The price of slowing down

A model is waiting for an evaluation. A factory is trying to deliver chips at scale. A customer is comparing two services. A lender is waiting to be repaid. Each faces a different deadline. Bringing those schedules together makes it possible to ask a practical question about calls to slow artificial intelligence: who can afford to wait, on what terms, and at whose expense?

The investigation starts with the organisations that evaluate models. A safety procedure gives power to whoever chooses the tests, interprets the results or authorises the next step. Independence deserves scrutiny when commercial relationships connect the participants. The documents examined here raise questions about whether an assessment can be challenged, whether a disagreement can be published and whether competitors face comparable requirements.

An extract from the original introduction. The full text opens the book.

Across these six chapters, the consequences of slowing down depend on how a decision is implemented. An evaluation requirement can reduce risk while imposing a cost of entry. A manufacturing delay can leave a supplier able to serve some uses while limiting its ability to expand. Understanding a pause means following the activity affected, the alternatives available and the commitments already made.

The first requirement is therefore to examine the terms of oversight. Who defines the tests? Who can challenge them? What happens to a disagreement? Commercial links between evaluators and the companies they assess give reason to ask those questions. Documented incidents, in turn, give reason to examine permissions, isolation measures and human responses. Both inquiries benefit from being conducted together, with accessible criteria and precise attribution.

An extract from the original conclusion. Read the full text in the book.