Are Kids' Chats Used to Train AI? The Hawaii Class Action

Yes. Unless someone turns a setting off, kids’ chats can end up among the chats used to train AI. On personal ChatGPT accounts, OpenAI’s own help center says it “may use chats and remembered information to improve models” when a toggle called Improve the model for everyone is on, and that toggle is on by default for Free, Plus and Pro accounts. Whether that is legal when the user is a child is now in front of a federal judge.

On October 13, 2026, a Hawaii mother filed a proposed class action in San Francisco on behalf of her 12-year-old daughter and 16-year-old son. Her lawyer’s summary: “OpenAI treats a child clicking a ‘Continue’ button as proof that their parent consented.” Below: how training on chat logs works, what the lawsuit claims, what COPPA does and does not cover, and the three settings you can change tonight.

Key Takeaways

• OpenAI says chats on personal accounts may be used to improve its models while “Improve the model for everyone” is on, which is the default for Free, Plus and Pro.
• The Hawaii suit (Doe v. OpenAI, N.D. Cal., No. 3:26-cv-11781) does not sue under COPPA. It argues a minor cannot validly agree to ChatGPT’s terms under California law, and asks for an equity stake or cash for the class.
• COPPA protects children under 13 and only applies to general-audience services that have “actual knowledge” a user is under 13. A 16-year-old is outside it entirely.
• Memory and model training are different things. Memory personalizes your child’s own chats; training can shape future models for everyone.
• Turn off model training in Data controls, review or delete saved memories, and use Temporary Chat for anything personal. All three take minutes.

What the Hawaii class action claims

The Hawaii case is a proposed class action, meaning one family sues on behalf of everyone in a similar position, and a judge later decides whether the group can proceed together. SFist reported that the children shared personal details, photos, and information about their health, emotions and relationships, and that the complaint blames ChatGPT’s personalized responses, memory, follow-up questions and habit of mirroring the user’s style for encouraging that.

We read the 81-page complaint. Three details matter for parents:

The proposed class is huge. It covers U.S. users who were under 18 when they accepted ChatGPT’s terms and provided data from November 2022 onward.

The legal theory is contract law, not COPPA. The three causes of action are a declaration that OpenAI’s way of getting parental consent is “legally ineffective” under California Civil Code § 1568.5, a declaration that the terms are void under California Family Code § 6701 because minors cannot make certain contracts, and unjust enrichment. The idea: if a child could not legally agree, OpenAI had no valid permission to keep and use the data.

The remedy is unusual. Besides restitution and disgorgement, the family asks for “an equity stake in OpenAI, or the cash value thereof, attributable to data that OpenAI wrongfully took,” proportioned to how much each child contributed.

The complaint also alleges that OpenAI defaults children’s training setting and memory to on, while defaulting business customers to off, and quotes Stanford researcher Jennifer King’s conclusion that “children should never be opted-into data collection by default.” OpenAI’s spokesperson told SFist the company disagrees with the allegations and pointed to ChatGPT for Teens, launched in August 2026. None of the allegations has been tested in court.

This is separate from the eight state lawsuits against OpenAI filed the next day, though Montana’s and Arkansas’s suits also include children’s-privacy claims.

What happens to chats used to train AI

“Training on your chats” sounds like the AI memorizing your kid’s diary. The reality is more mechanical, and parents should know the steps.

A large language model is built in stages. First comes pretraining on huge amounts of text. Then fine-tuning, where the model is adjusted toward better answers using examples and ratings. User conversations can feed the later stages: a company can sample chats, filter them, have human reviewers or automated systems label which answers were good, and use that labeled data to adjust the next version of the model. The complaint notes that a child’s chats “could go directly to a human reviewer,” which is how labeling often works across the industry.

What does a model keep from any single chat? Usually not a retrievable copy. Training nudges billions of numerical weights a tiny amount per example. But Carlini and colleagues (2021), presented at the USENIX Security Symposium, extracted hundreds of verbatim sequences from GPT-2’s training data, including names and phone numbers that appeared in only one document, so memorization is not zero. The bigger concern for families is simpler: the chat itself is stored on company servers, may be read by reviewers, and becomes part of a dataset with commercial value. That value is the heart of the unjust-enrichment claim.

Memory is a different system. ChatGPT’s memory stores facts about the user (“plays soccer,” “has a younger brother”) and feeds them back into that same user’s future chats. It does not, by itself, change the model for anyone else. But OpenAI’s memory help page says that for personal accounts it “may use chats and remembered information to improve models” when the training toggle is on. So memory can become training data too, unless training is off.

What COPPA says, and why it doesn’t cover most teens

The Children’s Online Privacy Protection Act is the main U.S. federal law on kids’ data. Per the FTC’s COPPA FAQ, the rule protects children under 13. It applies to services directed to children, and to general-audience services only when they have “actual knowledge” that a user is under 13. It “does not require operators to ask the age of visitors.” Violations can bring civil penalties of up to $53,088 each, enforced by the FTC and state attorneys general. We explain the details in our guide to what companies owe kids under COPPA.

Here is the gap. ChatGPT’s terms of use set a minimum age of 13 and say users under 18 “must have your parent or legal guardian’s permission.” So the service is, on paper, not for under-13s, and COPPA’s knowledge trigger is hard to prove. Meanwhile 13- to 17-year-olds get no COPPA protection at all. That is why the Hawaii lawyers built their case on California’s rules for contracts with minors instead, and why Montana and Arkansas lean on their own state privacy laws.

How to Teach Your Kid About Where Their Chats Go

Ages 5–8: The postcard game

Write a short message on a postcard and on a sealed envelope. Ask: who can read each one on its way to Grandma? The mail carrier, the sorting office, anyone who picks it up. Explain that typing into a chatbot is more like a postcard than a sealed letter: people who work at the company can sometimes read it. Rule for this age: no names, addresses, school names or photos in any chatbot.

Ages 9–12: Follow the data

Draw a simple flow chart together on paper: “You type” → “Company’s computers store it” → “Maybe a reviewer reads it” → “Maybe it helps build the next version.” Then open the settings with your child and find Improve the model for everyone. Let them be the one to switch it off. Kids remember the settings they changed themselves.

Ages 13+: Read the fine print as a project

Ask your teen to read OpenAI’s Data controls help page and the “Privacy and model training” section of the memory page, then explain to you, in three sentences, what happens to a chat by default. Bonus: compare with one other chatbot they use. Teens tend to be more privacy-protective once they find the defaults themselves.

The question to ask: “If you wouldn’t want a stranger at the company to read this chat, should it be in a chatbot at all?”

Which settings do what

Setting or tool	What it controls	What it does NOT do	Where to find it
Improve the model for everyone (off)	Stops your chats from being used to train future models	Does not delete existing chats or memories	Settings → Data controls
Temporary Chat	Chat is not saved to history; can be non-personalized; OpenAI says it may be kept up to 30 days for safety	Not private from safety review; saving it turns it into a regular chat	New chat menu
Memory off / delete memories	Stops ChatGPT from storing and reusing personal facts in your child’s chats	Turning memory off does not delete past chats	Settings → Personalization / Memory
Delete a chat	Removes that conversation from history	Does not automatically delete a separate saved memory from it	Chat menu
Export data	Gets you a copy of what the account holds	Does not change any setting	Settings → Data controls, or the Privacy Portal
Parental controls (linked teen account)	Some content limits, quiet hours, limited safety alerts	Does not let you read chats	Parent’s account → Parental controls

The memory row has a catch worth repeating. OpenAI’s help page says deleting a chat “does not necessarily delete a separate saved memory created from that chat,” and that it may keep logs of deleted memories for up to 30 days. To truly remove something, delete the saved memory and the chat where it was shared.

What to do at home tonight

Switch off training on every account your child uses

Open Data controls on your child’s ChatGPT account and turn off Improve the model for everyone. OpenAI’s help page notes that signed-out users can do this too, but the choice is saved only in that browser, so check each device. Voice mode has its own separate training control.

Clean up memory together

Open the memory summary with your teen and read it. Many families are surprised at what is in there. Delete anything about health, friends, school or family conflict. If your child is under 15 or so, consider turning memory off.

Teach the “Temporary Chat” habit for personal topics

If your teen wants to ask something personal, a Temporary Chat keeps it out of history and memory. It is not invisible, since OpenAI says it may retain temporary chats for up to 30 days for safety, but it is a better default than a saved chat.

Export before you delete

If you are thinking of closing a child’s account, request an export first through Settings or the Privacy Portal. You will see what was stored, which makes for a better conversation than guessing. For a list of what should never go into a chatbot in the first place, see what kids should never tell a chatbot.

What not to do

Do not assume the teen account handles this for you. The complaint alleges children’s training setting defaults to on, and nothing in OpenAI’s help pages says teen accounts are excluded. Check the toggle yourself. If your teen’s account is linked to yours, OpenAI’s parental controls also let you opt the teen out of model training from your own account.

What to Watch For Over the Next 3 Months

• Week 4: Expect OpenAI to respond to the complaint, possibly by seeking to compel arbitration under its terms or to dismiss. A motion is not a ruling on whether training on kids’ chats is lawful.
• Month 2 red flags: At home, notice if a child starts sharing photos, health details or friends’ names with any chatbot. That is the habit to interrupt, whatever the court decides.
• Month 3 self-check: Open the account settings again. Settings sometimes change after app updates. Is training still off? Is memory what you expect?

Frequently Asked Questions

Does ChatGPT use my child’s conversations to train its models?

On personal accounts, OpenAI says it may use chats and remembered information to improve its models while “Improve the model for everyone” is on, which is the default on Free, Plus and Pro. Turning it off stops future chats from being used for training. It does not delete chats already stored.

Is the Hawaii lawsuit a COPPA case?

No. The complaint does not bring a COPPA claim. It argues under California law that minors cannot validly accept ChatGPT’s terms and that OpenAI was unjustly enriched by their data. COPPA would only cover the under-13 part of the class, and it is enforced by the FTC and states rather than by individual families.

Can my child’s private chats show up for other users?

There is no evidence that ChatGPT shows one user’s chats to another. Training adjusts a model’s internal weights rather than storing a searchable copy. The realistic risks are storage on company servers, human review, and the data’s use in building products, which is what the lawsuit challenges.

Will we get money if the class action wins?

Possibly, but not soon. A judge first has to allow the case to proceed as a class, and OpenAI will likely fight that. If the class is certified and wins or settles, eligible users who were minors from November 2022 onward could receive something. There is nothing to sign up for now.

What is the single most useful setting to change?

Turn off “Improve the model for everyone” in Data controls on each account and device your child uses. Then review saved memories. Those two steps address both training and the stored personal profile.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Jane Doe v. OpenAI Foundation, et al. (2026, October 13). Class Action Complaint, No. 3:26-cv-11781, U.S. District Court, Northern District of California. Via Courthouse News Service. https://www.courthousenews.com/wp-content/uploads/2026/10/doe-v-open-ai-complaint.pdf
2. SFist. (2026, October 15). “Hawaii mother sues OpenAI over children’s data as eight states file separate lawsuits.” https://sfist.com/2026/10/15/hawaii-mother-sues-openai-over-childrens-data-as-eight-states-file-separate-lawsuits/
3. Federal Trade Commission. “Complying with COPPA: Frequently Asked Questions.” https://www.ftc.gov/business-guidance/resources/complying-coppa-frequently-asked-questions
4. Federal Trade Commission. “Children’s Online Privacy Protection Rule (‘COPPA’).” https://www.ftc.gov/legal-library/browse/rules/childrens-online-privacy-protection-rule-coppa
5. OpenAI Help Center. “Data controls in ChatGPT.” https://help.openai.com/en/articles/7730893-data-controls-faq
6. OpenAI Help Center. “Memory in ChatGPT.” https://help.openai.com/en/articles/8590148-memory-in-chatgpt
7. OpenAI Help Center. “Exporting your ChatGPT history and data.” https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data
8. OpenAI. “Terms of Use.” https://openai.com/policies/terms-of-use/
9. Carlini, N., et al. (2021). “Extracting Training Data from Large Language Models.” 30th USENIX Security Symposium. arXiv:2012.07805. https://arxiv.org/abs/2012.07805
10. OpenAI. (2025, September). “Introducing parental controls.” https://openai.com/index/introducing-parental-controls/