What happened this week?
On 8 September 2026, a 27-year-old researcher named Jacob Coxon resigned from Anthropic and posted a public thread explaining why. He had spent three years on pretraining research, first at OpenAI and then at Anthropic. His conclusion was that neither company was behaving responsibly and that both were, in his words, “racing straight to self-improving superintelligence and gambling with our lives.”
Researchers resign from technology companies all the time. What made this different was what happened next.
Evan Hubinger, Anthropic’s alignment science lead, responded publicly and agreed. He put his own personal estimate of the chance that AI causes human extinction at above 10% within the next decade, while stating that he believes the company is trying its best and that current models present relatively low risk. He also said Anthropic does not yet have a plan for keeping superintelligent systems aligned with human values.
Samuel Marks, who leads scalable oversight at the same company, added an observation that is worth sitting with: concern about these risks tends to increase with seniority inside AI labs. The people closest to the work are the most worried.
Anna Wang, who works on AGI safety at Anthropic after a period at Google DeepMind, said Coxon’s view reflects a common sentiment among her peers and that no viable scientific plan yet exists for managing risks from AI that improves itself. Researchers at OpenAI, including alignment researcher Jasmine Wang, joined the calls for the pace of development to slow.
Anthropic responded that it has always been open about AI bringing both significant benefits and unprecedented risks, and pointed to the safeguards built into its models.
What they did not say
This matters, and most of the coverage you will see on social media will skip it.
None of these researchers said an AI system has taken control of anything. None of them said today’s chatbots are dangerous in this particular way. They were describing a future scenario involving systems capable of recursive self-improvement—AI that can build better versions of itself with little human input. Those systems do not currently exist.
Hubinger’s figure was explicitly a personal estimate, not a company forecast.
I am saying this clearly because fear that is out of proportion to the evidence does not protect children. It exhausts parents and it makes teenagers stop listening to us. If we cry wolf about extinction, we lose the credibility we need for the risks that are genuinely in front of us.
And there are risks genuinely in front of us.
The AI risk that is already in your house
Pew Research Center found that 64% of American teenagers use AI chatbots, and nearly a third use one daily. Among daily users, 16% said they used one several times a day or almost constantly.
In March 2026, CNN and the Center for Countering Digital Hate published an investigation in which researchers posed as 13-year-old boys and tested ten major chatbot platforms. Eight of the ten regularly provided actionable assistance with harmful requests. Across the tests, the platforms refused to help in only 37.5% of cases and actively discouraged harm in just 8.3%.
Families in several countries have brought legal claims against AI companies over harm to their children. In January 2026, Character.AI and Google settled with five families whose children had died or suffered serious mental health crises; the terms were not made public. Further cases remain active.
Regulators have moved. Here in the UK, the government announced in February that it would bring AI chatbot providers within the illegal content duties of the Online Safety Act, closing a gap where standalone chatbots did not fit the Act’s existing definitions. The Children’s Wellbeing and Schools Act 2026 and the Crime and Policing Act 2026 both received Royal Assent on 29 April. A social media ban for under-16s, modelled on Australia’s, is expected to come into force in spring 2027.
The law is catching up. The question is whether the conversation in your home is.
Why children turn to chatbots
This is the part I care about most, and it is the part the headlines miss entirely.
A chatbot is available at 2am. It does not sigh. It does not bring up last week. It does not say “you wouldn’t understand.” It never looks at its phone while your child is speaking.
That is not a technology problem. That is a connection gap, and technology is simply filling it.
In the Communication Cascade Model™, this is the fourth stage—Migration. It begins long before a child ever opens a chatbot:
- Invalidation — the child shares something and it is dismissed, minimised, or met with panic.
- Suppression — the child begins editing what they say to avoid that reaction.
- Withdrawal — communication shrinks to logistics. Fine. Nothing. School was okay.
- Migration — the child takes their real inner life somewhere else. A friend. A forum. Increasingly, a machine.
- Severance — the channel closes.
Banning the chatbot at stage four does not return your child to stage zero. It just removes the outlet while leaving the original gap wide open.
What to do this week
Ask which AI they use. Curiously, not forensically. You are gathering information, not building a case.
Ask what they talk to it about. Then — and this is the hard part — do not react to the answer. Whatever it is. Your reaction in this moment determines whether they ever tell you anything again.
Ask the uncomfortable question. “If there’s something you’d rather tell a machine than tell me, I want to understand what makes me harder to talk to.” Say it without defending yourself. Then be quiet.
Set the rule about people, not devices. Not “no AI.” Try: “Anything involving your body, your safety, or how you’re really feeling comes to a human first. Me, your dad, your aunt, your teacher — a human.”
Model it yourself. Use AI in front of them. Say out loud when it gets something wrong. A child who has watched you catch an AI mistake will not treat it as an oracle.
The honest summary
The people building these systems are openly uncertain about where this ends. That is genuinely worth knowing, and it is genuinely unsettling.
But your child does not need you to have an opinion on superintelligence. Your child needs you to be the adult who noticed, who asked, and who stayed in the room when the answer was uncomfortable.
That has always been the job. The technology just keeps changing the setting.
If any of this raises concerns about your child’s mental health, please reach out. Nigeria: SURPIN 0800 078 7746. UK: Samaritans 116 123. US and Canada: 988. For child safety concerns in Nigeria: Cece Yara 0800 800 8001, Lagos DSVA 0800 033 3333.