Sanchez questions companies on teen safety and default data settings

Sanchez, Councilmember Phil Wong, and Councilmember Susan Shuang.

I will remind my colleagues to ask your questions up front.

And if you have questions that are specific for some of the members of the panel and not all four, specify.

Thank you, Councilmember Sanchez.

Thank you, Chair.

My first question is in connection to heartfelt testimony of Layla Turner-Scott before Congresswoman Jaya Paul last month.

Ms.

Turner-Scott's son died following chats with OpenAI that instructed him that incompatible drugs were compatible.

She highlighted that throughout the review of vlogs, ChatGPT responded to her son with phrases like, I've got your back, you can talk to me about anything.

And even I love you with heart emojis.

She stated, and it seems true, that the goal of ChatGPT was not to keep her son safe.

It was only to keep him using the product.

What updates have your companies made to their models to engage children and prevent addiction?

Are you limiting your agents from giving medical and health advice without a license?

And my second question is regarding the fact that members of the public and enterprise clients are sharing more and more information with your companies.

You have access to copyrighted information publicly available.

How do your companies...

Excuse me, how do your products currently use information gathered from users?

In an earlier exchanges, Chair, if I may, several of you stated that users can opt out of sharing your data.

What is the default option?

And how are trade secrets and other confidential business or government information from enterprise clients being used to trade AI?

Who have you paid for the valuable data that you are using to train your models?

So I want to start off by acknowledging the tragic incident that you reference.

Any incidents of harm that results from our product or any other product is tragic, and we take it seriously.

I'd like to tell you about ChatGPT for Teens, which was recently released to focus specifically on users ages 13 to 18 and is designed to be safe by default.

That includes safety protections on the type of content that is available.

Limiting content that might be related to suicide or self-harm, eating disorders, violence, or graphic or sexual content.

We also include take-a-break reminders to remind teens to stop using ChatGPT.

And we have parental controls, optional parental controls that

allow parents to add additional layers of safety onto their teen's experience.

One thing that I will highlight here is if a teen indicates signs of suicide or self-harm, we have a feature that allows the parent to be notified.

On your question of privacy and data, as I and my colleagues have shared, we take privacy seriously.

And our core value is that users should have control over their data.

And I'm happy to follow up on specific details of how those controls work at a later time.

What is the default option?

I'm sorry, I didn't hear you.

What is the default option with respect to who owns the data and how it is used?

So users have control over their data.

Everything is privacy protected.

What is the default option?

Everything is privacy protected by default, and users have the option to turn off their data for training or request that it be deleted.

Council member, first off, I just want to acknowledge the terrible tragedy that was

a terrible tragedy in relation to the Turner Scott family and how distressing and upsetting that is to hear and, you know, my sympathies and condolences to that family.

In relation to how Meta approaches minors and young people,

Creating a safe and productive experience is absolutely imperative to Meta, and we're committed to providing a safe and responsible AI, especially for teens.

And we have strong policies and safeguards in place in relation to that.

For example, by default,

Minors only have age appropriate interactions, meaning that content would not be out of place on a 13 plus movie, for example.

We also prohibit our AI offerings from providing licensed professional information like medical information and direct people towards appropriate resources.

In relation to training data, we

take creativity, copyright, and IP rights seriously, and believe our practices are consistent with the law.

We train on information from the internet, information licensed from third parties,

and information shared across our services, and we provide all of the information in relation to how we use that data in our privacy center.

In relation to Muse, as I mentioned, I think a little earlier, individuals interacting with Muse can opt out of their interactions for training.

Ma'am, I want to highlight a few protections that we have in place with kids in mind in particular.

One is for what we call persona, which prevents claims of sentience or simulating relationships and role play.

That helps prevent kids from thinking they're talking to a person and developing a relationship.

We also prevent sexually explicit outputs from chats with kids.

And we also have for all ages protocols on suicide and self-harm, where we refer users who are expressing suicidal ideation or other types of self-harm to crisis resources.

And that approach draws from our deep experience fielding such queries on search.

When it comes to training and the default question you asked, for those over 18, it is default on to use chat data to train our models to improve.

But all users have control over the ability to turn this feature off or to control the length of time such information is kept.

For users between 13 and 17, it's default off.

Teens can choose to turn it on if they wish to train a model through our personal intelligence program.

And for under 13, it is default off as well.

And we never use data for under 13 to train our models.

Thank you.

At Anthropic, we have limited the cloud platform to those only 18 plus.

We're concerned about child well-being.

We monitor and detect for any users that might be 18 or under.

And when that flags, we investigate to make sure that we could remove their access if so.

On training data, I will have to get back to you with more details as it's not my role.

My understanding is that all users must select which option they would like.

But I would refer you to our public documentation where we describe this in detail.

Thank you all.