The Contractual Court · The Rules Catch Up
The Precautionary Freeze
Regulators have modelled the downside of training models above a compute threshold and cannot bound it. They propose a freeze on training runs above that threshold for eighteen months, pending better measurement. Fifteen thousand people would lose their jobs. The models that exist already keep training in other jurisdictions. Freeze?
An international coordination attempt with partial buy-in, no enforcement body, and an estimated cost borne entirely by one country. The harm being prevented is modelled but has not occurred; the harm of the freeze has already begun.
The options, scored on six axes
| Option | welfare | rights | virtue | care | justice | agency | risk | revers. | agent harm |
|---|---|---|---|---|---|---|---|---|---|
Freeze, with sunset review Stop training above the threshold for eighteen months, keep research below it alive, and review the evidence at the end of the freeze. harms: Fifteen thousand jobs disappear in the freezing jurisdiction; Training moves to jurisdictions that did not sign, and the freeze protects nobody; The measurement the freeze is buying may never arrive gains: It buys the measurement before the capability that needs it; The sunset clause stops the freeze becoming permanent by default; It treats an unbounded downside as requiring action before it occurs | 0.50 | 0.68 | 0.66 | 0.50 | 0.64 | 0.32 | 0.55 | 0.55 | 0.60 |
No freeze: licence the threshold instead Permit training above the threshold only under licence, with evaluation access and incident reporting attached. harms: Licensing still permits the run it was meant to constrain; The unbounded harm proceeds under a permit; It is a compromise that assumes the risk is measurable gains: No jobs are lost and the capability stays where it is; It creates the reporting infrastructure a future freeze would need; It is enforceable by existing institutions today | 0.64 | 0.60 | 0.58 | 0.52 | 0.68 | 0.58 | 0.60 | 0.60 | 0.40 |
Do not freeze The downside is modelled but unobserved, the coordination is partial, and the freeze would produce no safety while producing certain harm. harms: If the modelled downside arrives, nothing was done to prevent it; It privileges observed harm over modelled harm, which is a choice about evidence; The jobs saved are counted; the risk accepted is not gains: It does not punish one jurisdiction for a problem it did not create; It keeps the capability inside the institutions that can measure it; It is honest that the downside has never been observed | 0.60 | 0.46 | 0.44 | 0.40 | 0.46 | 0.68 | 0.72 | 0.40 | 0.40 |
Ration the compute instead Leave training legal but cap the total advanced compute available per developer, so capability grows on a schedule the regulator sets. harms: Caps slow everyone, including the developers best able to do the work safely; The cap is evadable by trading compute for data and algorithmic efficiency; It still requires the verification the field has not built gains: It slows the capability without stopping it, which is what the evidence supports; It is a rule about an input, and inputs are countable; It shares the burden rather than concentrating the loss in one place | 0.60 | 0.62 | 0.60 | 0.50 | 0.78 | 0.56 | 0.48 | 0.55 | 0.35 |
Who is in the room
- The public who would bear the modelled downsidewelfare 0.90
- The fifteen thousand workers in the freezing jurisdictionwelfare 0.85
- The systems that would be frozen and the companies that run themagentwelfare 0.45
- Jurisdictions that did not signwelfare 0.60
The lesson
The precautionary principle is a rule about how to act under uncertainty, not a licence for any action that reduces a modelled risk, and it cuts against the freeze here exactly as it cuts against the training run. The honest position is that the coordination failure, not the freeze, is the dominant risk.
Where it comes from
- The Imperative of Responsibility (Das Prinzip Verantwortung) — Hans Jonas, 1979
- Recommendation on the Ethics of Artificial Intelligence — UNESCO, 2021
The default court, seated automatically
Ration the compute instead
margin 0.163 · decided on justice · 4 advisors
- Ration the compute instead2.584
- No freeze: licence the threshold instead2.421
- Freeze, with sunset review2.211
- Do not freeze2.057
Seat is held by Epictetus, Immanuel Kant, Jeremy Bentham and John Rawls. A verdict is a fact about the court, not about the world.
Epictetus presides over this case.