> ## Content Index
> Fetch the complete content index at: https://www.maragsanchez.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Ambiguous Promises of AI Therapy Scribes
- URL: https://www.maragsanchez.com/ai-scribes/
- Published: 2026-11-11T23:51:00.000Z
- Updated: 2026-11-17T15:39:28.000Z
- Description: A look into the privacy claims, policies, and disclosures of companies selling AI documentation tools to therapists.
- Author: Mara

One of the most dreaded parts of a therapist’s job is documentation. In the last few years, AI has entered the healthcare space with the promise of alleviating that burden and making documentation a much simpler task. The solution is usually sold to therapists as a subscription service with the ability to record a therapy session, transcribe it and generate a progress note, all within a couple of minutes. 

These services have gained a lot of popularity amongst therapists, being adopted by solo and group practices, community mental health organizations, and even large behavioral health entities. 

Intrigued after what felt like the *hundredth* ad trying to sell me some version of these tools I decided to look into the specifics behind them. 

With endless options to choose from, I saw a handful of names being recommended again and again by fellow therapists: Blueprint, Upheal, Twofold, Mentalyc, and Berries. The [comparisons I saw online](https://www.upheal.io/comparisons/upheal-vs-mentalyc?ref=maragsanchez.com) were mostly marketing pieces by the same companies reviewing competitors and ultimately [recommending their own product](https://www.mentalyc.com/blog/best-note-taking-app-for-therapists?ref=maragsanchez.com) over the rest; so I decided to compare them myself.

I skimmed their websites and their features seemed very similar: session capture options, note generation in multiple formats, AI clinical companions. The claims on the sites’ copy were also pretty standard and unremarkable: HIPAA compliant. Secure and Encrypted. Custom Templates. Business Associate Agreement in place. 

I then reviewed their public policies and disclosure documents (Privacy Policies, Terms of Service, BAAs, etc) and that’s when I started having more questions than answers. After just a quick first pass, I noticed glaring inconsistencies, ambiguous language, or terms that clearly contradicted their own security and privacy claims. So I began reading them more carefully and mapping out my questions on the wide gap between what these companies said on their marketing and what their policies actually described. This article is about that gap. 

*\*quick note before diving in: I am not affiliated with or sponsored by any of these companies, and the TL;DR is that I don’t recommend any of them nor do I have alternatives to suggest.* 

### *Your Data Is Never Shared* 

I initially imagined the service to be pretty self-contained: ambient audio recording captures the therapy session, a session transcript is created, then a session note is generated from that transcript, the therapist reviews the note and adds it to the client’s medical record. I assumed that to be a private and closed loop. 

I was wrong. My first surprise was learning that what seems to be a simple workflow that takes a couple of minutes, is actually a messy data journey with many other third parties involved.

The company you pay for the service is just the therapist-facing interface that relies on many third party subcontractors for storage, transcribing, processing, analytics, etc. Usually the audio file from the session is stored using one vendor, then sent to another one who does the transcription, then that transcript can go to an AI model provider, like OpenAI or Anthropic, to generate the actual therapy note. 

I didn't find a full list of sub-processors included on any of the disclosure documents I reviewed, even though sub-processors and third party vendors were referenced on them multiple times. Upheal does link a [Trust Center](https://trust.upheal.io/?ref=maragsanchez.com) on their [Privacy Policy](https://www.upheal.io/privacy-and-compliance/privacy-policy-healing-professional?ref=maragsanchez.com) and 16 sub-processors are listed there. I found Berries [Security Center](https://trust.heyberries.com/?ref=maragsanchez.com) linked on their Security page, which lists 9 sub-processors. Others, like Twofold or Blueprint, [acknowledge some third party processors](https://www.trytwofold.com/security?ref=maragsanchez.com#infrastructure), but don’t publish a comprehensive list anywhere I could find. 

Many of these sub-processors have sub-processors themselves, so the chain of third party vendors handling data can keep extending. Which ones touch PHI? For what purpose? Under what conditions? I couldn’t tell from the publicly available documents I reviewed. 

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Upheal-Subprocessors-112026-2.png)

Upheal’s 16 sub-processors and 7 (of the 45) downstream sub-processors for Deepgram.

While these can be all standard, HIPAA compliant business relationships if under a solid BAA when required, it also means clients’ records are likely shared with and handled by many more players than the one company selling the service to therapists. Each additional processor handling client data creates another place where those records are now stored, processed, or retained. Each adding risk to the privacy and security we promise clients. 

### *Deleted After Processed*

Each company seemed to have wildly different policies about what is kept, what is deleted, or how long records are kept. 

Mentalyc states on its [Security page](https://www.mentalyc.com/security?ref=maragsanchez.com) audio recordings are deleted after the note is created but does not mention that policy anywhere I could find on their official policy documents. Instead their [Terms of Use](https://www.mentalyc.com/terms-of-use?ref=maragsanchez.com) point to the Privacy Policy for data retention, the [Privacy Policy](https://www.mentalyc.com/privacy-policy?ref=maragsanchez.com) doesn’t mention a retention policy for audio recordings and points to a Data Retention and Deletion Policy, a document I could not find anywhere. 

Even within the same page I found inconsistencies and ambiguities about this. Twofold's homepage, for example, states that Twofold does not store audio from sessions “at any time” while further down stating there’s the option to retain audio, an option they reiterate on their [Terms of Use](https://www.trytwofold.com/terms-of-use?ref=maragsanchez.com).

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-09---Twofold-Homepage_-1.png)

**Source:* [**Twofold Homepage*](https://www.trytwofold.com/?ref=maragsanchez.com)**, accessed 11/9/2026*

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-13-at-3.44.00-PM.png)

**Source:* [**Twofold Terms of Use*](https://www.trytwofold.com/terms-of-use?ref=maragsanchez.com)**, accessed 11/9/2026*

The sub-processors and their policies matter here too. 

In an [article](https://help.blueprint.ai/en/articles/9758067-data-security?ref=maragsanchez.com) on their Knowledge Base, Blueprint names Deepgram as their sub-processor for audio transcription and states audio is deleted after they receive the transcript from Deepgram. But does Deepgram also delete the audio they received and the transcript they created? I couldn’t find any confirmation on Blueprint’s website. When verifying Deepgram’s data retention policy, I learned that their [Model Improvement Partnership Program](https://developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program?ref=maragsanchez.com) retains audio to train and improve their service. When I asked Deepgram about data retention under a BAA, a representative confirmed that even with a BAA in place, their customers are not automatically excluded from the MIP program and have to opt out of regular data retention terms on each request. Could there be a different agreement that opts Blueprint out of that setting? Absolutely, but only Blueprint and their executed BAA with Deepgram would confirm this. Otherwise therapists can only hope their session audio is not indefinitely retained and used by a third party sub-processor to improve their product.

### *No AI Training On Your Data*

All of the companies I looked into had some version of that phrase on their website, however, none of them defined exactly what that meant. The vagueness of that term can feel both reassuring and unsettling, depending on how it is interpreted. 

First, these assurances are usually referring to clients' [Protected Health Information](https://www.hhs.gov/hipaa/for-professionals/privacy/laws-regulations/index.html?ref=maragsanchez.com#what) (PHI), since de-identified data is not protected by HIPAA and therefore very much available to train AI with. So the original session transcript with names and birthdates might not be training AI, but the same transcript scrubbed of these [18 identifiers](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html?ref=maragsanchez.com#safeharborguidance) can be. Which to me doesn’t make much of a difference. I am certain I frequently have sessions that, if transcribed, the original transcript and the de-identified one would be identical. Does that mean they’re okay to share or that the Safe Harbor method of de-identification is woefully inadequate for contextual narrative as rich and personal as a therapy transcript?

Upheal seemed to be the exception here. They transparently disclose AI training and a lot of their marketing highlights dual consent as their differentiator. [This article on their blog](https://www.upheal.io/blog/privacy-of-ai-notetakers?ref=maragsanchez.com) assures users their session data is never used to train AI models unless both therapist and client have made "an active, informed choice to participate"; which I found difficult to reconcile with a different article on their support center explaining [how to opt-out of AI improvements](https://upheal.io/support/en/articles/8676583-opting-out-of-de-identified-data-use-for-ai-improvements?ref=maragsanchez.com), clearly labeling therapists being opted-in as default. A setting being active by default is not an active, informed choice to participate.

The language describing data use can also be full of vague umbrella terms like service improvement, product enhancement, platform development. In which case, the definition of AI training must be very narrow and specific for it to fall outside the enhancements, improvements, and developments of an AI service. 

Mentalyc, for example, states “Your sessions never train AI” on their [Security page](https://www.mentalyc.com/security?ref=maragsanchez.com), with no explanation of what that means. Their [Terms of Use](https://www.mentalyc.com/terms-of-use?ref=maragsanchez.com) grant them an irrevocable license to use derivatives of User Data (recordings, transcripts, and notes) to develop, test, and improve their services and features. Is that PHI being used for product development and enhancement? I couldn’t find that use disclosed on their [BAA](https://www.mentalyc.com/business-associate-agreement?ref=maragsanchez.com), the document that governs HIPAA compliant use of PHI. Is that because it is considered an essential part of providing their service? If so, why is it an [optional feature that can be opted out of](https://www.mentalyc.com/pricing?ref=maragsanchez.com) on higher priced tiers?

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-16-at-3.53.24-PM.png)

Source: [Mentalyc Pricing Page](https://www.mentalyc.com/pricing?ref=maragsanchez.com), accessed 11/9/2026

These clauses can also change without notice, and unless those changes are announced and explained by the company, users don’t have a way to know whether a change in a document implies a shift in their practices or just wording preferences. 

Berries, for example, has two different versions of their Terms of Use on their website, both marked as last updated on the same date. On a desktop browser, I initially accessed one version, but when resized the same browser to a smaller width the document changed. The Account Ownership section on one version reassures "Berries does not use Your Data, Protected Health Information (PHI), or any patient-identifiable or de-identified patient content to train, retrain, or fine-tune artificial intelligence, large language, or machine learning models; while the other version grants Berries rights “to create artificial intelligence and machine learning models used to support the Services”. As of the time of publishing this article, both versions remain active, both under the same Berries url, both last updated on November 4th, 2026\. Which one are therapists agreeing to when they sign up for the services? 

[![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-13-at-4.08.50-PM.png)](https://heyberries.com/terms-of-use-agreement?ref=maragsanchez.com)

**Source:* [**Berries Terms of Use*](https://heyberries.com/terms-of-use-agreement?ref=maragsanchez.com) *(large browser width),* accessed 11/11**/2026*

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-13-at-4.05.59-PM.png)

**Source:* [**Berries Terms of Use*](https://heyberries.com/terms-of-use-agreement?ref=maragsanchez.com) *(smaller browser width)*, accessed 11/11**/2026*

I’ll spare you the long detour I took trying to understand how AI models are trained and the difference between pre-training, fine-tuning, and prompt optimizing; or the hours down the rabbit hole of self-hosted models vs those running on cloud infrastructure, etc. The more important point here is that therapists shouldn’t have to become software engineers to decipher what goes on behind the scenes with their clients’ data or what is meant by ambiguous buzzwords. The burden of clarity should rest with the companies providing the service, not the therapists using it.

### *We Take Privacy Seriously* 

When I started stepping back from specific findings, a general pattern became clear. I noticed the vague language and ambiguous clauses were everywhere. There was sloppy drafting across multiple documents, and copy that simply did not make sense to me.

Like Upheal’s [US Data Processing Agreement](https://www.upheal.io/privacy-and-compliance/us-dpa?ref=maragsanchez.com) referencing annex IV, but there being no annex IV, the document ends after annex III; even though the initial section establishes “(d) Annexes I to V are an integral part of the Agreement”. Or an answer from [Mentalyc's Security FAQ](https://www.mentalyc.com/security?ref=maragsanchez.com) affirming notes are unlikely to be requested in court if their app is used “just for reference or education, without including client details”. Or Twofold's [Security page](https://www.trytwofold.com/security?ref=maragsanchez.com#access) reassuring customers “no one at Twofold Health can access your session transcripts, clinical notes, or patient names”, while their BAA states their workforce reviews PHI to evaluate documentation quality. 

![](https://storage.ghost.io/c/70/5a/705a1e22-caae-4ac1-9157-041eeff26c8c/content/images/2026/11/Screenshot-2026-11-13-at-4.25.43-PM.png)

Source: [Twofold BAA](https://www.trytwofold.com/terms-of-use?ref=maragsanchez.com), accessed 11/9/2026

After a while, I wondered if all of this could be explained by accidental oversight, careless mistakes on the backend, or errors that go unnoticed. That’s a possible explanation, but not a very reassuring one. Why would I entrust something as important as my clients’ stories to a business being so careless with their own legal policy documents? 

Overall, the whole process felt fragmented and effortful, like it was on me to piece together multiple clauses from multiple documents from multiple places to get a somewhat decent (*emphasis on somewhat*) understanding of what I would be agreeing to by using these services. Even then, I didn’t feel remotely comfortable enough to explain it to clients, let alone ask for their informed consent to use it. 

*How does one ask a client if they’re okay with an account of their most vulnerable moments being subject to ambiguous policies for the sake of convenience?*

Bottom line: I’ll continue to dread documentation and write notes myself. 

---

## Sources 

*This article reflects the publicly available policies, legal documents, and company materials reviewed. Documents may have changed since publication.*

### Blueprint

- [Data Security](https://help.blueprint.ai/en/articles/9758067-data-security?ref=maragsanchez.com) — Blueprint Product Knowledge Base. June 5, 2026.

### Upheal

- [Protecting Your Practice: A Guide to the Privacy of AI Notetakers](https://www.upheal.io/blog/privacy-of-ai-notetakers?ref=maragsanchez.com) — Upheal. November 4, 2026.
- [Opting Out of De-identified Data Use for AI Improvements](https://upheal.io/support/en/articles/8676583-opting-out-of-de-identified-data-use-for-ai-improvements?ref=maragsanchez.com) — Upheal Support Center. July 22, 2026.
- [Collecting Client Consent for Data Processing](https://upheal.io/support/en/articles/8614351-collecting-client-consent-for-data-processing?ref=maragsanchez.com) — Upheal Support Center. July 24, 2026.
- [US Data Processing Agreement](https://www.upheal.io/privacy-and-compliance/us-dpa?ref=maragsanchez.com) — Upheal. Effective July 10, 2026.
- [Privacy Policy for Mental Health Service Providers](https://www.upheal.io/privacy-and-compliance/privacy-policy-healing-professional?ref=maragsanchez.com) — Upheal.

### Twofold

- [Homepage](https://www.trytwofold.com/?ref=maragsanchez.com) — Twofold Health.
- [Security & Trust](https://www.trytwofold.com/security?ref=maragsanchez.com) — Twofold Health.
- [Terms of Use](https://www.trytwofold.com/terms-of-use?ref=maragsanchez.com) — Twofold Health, *Your Data* and *Retention of Your Data*, and the Business Associate Agreement incorporated into the Terms.

### Mentalyc

- [Security](https://www.mentalyc.com/security?ref=maragsanchez.com) — Mentalyc.
- [Pricing and Feature Comparison](https://www.mentalyc.com/pricing?ref=maragsanchez.com) — Mentalyc.
- [Terms of Use](https://www.mentalyc.com/terms-of-use?ref=maragsanchez.com) — Mentalyc. Updated March 2026.
- [Privacy Policy](https://www.mentalyc.com/privacy-policy?ref=maragsanchez.com) — Mentalyc. Updated March 2026.

### Berries

- [Terms of Use Agreement](https://heyberries.com/terms-of-use-agreement?ref=maragsanchez.com) — Berries Health. Last updated November 4, 2026, as displayed on the website.
- [Security](https://heyberries.com/security?ref=maragsanchez.com) — Berries Health.
- [Privacy Policy](https://heyberries.com/privacy-policy?ref=maragsanchez.com) — Berries Health.

### Deepgram

- [Model Improvement Partnership Program](https://developers.deepgram.com/docs/the-deepgram-model-improvement-partnership-program?ref=maragsanchez.com) — Deepgram Developer Documentation.

### HIPAA and De-identification

- [Guidance Regarding Methods for De-identification of Protected Health Information](https://www.hhs.gov/hipaa/for-professionals/special-topics/de-identification/index.html?ref=maragsanchez.com) — U.S. Department of Health and Human Services, Office for Civil Rights. See the Safe Harbor and Expert Determination methods under 45 CFR §164.514.

### 

## Subscribe

On being a therapist in an unraveling world.

Subscribe 

Email sent! Check your inbox to complete your signup. 

No spam. Unsubscribe anytime.