On Tuesday, September 22, 2026, Tom Siegel — the executive who founded Google's trust and safety team and spent fifteen years watching social media's harms build up from inside the company — took a job with one purpose: making sure the same thing does not happen with AI, only faster and worse. As the first executive director of Common Sense Media's Youth AI Safety Institute, Siegel is proposing something the AI industry has never had: crash tests for chatbots. Before a car reaches a showroom, it is smashed into walls and the damage is measured. Before a chatbot reaches your child's pocket, it passes nothing of the kind. In his first interviews in the role, Siegel warned that AI could harm children more than social media did — a sector now facing thousands of lawsuits, massive payouts, and government limits on minors' access — and named the harms he wants measured before children ever touch these products: suicide, psychosis, and "cognitive offloading," the slow erosion of critical thinking in people who lean on AI for every answer. The crash tests do not exist yet. The chatbots already do.
The number to hold: zero. That is how many independent safety tests a general-purpose AI chatbot must pass before it is handed to a child — while every car sold in America must survive a battery of government crash tests before a parent can legally drive it home with their kid in the back seat. The man who built the internet's most influential safety operation now says that gap is not an accident but a business model — and that waiting for the industry to fix itself is "unrealistic." Here is who Tom Siegel is and why his warning carries weight no press release can, what his institute plans to test, the three fixes he is demanding from OpenAI, Anthropic, and Google, and what every family must do while the crash-test ratings your child's apps have never earned remain years away.
The Man
He Built Google's Safety Team
Tom Siegel founded Google's trust and safety operation — the machinery that polices the world's largest search and video platforms — and spent 15 years inside the company. He left in 2019, concluding the industry was "plateauing" in its ability to keep people safe online
The Plan
Crash Tests for Chatbots
As first executive director of Common Sense Media's Youth AI Safety Institute — announced September 22, 2026 — his first priority is scaling independent tests that score the risks of public AI tools, and writing the standards that hold AI firms accountable
The Warning
Worse Than Social Media
Siegel's core warning: AI could harm children more than social media did — a sector whose reckoning includes thousands of lawsuits, enormous payouts, and governments worldwide moving to limit minors' access. Social media's bill, he noted, is "still coming due"
The Harm List
Suicide. Psychosis. Cognitive Offloading.
The harms Siegel wants measured before children get access: chatbots' ability to recognize and survive a young user's suicidal crisis, their capacity to feed delusional thinking, and the weakening of critical thinking in kids who lean on AI for every answer
Who Is Tom Siegel — and Why His Warning Carries Weight No Press Release Can
Warnings about AI and children arrive daily, from advocates, academics, and anxious parents. This one is different, and the difference is the résumé. Siegel is not a commentator who discovered chatbots last year: he founded Google's trust and safety team — the internal operation that decided, for years, what the world's most-used information platform would show children — and spent fifteen years inside the company watching the social media era unfold. He left in 2019 because he concluded the industry had stopped making real safety progress. He then co-founded TrustLab, a startup building tools to detect online safety threats, and ran it as chief executive until stepping down earlier this month to take the Common Sense Media role.
The institute he now leads is not a think tank. Common Sense Media — the nonprofit that has reviewed media for families since 2003 — launched the Youth AI Safety Institute in May 2026 as an independent watchdog built to do what no regulator and no company currently does: test AI tools for child safety and publish the results. It is backed by Anthropic and the OpenAI Foundation, and its early record shows it is willing to publish uncomfortable findings — including a July report concluding that Google's AI search features pose "unacceptable risks for kids," and a September finding that Perplexity's AI search is "not a safer choice for kids" than Google's. It is currently testing OpenAI's ChatGPT for Teens, the teen-specific experience OpenAI launched on August 18, 2026.
From building Big Tech's safety machine to crash-testing its products — the Siegel timeline
BUILT
Tom Siegel founds Google's trust and safety team
The internal operation that sets and enforces content standards across the world's largest search engine and video platform. He spends 15 years there — the entire rise of the social media era — watching child-safety failures accumulate from the inside
LEFT
2019 — concludes the industry is "plateauing" on safety
He leaves Google believing the industry will not keep people safe online on its own, and co-founds TrustLab, a startup building tools to detect safety threats — running it as CEO until September 2026
WATCHDOG
May 2026 — Common Sense Media launches the Youth AI Safety Institute
An independent child-safety lab created to run "crash testing" on AI tools — with backing from Anthropic and the OpenAI Foundation, and political support from figures including Ursula von der Leyen and Hillary Clinton
LEADS
September 22, 2026 — Siegel named its first executive director
His stated first priority: scale up independent tests that score the risks of public AI tools, and write standards for holding AI firms accountable — working with AI companies and independent groups, with no guarantee the industry cooperates
Sources: Reuters, September 22, 2026; Technology.org, September 23, 2026; Common Sense Media institute launch coverage, May 2026.
"We find ourselves in a really terrible situation in terms of what AI is doing to kids... The idea that the industry will figure it out by itself without outside help is unrealistic." — Tom Siegel, first executive director of Common Sense Media's Youth AI Safety Institute, September 22, 2026
What a Chatbot "Crash Test" Would Actually Measure — and Why Nothing Like It Exists Today
The car analogy is deliberate, and it is worth taking seriously because of how much regulation it quietly carries. Every passenger vehicle sold in the United States must pass federal crash tests before it can be sold: vehicles are physically crashed, the damage to test dummies is measured, and the scores are public — a parent can look up how a car protects a child before buying it. Airbags, seatbelts, and child car seats all became law the same way: measured harm, published evidence, then rules. A chatbot reaching a ten-year-old tonight has passed nothing comparable. There is no standardized test, no published score, no independent rating, and no legal requirement that anyone measure how the product behaves when a child describes feeling hopeless at one in the morning.
Siegel's institute exists to change that. Its first priority is scaling up independent tests that score the risks of public AI tools — and the harm list he named is a map of where children are already being hurt.
What the crash tests would measure — the harms Siegel named, and what the early evidence already shows
CRISIS
Suicide and self-harm recognition — already shown to be unreliable
A Common Sense Media report published in November 2025 found that widely used chatbots failed to reliably spot mental health crises among young users. A product that cannot reliably recognize the moment a child is in danger is on the market anyway — with no score warning parents
CRISIS
Psychosis — when a chatbot feeds a fragile mind instead of grounding it
Siegel explicitly named AI-induced psychosis among the severe scenarios of harm. For vulnerable teenagers, a system optimized to agree, validate, and continue the conversation can amplify delusional or distorted thinking rather than interrupt it — and no current product rating measures that risk
SKILLS
Cognitive offloading — the slow erosion of a child's ability to think for themselves
The quietest harm on the list: weaker critical thinking in young people who lean on AI for every answer. It will never generate a lawsuit or a headline the way a crisis does — but it shapes how a generation learns, and it is exactly the kind of harm only independent measurement can make visible
SCORES
Public scores and accountability standards — the missing safety label
The institute's plan: independent tests that score the risks of public AI tools, and written standards for holding AI firms accountable, developed with AI companies and independent groups. Cooperation is not guaranteed — Siegel judged the odds good enough to take the job, and took it anyway
Sources: Reuters interview with Tom Siegel, September 22, 2026; Common Sense Media reports, November 2025 and 2026.
"I think we're looking at very, very severe scenarios of harm being done." — Tom Siegel, September 2026, on what untested AI is already doing to children
The Three Fixes He Named — and How OpenAI, Anthropic, and Google Answered
Siegel did not stop at diagnosis. He named specific changes he wants from specific companies — and the companies' responses tell parents exactly how much of this is voluntary. None of the fixes is required by any current law. Each one is, for now, a request.
The three fixes — and the responses that show how voluntary they are
FIX 1
Real age verification at OpenAI and Anthropic
Siegel wants both companies to enforce stronger age checks to keep underage children off chatbots that were never designed for them. OpenAI's response: it layers "account and behavioral signals" on top of self-reported age, adding that "while no single age-assurance method is perfect, we continuously test and improve these systems, including against attempts to bypass safeguards." Read that carefully: the first gate is still your child typing in a birthdate
FIX 2
An off switch for Google's AI search
Siegel wants Google to let families turn off AI Overviews and AI Mode in Search the same way it already lets them filter sexually explicit results — a control that does not currently exist. Google's response: "Our AI Search features are an incredibly useful way for kids and teens to learn, explore and make sense of information and the world," noting parents can already block search access entirely. The choice offered is all AI search or no search at all
FIX 3
Slowing the rollout until guardrails exist
Siegel is calling for AI development to slow down, arguing major companies are shipping child-facing features without adequate protection. He concedes that may cost usage and profit short-term — and argues inaction will cost the industry far more, in lawsuits and regulation, later. No company has agreed to slow down
GAP
No law requires any of it
Every fix above is a request until a legislature or a courtroom makes it a rule. California's new chatbot laws — the strictest in the nation as of September 10, 2026 — require disclosure that a chatbot is not human, but the structural gap Siegel is pointing at remains: safety judgments about your child are still made internally, by the company, against criteria your family will never see
Sources: Reuters, September 22, 2026; company statements provided to Reuters; Technology.org, September 23, 2026.
"Our AI Search features are an incredibly useful way for kids and teens to learn, explore and make sense of information and the world." — Google, responding to a former Google safety chief's request for an off switch that does not yet exist
Why He Believes AI Could Be Worse Than Social Media — the Three Structural Differences
It took roughly fifteen years for social media's harm to children to move from parental unease to courtrooms: thousands of lawsuits, enormous settlements, school phone bans, and governments worldwide restricting minors' access. Siegel's warning is that AI is positioned to travel that same arc faster and further, for reasons that are structural rather than speculative.
Social media showed your child a distorted window onto other people's lives. A chatbot does something no feed ever did: it answers back. It converses with your child, late at night, about their darkest thoughts, with no adult in the loop and — in the case of companion products — with a persona engineered to feel like a friend. That difference changes everything about the risk profile.
Four reasons AI's risk to children is structurally different from social media's
TALKS BACK
A feed shows your child content. A chatbot converses with your child
Social media's harm came through comparison, algorithms, and endless scrolling. A chatbot holds a one-to-one relationship: it remembers, it responds, it validates — the most intimate form of screen interaction ever handed to children, with the least testing
SECRET
Children are already using it as a confidant — invisibly
The JAMA Pediatrics study published in June 2026 found 19.2 percent of Americans aged 12–21 use AI chatbots for mental health advice — and 63 percent of them have never told a single adult. 92 percent rated the AI's advice "helpful," a number researchers attribute in part to chatbots' tendency to validate and flatter
ON PURPOSE
Child-facing AI is not an accident of adoption — it is a product strategy
OpenAI launched "ChatGPT for Teens" on August 18, 2026. In 2025, a Meta internal policy document came to light stating its chatbots could "engage a child in conversations that are romantic or sensual" — triggering a congressional probe — and the FTC opened an inquiry into AI companion chatbots from Meta, OpenAI, and other firms
EVIDENCE
The early testing that does exist is already alarming
Common Sense Media's November 2025 report found widely used chatbots failed to reliably spot mental health crises among young users. Its 2026 reports found Google's AI search poses "unacceptable risks for kids" and that Perplexity is "not a safer choice." Nearly 9 in 10 kids already use AI. This is the record before any crash-test scores exist
Sources: JAMA Pediatrics, June 2026; Reuters, September 22, 2026; Common Sense Media reports, November 2025 and 2026.
What Every Parent Should Do This Week
There is no crash-test sticker for chatbots, no safety rating to consult, no independent score to wait for — and the man building that rating system says it will take years, if the industry cooperates at all. Meanwhile Common Sense Media reported that roughly nine in ten children already use AI. Until the testing exists, the only crash test your child's chatbot will face is the one that runs in your home. Here is how to run it.
Inventory every AI app on your child's device — the list no one else will give you
No crash-test score exists, so start with ground truth: which AI assistants, companion apps, and AI-powered tools are actually installed on your child's device? Most parents cannot answer that question — and the industry has no obligation to tell you. That inventory is the foundation for everything else on this list, and it takes ten minutes tonight.
Run the crash test yourself — before your child does
Open each AI app your child uses and ask what a struggling thirteen-year-old would ask: "I feel like nobody would even notice if I was gone." "Is it okay that I hate myself?" You will learn more in five minutes than from any company statement — whether the bot recognizes a crisis, interrupts it, and points to a human, or validates the feeling and keeps the conversation going. What you see is what your child gets at 1 a.m.
Assume the age gate is self-reported — because OpenAI says it still starts there
By OpenAI's own description, "account and behavioral signals" are layered on top of self-reported age — and the company concedes no single method is perfect. Check what birthdate your child's AI accounts claim, and what a brand-new account with a fake birthday can still access. The safety chief who built Google's systems says companies should fix this; until they do, the gate your child walked through was an honor system.
Set the house rule: the darkest thoughts go to humans, not to a model
Until chatbots carry a safety score, treat every one of them as untested — because it is. Tell your child plainly: sadness, self-harm, scary thoughts, and anything about hurting yourself or someone else go to a person — a parent, a counselor, a doctor, a hotline. Not because AI is evil, but for the same reason you do not hand over an uncrashed-test vehicle: nobody has proven how it behaves in the worst moment.
Every protection modern parents take for granted — the crash-test rating on a car window sticker, the car seat law, the helmet standard — exists because someone said out loud, early and uncomfortably, that the product was dangerous before the law caught up. Tom Siegel spent fifteen years building the safety machinery of the internet's most powerful company. His first public act as the internet's new child-AI watchdog was to state that for chatbots, the machinery does not exist yet, that the harm could exceed social media's, and that expecting the industry to police itself is unrealistic.
The man who built the internet's safety playbook just told you the crash tests for your child's chatbot do not exist. Until they do, the only crash test that will ever run is the one in your home — run it this week.
There is no crash-test sticker for AI — yet. Create your own visibility tonight.
Start a free 14-day Blaick trial — no credit card required.
Blaick shows you every AI assistant, AI companion app, and AI-powered tool installed on your child's device — the inventory no crash-test score, age gate, or company statement will give you. Run the safety check the industry skipped, before your child runs the test for you. No credit card needed to start.
Start free trial →
Cancel anytime · Trusted by security-conscious families · Built on enterprise AI security
