ai therapy apps and your privacy

the thing about telling a chatbot the worst of it at 3am is that it feels like a diary. it is not a diary. it is a company's server, governed by a privacy policy you clicked through and a set of federal rules that were mostly written before chatbots existed.

this page lays out who is actually responsible for what you type: what hipaa covers and does not, the ftc rule that reaches health apps instead, the three ftc orders against mental-health and health companies, what one state now requires of mental-health chatbots, what four ai companies' own policies say, and what researchers found when they checked apps against their policies. no company is graded here. each policy is quoted for what it says about itself, with the date we read it.

does hipaa protect what you tell an ai therapy app?

usually not, because hipaa is about who holds the data, not what the data is about. the regulation defines a “covered entity” as a health plan, a health care clearinghouse, or “a health care provider who transmits any health information in electronic form in connection with a transaction covered by this subchapter”, which in practice means billing insurance^[1]. a general chatbot, a companion app or a wellness app that you sign up for yourself is normally none of those. the same sentence about your depression is protected health information in your psychiatrist's chart and ordinary consumer data in an app.

the other direction: hipaa-covered digital care exists. a telehealth service that bills your insurance, or an app your health system provides, is generally covered or a business associate of a covered entity^[1]. and utah's law, below, borrows hipaa's rules for the narrow case where a mental-health chatbot shares data with a contractor^[8].

what does cover health apps, then?

the ftc's health breach notification rule, which the commission amended in april 2024 by a 3-2 vote specifically to “underscore the final rule's application to health apps and similar technologies not covered by hipaa”^[3]. the amended rule was published on 30 may 2024 and took effect on 29 july 2024^[2]. two definitions do the work. “health care services or supplies” now means “any online service such as a website, mobile application, or internet-connected device” that tracks a list of things that names mental health explicitly. and a “breach of security” now “includes an unauthorized acquisition … that occurs as a result of a data breach or an unauthorized disclosure”^[2]. in plain terms: an app that shares your health data with an advertiser without your authorization has had a breach, even if nobody was hacked, and must notify you. violations are treated as violations of an ftc trade regulation rule and are subject to civil penalties^[2].

the limit is the scope. the rule is written for “vendors of personal health records” and related entities. whether a general-purpose chatbot that you happen to talk to about your mood is such a vendor has not been tested in any ftc action we could find.

what has the ftc actually done?

three orders, all settlements, none against a chatbot, all about health data used for advertising.

• goodrx, february 2023. the ftc's release says the commission took “enforcement action for the first time under its health breach notification rule,” alleging the company shared users' prescription medications and health conditions with facebook, google, criteo, branch and twilio and failed to notify anyone. civil penalty: $1.5 million^[4].
• betterhelp, march and july 2023. the ftc alleged the online counselling service “used and revealed consumers' email addresses, ip addresses, and health questionnaire information to facebook, snapchat, criteo, and pinterest for advertising purposes.” the final order, approved 3-0 on 14 july 2023, bans sharing health data for advertising and required $7.8 million for partial refunds^[5].
• cerebral, april 2024. the ftc alleged the telehealth company disclosed the “sensitive information of nearly 3.2 million consumers” to third parties including linkedin, snapchat and tiktok through tracking tools. the stipulated order permanently bans most advertising uses of health data, with nearly $5.1 million for refunds and a $10 million civil penalty, $2 million of it due^[6].

these are the ftc's allegations, settled by consent; no court found liability after a trial, and the ftc says it files when it has “reason to believe” the law is being broken^[6]. they matter here because they show what the regulator treats as a breach of trust in mental-health data, which is exactly the kind of data people now type into chatbots.

what did the ftc ask the companion-chatbot companies?

in september 2025 the ftc voted 3-0 to send 6(b) orders to seven companies: alphabet, character technologies, instagram, meta, openai, snap and x.ai^[7]. among the things it asked is how each company uses or shares “personal information obtained through users' conversations with the chatbots,” how it monetizes user engagement, and how it tells users and parents about “data collection and handling practices”^[7]. a 6(b) study is a request for information, not an accusation. no report had come out of it when we checked.

what do state laws add?

utah is the clearest. its 2025 law says a supplier of a “mental health chatbot” “may not sell to or share with any third party” a utah user's individually identifiable health information or “user input,” with exceptions for a health care provider the user consents to, the user's own health plan, and contractors needed to run the service, who must then follow hipaa's rules as if they were covered^[8]. it also bars using a user's input to decide which ads to show them, and requires any ad inside a conversation to be labelled. fines run up to $2,500 per violation^[8]. the law covers only apps that present themselves as able to help with mental health, and only users in utah. california's 2026 children's chatbot law adds similar limits for users under 18 from 1 july 2027: no sale of a child's personal information gathered through a companion chatbot and no targeted advertising based on what the child says in chat^[9]. the full state-by-state picture is on state laws on ai therapy.

what do the companies' own policies say?

each of these is the company describing itself, read on 25 october 2026. a policy says what a company permits itself to do; it is not an audit of what it does.

• openai (chatgpt), us privacy policy updated 10 september 2026: “we may use content you provide us to improve our services, for example to train the models that power chatgpt,” with instructions to opt out. temporary chats “will not appear in your history or be used to improve openai's models” and are deleted within 30 days unless kept for safety or legal reasons; deleted chats are removed within 30 days with the same exceptions. the policy also says free and go users may see personalized ads, and that openai does not “sell” personal data but may share limited data with marketing partners for targeted advertising unless you opt out^[10].
• anthropic (claude), announcement of 28 august 2025: consumer users of claude free, pro and max choose whether their chats are used for training. if they allow it, retention is extended “to five years”; if not, “our existing 30-day data retention period” continues. “if you delete a conversation with claude it will not be used for future model training,” and “we do not sell users' data to third parties.” the change does not apply to business or api use^[11].
• google (gemini apps), privacy hub: human reviewers read some chats, and google asks users, “please don't enter confidential information that you wouldn't want a reviewer to see.” chats reviewed by humans “are not deleted when you delete your activity. instead, they are retained for up to three years.” with keep activity off, chats are kept for 72 hours^[12].
• replika (luka, inc.), policy updated 27 may 2026: “we will never use or disclose the content of your replika conversations for marketing or advertising purposes.” sensitive information you include is processed “including by third-party ai language model providers, solely to generate conversational responses.” the policy separately says it shares website-visitor data such as ip address and advertising id with advertising companies^[13].

we tried to read character.ai's policy as well; it renders only in a browser and we could not read it, so it is left out rather than summarised second-hand.

what do researchers find when they test the apps?

that the policy and the behaviour do not always match. a 2019 study in jama network open intercepted the data sent by 36 top-ranked depression and smoking-cessation apps: 33 of the 36 transmitted data to a third party, and 29 of 36 sent data for advertising, marketing or analytics to just two companies, google and facebook. of the apps sending data to google, only 12 of 28 disclosed it; of those sending to facebook, 6 of 12^[14]. a 2023 analysis of 27 top-ranked mental health apps on google play found “unnecessary permissions, insecure cryptography implementations, and leaks of personal data and credentials in logs and web requests,” and noted that data sharing with third parties and advertisers aggravated the risk of profiling^[15]. both studies predate today's chatbots; what they establish is that a privacy policy is a floor for what to expect, not a description of what happens.

where talking to people fits

we run a peer-support app, so weigh this paragraph accordingly, and hold us to the same standard as everyone above. here is what our own privacy policy says, no more: you post under a handle you choose, a member account does not require your real name or phone number, and your email is not shown to other members. it says we do not sell personal information. it also says we use cookies, and service providers for hosting, data analytics, customer service and marketing assistance, and that veisund is not a hipaa covered entity, so what you share is not protected health information. it is peer support, not therapy: people answer, not a model. no ai writes replies, with one automated exception we name: if a post suggests immediate risk, an automated safety response surfaces crisis resources. you may not get an answer, and what you post is visible to other members, which is its own privacy trade. the evidence on peer support is on is peer support effective, and the wider question of chatbots as therapists is on is it ok to use chatgpt as a therapist.

what to ask before you type the hard thing

questions, not advice. each one comes from a rule or a finding above.

• does the policy say the service is hipaa-covered, or is it an ordinary consumer app?
• are my conversations used to train models by default, and where is the switch?
• can a human employee or contractor read my chats, and for how long are reviewed chats kept?
• when i delete a chat, is it gone in 30 days, or is there a longer exception?
• is there a temporary or incognito mode, and what does the company say it still keeps?
• does the app share anything with advertising partners, and is my conversation content excluded?
• if i am in utah, is this app a “mental health chatbot” that may not sell or share what i type?

adam