AI-generated sexual deepfakes are no longer only a celebrity or national-politics story. In a new nationally representative survey, U.S. teenagers described images involving friends, ex-partners, classmates and themselves—and some said they had created or shared the material.
That changes the conversation parents need to have. “Be careful what strangers do with your photo” is not enough when a child may also encounter an image in a group chat, know the person depicted, forward it, or try a tool that can make one.
The survey puts peers at the center of the story
Common Sense Media surveyed 1,314 teenagers ages 13 to 17 in fall 2025. The quantitative survey followed 20 in-depth interviews, and two teen focus groups later reviewed the findings. The report was published July 21, 2026.
Forty-four percent of respondents said they had seen sexual content they believed was generated by AI. Among those teens, 24% said the content depicted themselves or someone they knew personally.
The report also separated exposure from creation. Eight percent of all respondents said they had personally made AI-generated sexual content. Thirteen percent knew someone who had. Because some teens belonged to both groups, the combined measure was 18%, not 21%.
Among that 18%, 63% said the material depicted themselves or people they knew, including peers, friends, and current or former romantic partners. The finding does not show that 63% of all teenagers were personally targeted. It shows that, within the smaller group connected to creation, known people were usually part of the story.
Sharing was also embedded in ordinary communication channels. Nine percent said they had personally sent or shared AI-generated sexual content, and 22% knew someone who had. Among those connected to sharing, disappearing social posts, private group chats and direct messages were common routes.
The missing conversation is larger than a warning about victimization
Only 20% of respondents said they had ever discussed AI-generated sexual content with a trusted adult. Thirty-four percent said they wanted to talk with an adult about the issue but did not know how to start. Forty-two percent said they would feel comfortable speaking to a parent or guardian about disturbing AI-generated content.
The gap matters because the roles can overlap. A teenager might be depicted, receive an image, know who made it, forward it without thinking through the harm, or be afraid that asking for help will trigger blame or the immediate loss of a phone.
Common Sense Media’s parent guidance recommends beginning with a less accusatory question about what teens have heard or seen among peers, rather than opening by asking whether the child committed an offense. That is guidance from one child-media organization, not a tested universal script. Its useful principle is to keep the route to help open while making the boundaries explicit.
A complete family conversation can name four things plainly:
- Do not create a sexual image of another person, whether the image looks real or not.
- Do not forward or save one because it arrived as a joke, a dare or gossip.
- Consent still applies when software generated or altered the image.
- If an image appears, the first job is to get help and protect the person involved—not to begin with punishment or public confrontation.
Laws and school procedures vary by country and jurisdiction. Families facing an actual incident may need prompt help from the school, platform, local authorities, or a qualified legal or child-safety professional. The preventive point is simpler: a child should know before anything happens that asking for help will not make the parent disappear behind anger.
What this survey does—and does not—establish
The sample represents U.S. teenagers ages 13 to 17, not younger children or teenagers everywhere. It was fielded in fall 2025, several months before publication. The measures rely on self-report about a sensitive subject, and respondents classified content they believed was AI-generated rather than having every image technically verified.
The survey cannot establish how any one image affected a teenager, whether exposure caused later beliefs, or whether a particular conversation prevents harm. It does establish that the issue is not confined to distant celebrities: reported creation and sharing frequently involved people inside teenagers’ own social circles.
For a parent, that is enough to change the opening question. Ask not only, “What if someone does this to you?” Also ask, “What would you do if a friend sent one, if you knew the person in it, or if someone asked you to make one?”
Four more AI product changes worth understanding
Meta AI is moving from replies to scheduled and connected actions
Meta says its AI app and website have begun rolling out planning, email and calendar connections, recurring briefings, research reports, slide generation and real-time steering in select markets. A user can set a weekly plan or preferred briefing time once and receive later updates without repeating the request.
The announcement is a product rollout, not independent evidence that the assistant will complete every plan accurately. The visible change is still substantial: consumer assistants are moving from a conversation that ends when the tab closes toward work that retains context, connects to other services and returns later. Permission, data access and review become part of ordinary household use, not just enterprise administration.
OpenAI Presence packages the rules around a customer-service agent
OpenAI launched Presence for enterprise voice and chat agents. Each deployment begins with a defined job, limited knowledge and system access, approved actions, policies, evaluations and conditions for escalating to a person.
OpenAI says Presence handles its English-language phone-support channel and resolves 75% of inbound issues without human assistance. The company also reports that a Codex-powered improvement process reduced human handoffs by 15 percentage points in 10 days. These are first-party operating figures, not an independent audit.
The product makes an otherwise abstract shift easy to recognize: the “person” answering a support call may increasingly be software. The important questions are what it can see, what it can change, and how quickly a human can take over.
Claude for Teachers is for educators, not students
Anthropic launched Claude for Teachers for verified K-12 educators in the United States. The offer includes free premium access, teaching skills and a Learning Commons connection that maps curricular materials to academic standards in all 50 states. It also connects to tools for lesson materials, assessment and classroom feedback.
Anthropic’s published terms say the offer is not for students and that it does not train models on participating teachers’ inputs or outputs. A separate written agreement with a school or district takes precedence over the general offer terms.
Those facts give parents a more useful set of questions than “Does the teacher use AI?” Ask which materials ground the output, what the teacher reviews, whether student information enters the system, and which school agreement governs the use.
Opus 5 trades the top Anthropic tier for a cheaper everyday model
Anthropic released Claude Opus 5 on July 24. It is now the default model for Claude Max and the strongest model for Claude Pro, priced at $5 per million input tokens and $25 per million output tokens. Anthropic positions it near its higher-end Fable 5 model at half the price and highlights coding, computer-use and professional-work evaluations.
The company also describes Opus 5 as more persistent at checking and revising multi-step work. Most of the detailed examples and performance comparisons come from Anthropic and early-access customers, so they should be read as launch evidence rather than neutral proof of superiority.
For non-specialists, the useful change is not another leaderboard. More capable everyday models are being designed to keep a task in view, use tools, test their work and continue longer. As that happens, knowing when a person must review or interrupt the work matters as much as knowing which model scored highest.
This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.