked the privacy policies and security pages themselves, in November 2026, not a roundup's summary of them. Policies change, and a roundup from August is already stale in places, so treat the dates below as part of the answer. If a policy changes after this piece, the policy wins, not this article.
One limit up front, because it applies to every verdict here: a privacy policy is a promise, not an audit. I found no independent study measuring whether any of these tools actually train in practice. What follows is what they promise, in their own words, and what you can check yourself in about a minute.
The short answer
| Tool | Trains on meeting content by default? | How to stop it |
|---|---|---|
| Otter.ai | Yes, de-identified audio, transcripts may contain personal information | Enterprise only, off by default there |
| Fathom | Yes, de-identified, its own in-house models only | My Settings, all plans, or org-wide on Team Edition |
| Granola | Yes, anonymised, its own models only | Settings, Preferences, Data and sharing, all plans |
| Fireflies.ai | No, on every plan including Free | Nothing to switch off |
| tl;dv | No | Nothing to switch off |
| Zoom AI Companion | No, its own or third-party models | No training toggle, admins turn features on or off instead |
| MeetGeek | No, on any plan, unless you opt in in writing | Nothing to switch off |
| Jamie | No general-model training, but it personalises your own account | Nothing to switch off for general training |
That table is the whole piece in one glance. The rest is why the "yes" rows are not as harmless as they sound, and why the "no" rows are not all the same.
What "de-identified" actually means on a client call
The three "yes" tools do not say they train on your meeting with your name attached. They say they train on de-identified or anonymised data. That sounds safe until you think about what is in a meeting.
De-identified usually means names, emails, and direct identifiers are stripped before training. It does not mean the facts are stripped. A de-identified transcript of a client call can still contain the client's pricing, their product roadmap, their hiring plan, and the exact words they used to describe a competitor. The confidential part of a meeting is rarely the name at the top. It is the substance in the middle.
That is the freelancer problem in this whole question. You are not just protecting your own data. You are protecting a client's data, on a call they did not agree to feed into a training set. If your notetaker trains by default, you have made that decision for them, unless you opted out first.
The other trap is the opt-out that only exists on the expensive plan. Otter is the clearest case. More on that below.
The three that train by default
Otter.ai: yes, de-identified, and the opt-out is Enterprise only
Otter's privacy policy, effective June 16, 2026, section 2, says it uses data to "improve and monitor the Services, including training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)."
Read that second clause again. Transcriptions may contain personal information. The audio is de-identified. The transcript is not promised to be.
Otter also says it uses a proprietary method to de-identify user data before training so that an individual user cannot be identified, and that no customer data will be used to train its AI service providers' models. Those are real limits, but they are not the same as not training.
The opt-out is the problem. There is no self-serve training toggle documented in the privacy policy for Basic, Pro, or Business. Enterprise workspaces are opted out of AI model training by default, and an Enterprise workspace has to contact its Otter account manager to opt in. So on the plans most freelancers actually buy, the training is on, and the documented way out is to buy Enterprise.
There is also litigation in the background, and I will state it as allegations, not findings. In August 2026, a judge denied Otter's motion to dismiss a claim under California's wiretapping law in In re Otter.ai Privacy Litigation. Otter denies the allegations. A denial of a motion to dismiss is not a finding that Otter did anything wrong. It is a reason to read the policy yourself, not a verdict.
Fathom: yes, de-identified, its own models only, opt-out on all plans
Fathom's privacy policy, last updated August 16, 2026, says: "Based on how you configure your account settings, we may use and create de-identified data generated from Meeting Content Information to improve our Service by training, improving, and customizing our in-house artificial intelligence models. You can opt out from this use of your data in your account settings. We do not authorize third parties (e.g., OpenAI, Anthropic, Google, etc) to use your personal information or Meeting Content Information to train their artificial intelligence models."
Fathom's help page is clearer about the default: "Fathom uses de-identified customer data to improve the accuracy of our proprietary AI models in order to improve our service for all users. You can opt out of this anytime in your My Settings, and Organizations on Team Edition can opt out all users on their account by going to Organization Settings."
So Fathom trains, but only its own in-house models, not OpenAI, Anthropic, or Google, which are contractually barred. And unlike Otter, the opt-out is on every plan, in My Settings, not locked behind Enterprise. That is a meaningful difference if you like the product otherwise. It still means the default is on, and you have to turn it off.
Fathom also says it is HIPAA, SOC 2 Type II, and GDPR compliant, and that all Fathom data is stored in the United States. No fixed retention period is stated on the primary pages I read.
Granola: yes, anonymised, its own models only, opt-out on all plans
Granola's model-training page, October 16, 2026, is the clearest vendor page in this set, which is to its credit. It says: "We never let third-party AI providers train on your meeting data. That is contractually prohibited on every plan. We do use anonymized data to improve our own models internally, and this is on by default on Basic (Free) and Business plans. You can opt out at any time under Settings, Preferences, Data and sharing. The toggle is labeled 'Use my data to improve models for everyone.' Enterprise plans default to model training off org-wide, with admins managing the setting at the workspace level."
Its privacy policy, effective September 8, 2026, says: "We use de-identified, aggregated data, not Personal Data, for our AI model training."
Granola also says audio is cached temporarily during the meeting for transcription only, then deleted once transcription completes, data is hosted in US-based AWS infrastructure, and it earned SOC 2 Type 2 certification in July 2025, vendor spelling.
A separate case, Chamberlain v. Granola, filed in July 2026, alleges capture without knowledge and default-on training. That is an allegation, not a finding, and I am not treating it as one.
The four that say they do not train
Fireflies.ai: no, on every plan including Free
Fireflies' security page says: "You own your data. We don't train on it by default unlike other AI companies." And: "Fireflies does not use your meeting data to train AI models, and our AI vendors don't retain it after processing."
Its privacy policy, last updated September 28, 2026, section 2, says: "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." Section 6 on retention describes zero data retention for meeting content by third-party vendors after processing.
Fireflies also says it is certified for GDPR, SOC 2 Type II, and HIPAA compliance, with AES-256 at rest and TLS in transit. HIPAA BAA and private storage are on Enterprise.
tl;dv: no, and its providers may not either
tl;dv's privacy policy, published September 9, 2026, section 9, "AI Model Training," says: "Tldx Solutions GmbH does not use Customer Content, including meeting recordings, transcripts, notes, files, or other data processed through the Services, to train, fine-tune, or improve foundation models, large language models, or other generative AI models for the benefit of tldx Solutions GmbH or any third party. Where we use third-party AI service providers to deliver features of the Services, Customer Content is processed solely to provide the requested functionality and is not used by tldx Solutions GmbH or such providers to train or improve their general-purpose AI models."
tl;dv also says data is primarily processed and stored in the EEA, on Google Cloud, Wasabi, and Hetzner in Germany and Finland, with AI processing in the EU or US depending on the hosting location you select. Recordings and transcripts are kept 3 months on free plans and until account deletion on paying plans. I did not find a HIPAA claim on its policy or security pages, and its SOC 2 wording is inconsistent across its own pages, so I am not citing a SOC 2 level for tl;dv here.
Zoom AI Companion: no, its own or third-party models
Zoom's support page, "How Zoom AI features handle your data," says: "Zoom does not use any of your audio, video, chat, screen sharing, attachments, or other communications-like customer content (such as poll results, whiteboard and reactions) to train Zoom's or its third-party artificial intelligence models."
It also says its third-party model providers have zero data retention policies, with limited exceptions for trust and safety, and are not allowed to use your data to improve or train their models. That zero-retention wording has the trust-and-safety exception in it, so do not read it as absolute zero retention. Meeting summaries persist until deleted by the user or admin, or until account termination.
There is no training opt-out toggle for Zoom, because there is nothing to opt out of. Admins turn AI Companion features on or off instead.
Zoom is the odd one in this table, because most people do not shop for Zoom as a notetaker. It is here because the question people actually ask is whether the tool already in their meetings is training on them, and for Zoom AI Companion, the published answer is no.
MeetGeek: no, on any plan, unless you opt in in writing
This is the affiliate payoff in this cluster, so I will state the limit plainly: MeetGeek wins this comparison on the policy, not on a test I ran. I tested MeetGeek's bot-free mode in the earlier piece, and that test was about the bot, not about training. No one outside MeetGeek can audit whether its training promise is kept in practice. What I can verify is what it promises, and the promise is the cleanest in this set.
MeetGeek's privacy policy, last updated September 9, 2026, section 4, says: "We do not use meeting content (recordings, transcripts, summaries, chat prompts or AI outputs) to train, fine-tune or otherwise improve AI models, whether ours or those of our AI providers, unless you explicitly opt in in writing. Our AI providers are contractually prohibited from using it for training."
Its data processing agreement, August 27, 2026, section 1(m), says the same in contract language, including that subprocessors must retain customer data no longer than necessary and apply zero data retention where they offer it.
MeetGeek also says its EU instance is stored in AWS EU and its US instance in AWS US, retention is as set by your plan and workspace retention settings, including custom retention on Enterprise, and it is SOC 2 Type II audited annually. Its primary policy does not state a plan limit for a HIPAA BAA, and secondary sources disagree on that, so I am not asserting a plan limit here.
Try MeetGeek here. The reason to try it, on this question, is not that it is the best notetaker in every way. It is that on training, the answer is no on every plan, in writing, in the DPA, with no toggle to remember to turn off.
The one with a qualification: Jamie
Jamie is the reason I do not repeat a roundup's flat "no" for any tool without reading the vendor's own pages.
Jamie's data processing agreement, section 4(4), says: "The Processor shall not use Personal Data for the purpose of training, retraining, fine-tuning, or otherwise improving any machine learning or artificial intelligence models." Its enterprise page says: "No model training, Your data is never used for the training of models, not by us or any of the model providers."
But Jamie's own privacy blog says: "We don't use your sensitive data to train any of our or third party algorithms. Our sole purpose in training jamie is to enhance your user experience, particularly in the domains of speaker identification and recognizing custom words. The information you share with jamie is strictly linked to your account and isn't shared or combined with data from other users."
So the accurate wording is: no general-model training, and no cross-account training, but your voice and custom words are used to personalise your own account. That is a reasonable feature, and it is not the same as feeding your meetings into a general model. It is also not what "never used for training" means on its own. If you use Jamie, know which one you are getting.
Jamie also says it is ISO 27001 and GDPR compliant, with EU data residency, AES-256 at rest, TLS 1.2 in transit, audio deleted after transcription, and storage in the EEA, Switzerland, and UK. I did not find a verified SOC 2 Type II claim for Jamie on a primary page, so I am not citing one.
How to check any notetaker in 60 seconds
You do not need to read a whole privacy policy. Search the policy page for four words, in this order.
First, search for "train." That is the word vendors use when they train, and the word they use when they promise not to. If the first hit says "we do not use your data to train," you are mostly done. If it says "training our models on de-identified data," you have your answer, and the next search matters more.
Second, search for "de-identified" or "anonymized." If that word appears, the training is on de-identified data, which still leaves the substance of your meetings in the training set. Ask what is actually stripped, and whether transcripts are included.
Third, search for "opt out." If the opt-out exists, check two things: is it on every plan, or only Enterprise, and is it off by default or on by default. An opt-out that only exists on Enterprise is not an opt-out for most freelancers.
Fourth, search for "third party" or "subprocessor." A tool that does not train itself can still send your meetings to a provider that does, unless the provider is contractually barred. The clean policies say that bar in plain language. The vague ones do not say it at all.
If a policy does not answer those four searches, that is an answer too. A vendor that will not say plainly whether it trains is a vendor to avoid for client calls.
The consent problem no policy fixes
Even a notetaker that never trains is still recording people. That is a separate problem from training, and it is the one most articles skip.
If you record a client call, you are recording the client, not just yourself. In some states, that requires the other person's consent, not just a notice. A bot in the participant list is at least visible. A bot-free recording, which I covered in the earlier MeetGeek piece, can feel private in a way that makes consent easy to forget. Tell people you are recording, and follow the consent rules where you work, whatever tool you use and whatever its training policy says.
For a client call under an NDA, the training question is the consent question in another form. If your notetaker trains by default, you have agreed, on the client's behalf, to put their confidential information into a training set. That is not a setting to leave on by accident. Turn it off before the call, or use a tool that does not train in the first place.
Bottom line
If you take client calls and you do not want your meetings training AI models, the safe choices in this set are Fireflies, tl;dv, Zoom AI Companion, and MeetGeek, plus Jamie with the per-account personalisation qualification above. Of those, MeetGeek is the cleanest on paper: no training on any plan, opt-in only in writing, in the DPA, with its providers contractually barred.
If you already use Otter, Fathom, or Granola, you are not necessarily doing anything wrong. You are using a tool that trains by default on de-identified data. Check whether you opted out, and whether your plan even lets you. Fathom and Granola let you opt out on every plan. Otter, on the plans most freelancers buy, does not document a self-serve way out.
The rule that covers all eight, and any tool I did not check, is simple: search the policy for "train," check the default, check the plan, and check the third parties. If any of those four answers is unclear, do not put a client call on it until it is clear.