AI child safety needs independent testing and tougher age controls, according to former Google safety executive Tom Siegel, who says self-regulation is not enough.
A former Google trust and safety chief is warning that AI child safety could become a bigger problem than the damage critics have blamed on social media.
Tom Siegel, who founded Googleâs trust and safety team and later served as vice president for the division, said AI companies are moving too quickly while protections for children remain incomplete.
Siegel told Reuters that the technology industry risks repeating mistakes made during the social-media era, when products reached huge numbers of young users before regulators and safety systems caught up.
His warning is broader than screen time.
Siegel pointed to risks involving self-harm, unhealthy psychological dependence and what researchers call cognitive offloading, where people rely on an AI system to do thinking they would previously have done themselves.
Those are concerns and risk scenarios, rather than proof that AI chatbots cause those outcomes in every young user.
The difference matters.
AI systems can also help students understand difficult concepts, practise languages and find information quickly. The debate around AI child safety is increasingly about how to preserve those benefits without treating general-purpose AI products as automatically safe for minors.
Siegel has now joined Common Sense Media as the first executive director of its Youth AI Safety Institute.
The Youth AI Safety Institute says it plans to create independent standards, test AI products used by children and publish assessments in a model inspired by vehicle crash testing.
The comparison is deliberate.
Car companies do their own safety work, but buyers do not rely entirely on manufacturers to decide whether a vehicle is safe. Independent testing gives consumers another source of evidence.
Siegel wants a similar system for AI child safety.
The institute launched in May and has already evaluated products from major technology companies. OpenAI Foundation and Anthropic are among its backers, according to Reuters.
Siegel argues that AI companies need stronger age checks before allowing minors to use products designed mainly for adults.
OpenAI has already moved in that direction.
The AI Decode recently covered ChatGPT for Teens, which automatically applies stronger restrictions when OpenAI identifies or suspects that a user is between 13 and 17.
The teen experience adds restrictions around sensitive conversations, expanded parental controls and limits on some features.
OpenAI told Reuters that it uses account information and behavioural signals alongside self-reported age to identify younger users. The company also said no age-assurance method is perfect and that it continues testing its safeguards against attempts to bypass them.
Google pushed back on the idea that its AI search tools leave parents without options.
A Google spokesperson told Reuters that parents can turn off a childâs access to Google Search entirely and argued that AI-powered search can help children learn and explore information.
Siegel wants something more specific. He suggested Google should allow users to disable AI Overviews and AI Mode directly, similar to other content controls.
That disagreement gets to the centre of AI child safety.
Technology companies generally argue that AI can be useful for younger users when appropriate protections are in place. Critics question whether those protections are strong enough when chatbots can hold long, personal conversations and produce confident answers on almost any subject.
Education provides a good example of the trade-off.
A student can ask an AI system to explain algebra step by step.
The same student can ask it to complete the homework.
The AI Decodeâs reporting on AI student learning examined this problem, including concerns that students can skip the difficult reasoning that normally helps them build problem-solving skills.
Common Sense Mediaâs own August survey found that seven in ten U.S. teenagers surveyed used AI for schoolwork. Among teen AI users, 63% said they used it to get answers to school assignments, while many also used it for brainstorming and checking their work.
This is why AI child safety covers more than blocking obviously harmful content.
It includes whether systems encourage dependence, how they respond when a child appears distressed, what information they collect and whether parents understand the productâs limitations.
Siegel also doubts that companies can solve those questions entirely through self-regulation.
His concern is shaped by what happened with social media. Platforms spent years adding protections after products had already become deeply embedded in childrenâs lives, while lawsuits and government investigations followed.
AI companies have an opportunity to act earlier.
Independent testing could help, but only if companies cooperate and evaluators have enough access to test the products realistically.
The next stage of AI child safety will therefore depend less on promises that a chatbot is suitable for young users and more on measurable evidence showing how it behaves when children actually use it.
That evidence is only starting to arrive.
