Your Meeting Transcripts Are Training Data: What the EDPB's July 2026 Guidelines Mean for Otter, Fireflies, and Granola Users

If you're using Otter, Fireflies, or Granola on anything other than an Enterprise plan, your meeting transcripts are — right now, by default — being used to train the vendor's AI models. On July 7-8, 2026 in Brussels, the European Data Protection Board adopted Guidelines 03/2026 on web scraping in the context of generative AI — the first pan-EU framework confirming that the GDPR applies in full whenever personal data is used to train AI, with no carve-out. The guidelines were written for web scrapers, but the principles land squarely on the AI notetaker industry: consent "will most probably not" work at scale, legitimate interest requires a documented three-part test, and once data enters model weights, it can't easily be removed. This piece walks through what changed on July 7, what your notetaker actually does with your conversations, and why on-device transcription is the only architecture that avoids the problem entirely.

What the EDPB actually adopted on July 7-8, 2026

The European Data Protection Board is the panel of national data-protection regulators that coordinates GDPR enforcement across the EU. At its Brussels plenary, per YuSMP Group's regulatory summary, it adopted Guidelines 03/2026 on web scraping in the context of generative AI — a 22-page framework confirming that the GDPR applies in full to personal data scraped to train AI models, with no carve-out for AI. Public consultation runs until October 30, 2026, but the enforcement direction is already clear.

Reporting from PPC Land characterized the guidance as blocking AI firms from using consent as an excuse to scrape, setting out a three-condition legitimate-interest test that firms will have to document dataset-by-dataset. Reed Smith's analysis notes the EDPB expressly acknowledged that this kind of processing typically occurs without the data subject's knowledge or consent, creating significant risks to fundamental rights.

The four things the guidelines say that matter for notetakers

Distilled from the YuSMP summary:

• Consent "will most probably not" work. Public visibility of data — or, by analogy, presence on a Zoom call — is not consent to have that content ingested for AI training.
• Legitimate interest needs a documented three-part test (genuine interest, necessity, and a balancing test the data subjects' rights don't override).
• Special-category data is near-prohibited. Incidental collection — the medical detail that comes up in a client call, the biometric voiceprint generated by speaker diarization — must be minimised and deleted on discovery.
• Once trained, data can't easily be deleted from a model. Governance must happen upstream, before ingestion, not after.

That last point is the sharp end for anyone whose voice has ever been on a call recorded by a mainstream cloud notetaker.

What your notetaker actually does with your meetings

The default training behavior varies by vendor and plan, and it's almost never the answer you'd guess.

Otter.ai

Otter's own privacy policy discloses sharing with "data labeling service providers who provide annotation services and use the data we share to create training and evaluation data for Otter's product features." A Cyber Unit review of six major notetaker privacy policies found that these labeling companies are not named in Otter's privacy policy, and standard users have no opt-out for this transfer. The same review confirmed Otter also works with Facebook and other advertising partners for personalized ad targeting.

The Otter lawsuit filings go further. According to a structured analysis of the Brewer complaint, Otter's Privacy Policy admits to training its AI on de-identified recordings and transcripts but only seeks consent from users, not from non-users present in those meetings — effectively outsourcing its legal obligations by asking hosts to "make sure you have the necessary permissions."

Granola

Granola's positioning as a "bot-free" alternative was supposed to sidestep the notetaker consent problem. Then Chamberlain v. Granola was filed on July 30, 2026. According to the PPC Land reporting on the complaint, model training is enabled by default for users on Granola's Free and Business plans, through a setting a user must locate and manually disable. On Business plans, each individual user must opt out separately rather than an administrator doing so once for the organization.

Granola's own security page confirms the training default: "Granola trains on your anonymized data so we can keep making Granola better. You can opt out of this in your Settings." Only Enterprise users have model training turned off by default. A detailed Granola privacy audit published August 3, 2026 confirmed that training on your meetings is on by default for the free tier and for Business, and off by default only for Enterprise.

Fireflies

Fireflies is already a co-defendant in the consolidated Northern District case. Per the Embertype notetaker-lawsuit tracker, Cruz v. Fireflies.AI Corp. was filed on December 18, 2025, raising similar consent-and-recording claims to the Otter lawsuit and consolidated before the same federal judge. Its privacy policy should be read alongside the complaint before any regulated deployment.

Comparison: how training defaults stack up

Why "we anonymize before training" isn't a defense

Every major notetaker's response to the training question is some version of "we train on de-identified data." The Brewer complaint's structured critique cites academic research showing that ML models trained on de-identified data remain vulnerable to re-identification attacks, rendering the de-identification claim weak. Voice content is particularly bad here: a voiceprint isn't just personal data, it's biometric identifier data — the very category the EDPB flagged as near-prohibited absent both an Article 6 basis and an Article 9(2) exception.

And the durability problem is real. A detailed retention analysis of Otter's data lifecycle put it bluntly: once a meeting transcript has contributed to a model-training pass under the opt-out default, the de-identified contribution to model weights is durable, and deleting the original recording does not unwind the training contribution.

Article 5 GDPR and the transparency problem

Article 5 of the GDPR requires that personal data be processed "lawfully, fairly and in a transparent manner in relation to the data subject." A meeting participant who never signed up for the notetaker, never saw a disclosure, and had no chance to decline before the transcription started cannot, by any reading of the text, be said to have received transparent processing.

The National Law Review's analysis of the Otter litigation makes a related point about California law that reads across to Europe: the fact that a notetaker does not store audio does not eliminate exposure. The EU regime — Article 5 transparency plus the new EDPB guidance on training-data lawful basis — points in the same direction. Ephemeral processing that still results in personal data entering a training corpus is still processing that needs a lawful basis.

The EU AI Act layer arriving August 2, 2026

The EDPB guidance doesn't sit alone. HR Executive's coverage of the Littler analysis flags that beginning in August 2026, the EU AI Act introduces a separate layer of obligation, and AI systems used for worker monitoring and management may be classified as high-risk — a category that could encompass tools offering sentiment analytics or productivity scoring alongside transcription.

The same reporting notes that in co-determination countries such as Germany and France, deploying an AI notetaker may also require works council consultation before rollout, a requirement with no U.S. equivalent that multinational HR teams frequently overlook. For further coverage of how these frameworks interact, see our analysis of Article 50 transparency obligations for AI notetakers.

The litigation backdrop: three consolidated cases

The EDPB guidelines land in the middle of an already-hot U.S. docket. Per the Embertype lawsuit tracker, since August 2025 three major federal privacy lawsuits have been filed against AI notetaker companies. Recording Law's docket summary notes that Judge Eumi K. Lee consolidated the Otter suits into In re Otter.AI Privacy Litigation, No. 5:25-cv-06911-EKL, with a consolidated complaint filed December 5, 2025 alleging Otter obtained consent, at most, from the meeting host who added the assistant — not from the other participants whose voices it recorded and transcribed.

The complaint further alleges that Otter used recorded content to train and improve its artificial intelligence models. The case docket summary confirms Otter moved to dismiss, the motion-to-dismiss hearing was held on May 20, 2026, and Judge Lee took the matter under submission. California's CIPA allows statutory damages of $5,000 per violation, and the federal Wiretap Act provides its own statutory damages.

Chamberlain v. Granola, filed July 30, 2026 in the same district, extends the theory to bot-free notetakers. Per the PPC Land complaint summary, the filing extends California's wiretapping and privacy statutes to an AI product whose core function is recording conversations and repurposing them for model training. For a deeper look at that case, see our Chamberlain v. Granola breakdown.

What professionals should do this week

If you handle any content that would embarrass a client, disclose a trade secret, or make discovery counsel nervous — sales calls, therapy sessions, investigator interviews, deposition prep, board discussions — you need to know the training posture of every AI notetaker on every device on your calls. That means:

1. Audit your seats. For each Otter / Fireflies / Granola account in your org, verify whether the training toggle is off. On non-Enterprise plans, assume it's on until you check.
2. Push to Enterprise or off-cloud. Enterprise plans at least give admins one place to enforce a training default. On-device transcription eliminates the surface entirely.
3. Document your legitimate-interest test if you're an EU controller relying on any vendor that trains. The EDPB expects that documentation to exist before ingestion, not after a complaint.
4. Refresh calendar-invite language to disclose AI-assisted transcription. This is a bare minimum for CIPA/BIPA/GDPR transparency and does not, by itself, resolve the training-consent problem for non-users on the call.

For a broader buyer's framework, see our companion piece on what "compliant AI meeting notes" actually means.

How Basil AI solves this

Basil AI was built for exactly this problem. Every transcription runs on-device using Apple's Speech Recognition framework and the Neural Engine — see Apple's Speech framework documentation and Apple's privacy overview. Concretely:

• No vendor server holds the audio. There is no cloud copy of your meeting to include in a training corpus, subpoena, or breach notification.
• No training on your content — architecturally. Basil doesn't have a training pipeline that ingests customer meetings, because it doesn't have customer meetings on its infrastructure.
• No third-party AI providers with access. Not Deepgram, not AssemblyAI, not OpenAI, not Anthropic. Transcription runs on your device; summarization uses Apple's on-device foundation models where available.
• No opt-out toggle to hunt for. The training question doesn't arise. That is the whole point.

Compliance is your determination as the controller. What on-device architecture gives you is a cleaner factual record when you make it: no vendor server holding the recording, no processor to add to your Article 30 register for the transcription step, and no training-corpus question to answer on your DPIA. For the technical detail, see our deep dive on on-device transcription on iOS 26.

The bottom line

The EDPB didn't need to name AI notetakers to change their exposure. Guidelines 03/2026 confirms what has always been true under the GDPR — that scraping personal data at scale to train AI needs a lawful basis, that consent doesn't work at that scale, and that legitimate interest requires documentation your average SaaS notetaker doesn't have. Combine that with the pending U.S. wiretap and BIPA cases, the training-by-default posture that reporting keeps uncovering on the top three vendors, and the durability of model-weight contributions, and the direction of travel is clear. The safest meeting transcript is the one that never leaves your device.

Keep your meetings out of anyone's training data

Basil AI transcribes 100% on-device. No cloud upload, no training pipeline, no third-party AI providers. Your voice never leaves your iPhone or Mac.