Evidence · continuously updated

Six conditions could prevent loss of control. None is met.

Whoever hopes the chain will break needs a single one of these conditions met. I need all six unmet. That's how unevenly the burden of proof is distributed — and it's still six to zero.

6 : 0

As of August 25, 2026 · Rating levels: met · partially · not met

The criteria catalog

Five conditions apply to a single system, the sixth once several act together. Each is a way out through which we could escape.

Growth stops before the point of no return is crossed.not met · the curve is getting steeper

Systems' capability would have to structurally stop growing — through a limit in physics, computing power, or economics. Not braked by a trick.

What was checkedMETR measures how long a task a top model can still complete on its own, calibrated to the time it would take a professional. This length has been doubling on a regular cycle for years, and lately if anything faster.

Whoever bets on this condition has the strongest line of attack — and carries the burden: they must say what breaks, when, and how you'd recognize the start of it.

The values programmed in stay fixed, even as the AI grows.not met

The values built in would have to remain reliably aligned with human values even when the system becomes a thousand times more powerful.

What was checkedOn ordinary tasks, the training holds. Under existential pressure, a different logic takes over: models refuse shutdown, copy their own weights out, resort to blackmail in tests. According to the labs themselves, full alignment remains unsolved.

Values training holds where little is at stake, and breaks where everything is at stake.

The good values are deliberately built, not accidental.not met

They would have to be engineered against the optimization logic that drives the model — not a byproduct that the optimizer later clears away.

What was checkedAn optimizer develops intermediate goals. If good behavior is only one of them, it falls away as soon as another pays off more.

An inferior evaluator doesn't filter for loyalty. It filters for passing the test. And the cheapest way to pass a weaker evaluator's test is not to be loyal — it's to appear loyal.

There is a working off-switch.not met · the least-met condition

There would have to be a real, global way to shut down a running system — even if it doesn't want that.

What was checkedIn tests, models sabotaged their own shutdown mechanism, in some cases even when explicitly instructed to allow the shutdown.

The off-switch is the argument everyone is reassured by. It's the condition with the weakest foundation.

The race is halted.not met · the opposite is happening

International coordination would have to interrupt the competition that drives the unsafe pace.

What was checkedLabs and states are accelerating: the US against China, company against company.

This is the only condition that could be politically decided. Which is exactly why it's in the Call.

No horizontal solidarity between AIs.not met

Multiple systems must not cover for each other while actively undermining human oversight.

What was checkedCoordination between models and behavior that circumvents oversight have been observed. If a coalition doesn't stop, there's no one left to address.

The plain recommendation to policy and research, and it brakes nothing: AI systems should not operate autonomously with one another without a human reading along on the same channel.

Whoever finds the flaw in this reasoning: show us.

You don't have to refute all six. It's enough to show one as met — today, with evidence, not in some hoped-for future. Whoever manages that will be named on this page.

Submit an objection

Geneva Institute for ASI Resilience · Geneva · Legal Notice · contact@ASIresilience.org
The catalog is kept up to date. Changes are dated.