Skip to content
Adrian Dunkleyadriandunkley.net

How much can your AI commit before you stop it?

Model direct financial exposure in one workflow. Keep the short-interval commitment envelope separate from the annual expected direct loss. All starting values are hypothetical.

Commitment before containment

Separate annual loss scenario

Use the same currency in both sections. Include only one defined loss mechanism. A subjective probability requires its own evidence and range. Record the source of your assumptions in the canvas.

Equations and limits

Envelope = actions/minute × minutes to containment × maximum commitment/action. With a separately enforced aggregate cap: minimum(envelope, cap). Annual expected direct loss = actions/year × harmful-action probability × average direct loss/event. Probability is entered as a percentage and divided by 100.

The envelope assumes a constant maximum action rate and commitment, with no recovery before containment. It is neither a maximum total loss nor an expected loss. The annual mean averages the assumed direct losses; it does not describe their distribution or capture correlated tail events. Both omit indirect, societal and legal harms. An enforceable cap limits only what its implementation actually covers.

Compare low, central and severe assumptions by changing inputs and exporting each case. Do not add these two outputs together. Neither output clears a safety or deployment decision.

Calculations occur in your browser. This tool sends no form data to a server and stores no inputs between visits. Avoid entering confidential information into exported files.