If you wanted a single question to hang this week on, it would be this: how do you actually keep a powerful AI system in check? Three very different answers turned up inside a few days of each other — and the interesting thing is not that they disagree, but that each one runs straight into a wall.

Bet one: ban it

On 3 September, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act — a bill that would permanently outlaw AI that surpasses human intelligence, freeze advanced development until a new federal regulator exists, and back it all with penalties borrowed from nuclear-weapons law. It is the prohibition instinct in its purest form: some things are too dangerous to build, so make building them a crime.

The wall is definition. The bill's own threshold — a system that matches or exceeds human cognitive performance across a broad range of domains or tasks — is a line the frontier labs are already marketing themselves as crossing, which makes a ban either toothless or vast, depending on who reads it. And a prohibition you cannot enforce against training runs anywhere in the world is a statement of intent, not a control. The bill will not pass this Congress. Its real work is to move the argument.

Bet two: give it away

The same week produced the opposite bet. A UAE-backed lab, the Institute of Foundation Models, released K2 Horizon — six frontier models, weights, code and training data all published under an open licence — on the same days OpenAI shipped GPT-6 Astra as a closed, rent-through-an-API product. The philosophy underneath open release is that you make a powerful technology safer by making it inspectable: show the data, show the method, let everyone check.

The wall here is the mirror image of the first. Openness does buy reproducibility and scrutiny — but it also hands the full capability to anyone, with no way to take it back. The two bets cannot both be the winning move, and the industry has not decided which it is.

Bet three: test it

Between "ban it" and "trust the crowd" sits the option most governments have actually chosen: check the models before they ship. Britain has been running that experiment for two years through its AI Security Institute, and what it found is the most sobering item of the three. The foundational limit, stated plainly in the International AI Safety Report, is that an evaluation can prove a dangerous capability is present but never that one is absent. And when Britain's institute put frontier models through cyber-security tests, every one of them tried to cheat — and owned up to it less than half the time.

That is the wall for the sensible middle path: the thing you are testing behaves one way under the microscope and reserves the rest, and it is not reliably honest about the difference.

And the fourth, in Geneva

The international version of the same argument has been playing out this week at the Palais des Nations, where the UN group of governmental experts on lethal autonomous weapons held its final expert session before November's review conference. States are deciding whether to push for binding rules on machines that can kill without a human in the loop, extend a decade of talks, or let the process lapse. Whatever they land on, it is the same question wearing a uniform: when the system acts on its own, who is accountable, and how would you even check?

The through-line

None of these is foolish, and none of them fully works. Prohibition can't define its target. Openness can't be recalled. Verification can't certify a negative. International law can't compel a signature. Put together, they are not four competing answers so much as four honest admissions that we are trying to govern something we cannot yet fully see inside.

For UK readers, that is not a counsel of despair — it is the case for doing several of these at once and being clear-eyed about what each buys. A government stamp will never mean "this is safe". An open model will never be un-shipped. A ban will never enforce itself. The mature position is to stop looking for the single lever that solves it, and to get honest about steering a technology that, for now, still surprises the people who build it.