The Gate Has a Hole in It

Here is what a fifteen-year-old with a gaming PC can do on a Saturday afternoon.

He downloads a free program. He types one line to fetch a free model, a few gigabytes, released by a company in Paris, Beijing or Menlo Park under a licence that asks nothing of him. He installs a free chat window on top of it. Forty minutes later he has an AI assistant running on his own graphics card, in his own bedroom, with no account, no login, no terms of service, no age check and no one on the other end.

Nobody can stop him. No law was broken. No platform was involved. There is nothing to fine, nothing to block and nobody to serve a notice on, because the thing he built is not a service. It is software, running on hardware he owns, the same as a game or a spreadsheet.

Australia has spent two years building age walls around AI. It has not noticed that the walls are penetrable.

What the gate actually gates

The under-16 social media ban, the eSafety Commissioner's March 2026 codes for chatbots and companions, and the Children's Online Privacy Code due in December all share one design. They regulate the provider. The company that runs the service must verify, filter, restrict and report, or face penalties that now run to $49.5 million per breach.

That design works on the companies it can see. OpenAI, Google, Microsoft and Anthropic have offices, lawyers and share prices. They have responded exactly as the incentives dictate: age gates at thirteen with parental consent, or eighteen flat, depending on how much risk each is prepared to carry. The most capable research assistant on the market, the one with the heaviest investment in safety training, now refuses anyone under eighteen, parent or no parent. That is the rational response to a $49.5 million stick.

The design does nothing to the thing it cannot see. An open-weights model is a file. Once it is on a hard drive, no regulator, no platform and no app store has any further say in what it does. There is no provider. There is no account. There is no compliance officer to lean on. The gate, by its nature, only stands in front of the companies that were already trying.

The safety is on the wrong side of the wall

Here is the part the policy never confronts.

The commercial models are the safe ones. Hundreds of millions of dollars have gone into training them to refuse harmful requests, to hand a distressed user a helpline, to decline sexual content with a minor, to flag self-harm. They log sessions. They offer parental controls. They are, by a wide margin, the most carefully supervised conversational software ever built.

The open model on the teenager's hard drive has none of that, or has it only as far as a volunteer thought to include it. Many of the most popular downloads are explicitly "uncensored" builds with the refusals stripped out, because that is what the hobbyist market asks for. There is no log unless he keeps one. There is no helpline. There is no parental control panel because there is no parent in the loop at all, unless the parent is standing behind the chair.

So consider what the age wall actually does to a capable, curious fifteen-year-old. It does not keep him from AI. It removes the supervised, safety-trained, logged option, and leaves him the unsupervised, unfiltered, unlogged one. The regulation does not reduce the risk it was written to address. It relocates the child to where the risk is highest, and then congratulates itself that he is no longer on the platform.

The hazard the codes describe is real. It is now simply happening somewhere the codes cannot reach, and the codes are the reason it moved.

What he does with it

None of this is an argument that the teenager should be stopped. He shouldn't, and he can't be. It is an argument about what we have pushed him toward and what we have taken away.

Because the local model, for all its gaps, is a serious tool. On a mid-range graphics card a free model will explain why an engine fires in the order it does, walk through a block of code that won't compile, lay out a phase diagram, check the arithmetic on a 3D-print tolerance, and translate a Latin inscription. It will do it at midnight, with no quota, for no money, and it will never report the conversation to anyone. For a withdrawn child who will not ask a teacher but will ask a machine, that is a tutor. For a regional kid three hours from the nearest specialist, that is the specialist.

And the act of building it is itself an education. To stand up an inference server a fifteen-year-old has to understand what a model is, what quantisation does, why VRAM matters, how a local web service works and how to secure it. That is a Year 12 computing syllabus compressed into a weekend, undertaken voluntarily, for the reward of a working machine.

The child who does this is not the problem the codes were written for. He is the outcome the education system says it wants and cannot produce.

The policy's only contribution has been to make sure he did it without guardrails.

The counter-argument, and what it misses

The regulator's defence is straightforward and not stupid. Most teenagers will not build a local model. The age wall still protects the many who would otherwise have drifted into a companion app, and if a few technically able kids route around it, that is an acceptable loss for a rule that is simple, enforceable and understood.

Three things are wrong with that.

First, "a few" is already not a few. Running a local model has gone from a hobbyist project to a one-click install in under three years, and the barrier drops with every release. The teenagers who route around the wall are not a rounding error; they are the leading edge of where everyone will be.

Second, the kids who route around it are precisely the ones the policy should most want inside the fence. They are the curious, the capable, the ones who will use the tool most and benefit most from the safety work the commercial providers have done. A rule that keeps the passive majority off a companion app while pushing the active minority onto an uncensored model has got its priorities backwards.

Third, there was a better rule available. A parental-consent pathway, with a parent-set content ceiling on graded axes rather than a single age switch, would have kept the supervised option open for the family that wants it while leaving the blunt wall in place for everyone else. That rule does not need to be invented. An open standard built on exactly that model already exists: the VEIL Standard's Guardian tier lets a verified parent issue a signed consent envelope for a young person, locks the content ceilings to the child's age profile, makes the session log visible to the parent, and binds any engine honouring it to six non-negotiable minor-safety obligations. It is published, free to implement, and readable by anyone at r8rly.org/veil.

The regulator chose the simpler rule. The simpler rule has a hole in it, and the hole is shaped like the most capable children in the country.

The wall is not the point

Every parent of a teenager already knows what a locked door does. It does not stop the child. It stops the conversation.

That is what the age wall has done. The supervised product, with its logs and its parental controls and its helplines, was the one place a parent could stand beside a child and see what was being asked. The regulation has closed that door and left the window open. The child has gone through the window, as children do, and now nobody's standing beside him.

If the aim was to keep young Australians away from AI, the policy has failed, and was always going to. If the aim was to keep them safe while they used it, the policy has made things worse. The only aim it has actually served is to let the adults who wrote it say that something was done.

Something was done. It was the wrong thing.

Companion piece: The Permission Slip Generation.