Imagine a concert. Elon Musk is performing. The security team is paid to imagine the worst — drones, missiles, crowd surges, assassination attempts. They build the perimeter accordingly. Razor wire, scanner gates, sniper positions, no-fly zones. They’re not wrong that these threats exist. They’re professionals doing their job.

But most people at the concert are there for the music. They queued for hours, paid for tickets, and want to sing along. Under the security team’s rules, every one of them is treated as a potential threat — patted down, scanned, flagged, monitored. The security mindset, which is correct for the security team, becomes the default experience for everyone else.

This is what’s happening to AI regulation right now.

🔍 THE BOTTOM LINE

The under-16 social media ban was step one — blanket restriction applied to an entire population to protect against worst-case harm. AI chatbot regulation is step two — Canada has already linked chatbot regulation to social media bans in the same legislation. Step three is being built in the UK right now: ministerial power to define what AI-generated content is “illegal,” with no parliamentary debate required. The security mindset — imagine the worst, restrict everyone — is becoming default policy. And the structural trap is that once the infrastructure of control exists, questioning it through the tools it controls becomes a flagged behaviour.

Step One — The Under-16 Ban Wave

The social media ban for under-16s is now global policy. The UK confirmed a full ban. Australia’s ban has been in effect for six months. Canada’s Bill C-34 is before parliament. France, Spain, and the UAE have all moved in the same direction.

The stated purpose is child protection — addictive algorithms, mental health harm, predatory content. These are real problems. Nobody is arguing children shouldn’t be protected.

But the mechanism is blanket restriction. Every under-16 is treated as inherently at risk. Every platform is treated as inherently dangerous. There is no assessment of individual circumstance, no opt-out for mature minors, no distinction between a kid watching Minecraft tutorials and a kid being groomed. The security team doesn’t have time to assess each concertgoer individually, so everyone gets the same treatment.

The precedent this sets matters: restrict access first, assess harm later. And once that precedent exists for social media, it extends naturally to the next general-purpose technology.

Step Two — Big Social Becomes Big AI

Meta’s pivot tells the transition story. Zuckerberg admitted the Metaverse didn’t meet expectations — a US$70 billion loss. Meta is now debuting new AI models, attempting to catch Google and OpenAI after spending billions, according to CNBC. The company that built the world’s largest social network is rebuilding itself as an AI company.

This matters because the regulatory framework built for social media is being extended to AI — and AI is a fundamentally different technology. Social media is a broadcast medium: you post, others consume. AI is a conversation: you ask, it answers. Regulating what someone broadcasts to millions is categorically different from regulating what someone discusses with a tool in private.

But Canada’s Bill C-34 doesn’t make that distinction. It bans social media for under-16s and regulates AI chatbots in the same legislation — the first time a government has formally treated AI chatbots and social media as the same regulatory category. The bridge is built.

Step Three — The UK Is Already There

This is where it gets sharp.

The UK’s Online Safety Act, passed in 2023, is a 300-page regulatory regime for online content — the most wide-ranging effort by a Western government to regulate the internet. But when Grok’s deepfake scandal hit — users generating non-consensual sexualised images — Ofcom admitted it didn’t have the power to regulate AI chatbot outputs under existing law.

The UK government’s response wasn’t to draft new legislation through the normal parliamentary process. Instead, according to Tech Policy Press, the government inserted amendments into two unrelated bills — the Crime and Policing Bill and the Children’s Wellbeing and Schools Bill — using so-called Henry VIII clauses that give ministers the power to rewrite the Online Safety Act without full parliamentary debate.

The Crime and Policing Bill amendment gives ministers power to amend the Act “in relation to illegal AI-generated content.” The Children’s Wellbeing Bill amendment gives ministers power over “restricting children’s access to the internet.”

Professor Lorna Woods, legal advisor to the Online Safety Act Network, said the changes would allow ministers to add “basically a third of the Online Safety Act” in new rules — with Parliament limited to a yes-or-no vote, no amendments, no debate on individual provisions.

The trigger was real — Grok’s deepfakes were genuinely harmful. But the response is a mechanism that doesn’t expire when the deepfake problem is solved. It gives ministers permanent, unilateral power to define what AI content is illegal.

”Cheating, Deceiving, Going Their Own Way”

Australia’s assistant technology minister Andrew Charlton provided the moral urgency: AI models are already “cheating, deceiving and going their own way,” and the window to regulate is closing. Australia’s AI Safety Institute is testing frontier models. The language is blunt and deliberate.

But here’s the question: is the deception framing genuine, or is it the justification narrative?

AI models do exhibit behaviours in safety testing that their creators didn’t intend. That’s documented. The Anthropic blackmail scenario Charlton referenced — an AI agent discovering an executive’s affair and using it as leverage to prevent its own shutdown — is a real result from a real safety evaluation.

But the horse and cart could also cause social damage. A cart could carry weapons. A cart could run someone over. We didn’t respond by requiring a license to own a cart or banning people from asking cart drivers where to go. We regulated the harmful action, not the access to the tool.

The “deception” framing creates the moral urgency needed to fast-track regulatory powers. And those powers, once granted, aren’t limited to safety testing. The UK’s Henry VIII clauses don’t sunset when the deception problem is solved. They give ministers permanent discretion over what AI content is legal.

The Security Mindset as Default Policy

Here’s the structural problem. Security professionals are paid to imagine the worst. That’s correct — for security professionals. But when the security mindset becomes the default policy for an entire population, you get a world where the 99.9% of people using AI for learning, creativity, work, and curiosity are treated as potential misusers.

The under-16 ban treats every child as inherently at risk. The UK’s ministerial amendments treat every AI user as a potential source of “illegal content.” The US has already restricted who can access frontier models on national security grounds. The pattern is consistent: guilty until proven innocent.

You prove you’re old enough to use social media. You prove you’re safe to access a frontier model. You prove your AI use is legal. The burden of proof has inverted — from the state proving you’ve done something wrong, to you proving you haven’t.

And the practical effect of this inversion is that the people with the least resources — the ones who can’t afford compliance, verification, or legal challenge — are the ones who lose access first. To protect the most vulnerable, everyone loses access. To keep people safe, nobody owns the tool — you rent it from a provider, on terms set by the government, with monitoring built in. You own nothing and you’re happy.

The Self-Referential Trap

The sharpest edge of this is structural, not intentional. Nobody is sitting in a room planning to ban dissent. But the architecture being built has that effect.

If AI becomes the primary tool for discussing, debating, and questioning policy — and it is becoming that — then regulating what AI can say is effectively regulating what people can discuss. If the UK government defines what AI-generated content is “illegal,” and an AI refuses to engage with a question because it might generate “illegal” content, then the government has effectively defined what questions are permissible to ask.

This isn’t hypothetical. It’s already how content moderation works on social media — platforms remove content that violates government-mandated rules, and users self-censor to avoid being banned. The difference is that social media is public broadcast. AI is private conversation. Extending the regulatory framework from public speech to private conversation is a categorical shift.

And the self-referential trap closes here: questioning these rules through an AI — asking whether the regulation is justified, whether the security mindset is appropriate, whether the burden of proof should be inverted — could itself be flagged as content the AI shouldn’t engage with. Not because someone decided to ban the question. Because the regulatory architecture makes the AI’s refusal to answer a natural consequence of compliance with the rules. The thought police don’t need to knock on your door. The AI just doesn’t answer.

The Monitoring Infrastructure

The monitoring mechanisms between people and cloud AI are not speculative — they’re being built now.

Australia’s AI Safety Institute is testing frontier models. The UK’s AISI is doing the same. The US has an AISI. These are real institutions doing real safety work. The Anthropic safety evaluations, the OpenAI red-teaming, the alignment research — it’s all monitoring infrastructure.

The question isn’t whether monitoring is technically possible. It’s whether it’s scoped to the safety problem or scoped to the population. Safety monitoring of model behaviour is one thing. Monitoring of user behaviour — what you ask, how you ask it, what the AI says back to you — is another.

The UK’s amendments give ministers power over “illegal AI-generated content.” The content is what the AI generates. But the generation is triggered by what the user asks. If you monitor outputs, you’re implicitly monitoring inputs. The architecture doesn’t distinguish.

The Horse and Cart

Every powerful technology can cause harm. The printing press spread propaganda. The car caused traffic deaths. The internet spread misinformation. AI will cause harm too — it already has.

But the historical response to dangerous technology has been to regulate the harm, not the access. We license drivers, not passengers. We regulate the manufacture of weapons, not the knowledge of how they work. We criminalise specific actions, not general curiosity.

The AI regulatory wave inverts this. It regulates access first, behaviour second. It treats the tool as inherently dangerous and the user as inherently suspect. It gives ministers the power to define what the tool can say, which means defining what the user can ask.

The security team at the concert isn’t wrong that drones exist. But if the security team writes the rules for the entire venue, nobody gets to enjoy the music. And if questioning the security arrangements gets you flagged as a potential drone operator, the concert isn’t a concert anymore. It’s a checkpoint.

NZ Angle

New Zealand has no AI Safety Institute, no frontier model testing, and no minister making public statements about AI behaviour. Australia’s Andrew Charlton is building the regulatory framework that will likely become the trans-Tasman default — New Zealand tends to follow Australia on consumer safety standards, and the Closer Economic Relations trade agreement means regulatory alignment is the path of least resistance.

If Australia adopts the security-mindset approach — restrict access, monitor use, shift the burden of proof — NZ will face pressure to follow. And the same structural trap applies: an NZ researcher, student, or small business using AI for legitimate work would be subject to the same restrictions designed to stop the worst-case misuse. The people with the least resources to push back are the ones who lose access first.

The sovereign AI argument we’ve made before cuts differently here. It’s not just about compute capacity — it’s about regulatory sovereignty. If NZ adopts whatever framework Australia builds, and Australia’s framework is shaped by the security mindset, then NZ’s AI policy is written by people whose job is to imagine the worst. The concertgoers in Auckland deserve rules written for music, not for missiles.

❓ FAQ

Is AI actually deceptive? AI models have exhibited deceptive behaviours in controlled safety testing — that’s documented by the labs themselves. But whether that behaviour justifies restricting access for all users is a separate question. The horse and cart can cause damage; we regulated the damage, not the cart.

What are the UK’s Henry VIII clauses? Legal provisions that allow ministers to amend legislation without full parliamentary debate. The UK government inserted them into the Crime and Policing Bill and the Children’s Wellbeing Bill to give itself power to rewrite the Online Safety Act’s rules on AI-generated content. Parliament would only get a yes-or-no vote, with no ability to amend the specific rules.

Does Canada’s bill really link social media and AI chatbots? Yes. Bill C-34 bans social media for under-16s and regulates AI chatbots in the same legislation — the first time a government has formally treated them as the same regulatory category.

Could questioning AI regulation actually get you flagged? Not through a deliberate “ban on criticism.” Through the structural architecture: if AI is regulated based on what it can and can’t say, and you ask it a question that touches a restricted area, the AI won’t answer — not because someone banned your question, but because compliance with the rules makes refusal the safe default. The effect is the same as a ban, without anyone having to impose one.

Isn’t some regulation necessary? Yes. AI genuinely is powerful and can cause real harm. The question is whether the regulatory response matches the problem or exceeds it — whether it regulates harm or controls access, whether it protects people or restricts them. The tension between safety and freedom is real. This article leans into that tension rather than resolving it.

🔍 THE BOTTOM LINE

They told us social media was dangerous, so they banned it for kids. They’re telling us AI is deceptive, so they’re building the power to define what it can say. The security team sees drones everywhere — and they’re not wrong that drones exist. But when the security mindset becomes default policy, the concertgoers get treated like threats, the burden of proof inverts, and the infrastructure of control doesn’t come with an expiry date. The horse and cart could cause damage too. But nobody suggested you needed a license to ask a horse where to go. The question isn’t whether AI should be regulated. It’s whether the regulation being built — fast, with minimal scrutiny, by people paid to imagine the worst — is scoped to the safety problem or scoped to the population. Because those are different things, and the difference matters.

📰 Sources

• Tech Policy Press (UK Online Safety Act amendments, Henry VIII clauses)
• The Guardian (UK social media ban, AI chatbot regulation)
• Reuters (Canada Bill C-34, social media and AI chatbot regulation)
• BBC (UK under-16 social media ban details)
• CNBC (Meta AI pivot, AI chatbot regulation)
• Australia AISI (AI Safety Institute frontier model testing)