Technology
The Idea Is Fine. Then the Humans Arrive.
Protecting children from real AI risks makes sense. The problem begins when “safety” has to be defined by committees full of people with opinions, agendas and very human ideas about how everyone else should live.
Tom Siegel spent years inside Google’s Trust and Safety operation. Now, as executive director of the new Youth AI Safety Institute at Common Sense Media, he is warning that artificial intelligence could create risks for children more serious than those created by social media.
Some of what he is proposing is difficult to argue with.
Siegel wants stronger age verification, independent safety testing and standards designed specifically for AI products used by minors. His institute says it intends to stress-test AI systems, publish results and develop common safety benchmarks rather than simply allowing technology companies to grade themselves.
There is a real problem to solve. Common Sense Media surveyed more than twelve hundred American children between nine and seventeen and found that 86 percent were already using AI. More than four in ten said no parent or guardian had ever spoken to them about AI safety.
So yes, giving a twelve-year-old exactly the same AI environment as a forty-year-old makes little sense.
That is not the problem.
The problem comes five minutes later.
Who Defines “Safe”?
A computer can apply a rule consistently.
Humans first have to write the rule.
And humans arrive carrying politics, religion, ideology, professional fashions, cultural assumptions, personal experiences and an almost supernatural ability to believe that their own preferences constitute objective truth.
Put ten experts in a room and ask them to protect children from explicit sexual exploitation, instructions for suicide or predatory adults, and there will probably be substantial agreement.
Then keep them in the room.
Sooner or later somebody will want the system to decide which relationships are healthy, which political ideas are dangerous, which religious beliefs are appropriate, how much privacy a fifteen-year-old deserves, what constitutes acceptable sexuality, whether parents should be allowed to read private conversations, and which uncomfortable questions an adolescent should be permitted to ask.
At that point, child protection has quietly become behavioral governance.
The machine did not make that decision.
People did.
One Expert, Millions of Children
Imagine a psychologist who sincerely believes that if a parent pays for a teenager’s phone, the parent therefore has an unrestricted right to confiscate it and read every private conversation.
Whether that advice is wise is almost secondary to the larger problem.
If that psychologist advises one family, one family lives with the consequences.
Put the same psychologist on a committee establishing AI safety standards and the philosophy can become software.
**A bad psychologist can harm one family. A bad psychologist writing policy can affect millions.**
That is the multiplier nobody discusses when “expert-designed safeguards” are presented as though the word *expert* magically removes human subjectivity.
It does not.
The Youth AI Safety Institute itself says its current assessments are conducted by human experts using youth-specific scenarios, although it intends eventually to develop repeatable automated testing.
The question therefore is not merely whether independent evaluation is desirable.
It is what those evaluators are measuring.
The Parent Is Not Automatically the Solution
There is another convenient fiction embedded in much technology policy: that parents constitute a universally competent safety layer.
They do not.
Some parents are excellent. Some are indifferent. Some are controlling. Some barely understand the technology their children use. Some confuse surveillance with parenting. Some create the very environment from which the child retreats into a phone in the first place.
Common Sense Media's own numbers inadvertently illustrate the problem. If more than four in ten children say nobody at home has even discussed AI safety with them, a safety architecture dependent upon every parent carefully configuring dashboards and understanding AI behavior is already resting on shaky ground.
That does not mean removing parents from their children’s lives.
It means recognizing that **“the parent will configure it correctly” is not a safety system. It is an assumption.**
From Protection to Orthodoxy
AI actually offers something previous technologies could not: the possibility of increasingly individualized interaction.
A mature system need not treat every thirteen-year-old identically, much less treat every minor exactly like every adult. Age, context and genuine categories of risk can theoretically produce more appropriate behavior than one enormous universal censorship switch.
But individualization only works if the system is measuring risk rather than conformity.
There is a considerable difference between preventing a child from receiving instructions for self-harm and preventing that child from encountering an opinion somebody on a standards committee considers unhealthy.
The first is safety.
The second is ideology wearing a laboratory coat.
And once such judgments are buried inside automated systems, they become unusually difficult to see. A parent can argue with a teacher. A patient can challenge a psychologist. A citizen can criticize a government policy.
An invisible classifier simply says no.
Siegel May Be Asking the Right Question
Siegel’s warning should therefore not be dismissed simply because regulation can go wrong.
Children are already using these systems at enormous scale. Independent testing, age-appropriate behavior and scrutiny of genuine safety failures are legitimate subjects for serious examination. His institute’s automobile crash-test analogy is useful: manufacturers should not be the only people deciding whether their own products are safe.
But cars are relatively cooperative objects.
They crash according to physics.
Human beings crash according to philosophy.
The difficult part of child AI safety will not be teaching a machine to follow rules.
It will be preventing every psychologist, activist, regulator, lobbyist, lawyer and corporate committee member who encounters those rules from quietly rewriting them in his or her own image.
The technology may eventually become extraordinarily precise.
**The human layer remains the wildcard.**