When Judgment Becomes an API Call

AI systems can now make structured judgments that can be embedded directly into software. As machine judgment becomes part of ordinary processes, we need to decide how much authority to give it.

From Generation to Judgment

In September 2026, TypeSafe AI introduced Jev, its first public “System One Model.”

Most of our experience with generative AI is conversational. We ask a question, the model produces an answer, and we decide what to do with it.

System One Models are designed differently. Rather than generating an open-ended response, they make structured judgments about unstructured information. Jev can return a probability, select between predefined choices or assign a score.

A company could give it a job application and ask:

Does this candidate meet the requirements?
Should this application receive human review?
Which role is the closest match?

Jev is designed to be self-consistent and its outputs can be consumed directly by software.

A probability that a candidate is suitable could be shown to a recruiter. It could determine whether the application receives human review. Or it could automatically reject the application.

The model is the same. What changes is the authority we give its judgment.

When I first read about Jev, I immediately thought about a book published almost exactly ten years earlier.

Weapons of Math Destruction

In Weapons of Math Destruction, Cathy O’Neil examined mathematical models used to make consequential decisions about employment, loans, education, insurance and policing.

She identified three characteristics that can turn such models into what she called Weapons of Math Destruction: opacity, scale and damage. The greater the authority we give such a model, the more these characteristics matter.

One hiring manager may reject a candidate because of their own preferences or prejudices. Another may see the same person differently. Someone rejected by one employer gets another chance elsewhere.

A model can apply the same assumptions across thousands or millions of decisions. When judgments are applied consistently at scale, biases that might once have been individual can become systematic.

Yet the resulting decision can appear objective because it comes from a mathematical model. And when the model is opaque, someone affected by its decision may have little ability to understand or challenge it.

How Much Authority?

This question extends beyond decisions about people.

In an engineering organisation, a model could advise an engineer that an issue should be escalated. Or it could decide which issues engineers ever see. In the first case, it supports human judgment. In the second, it becomes part of the organisation’s decision-making structure.

There may be no visible AI assistant making these decisions. There may simply be a workflow.

An organisation could therefore gradually delegate thousands of small judgments without explicitly deciding where machine authority should end.

This is where we need to make deliberate choices about what the model is allowed to judge, what happens to its output, when a human becomes involved and how its decisions can be challenged.

Weapons of Math Destruction showed what can happen when mathematical models acquire authority over consequential decisions while their assumptions remain difficult to see. System One Models make that question newly relevant because machine judgment can become an almost invisible component of ordinary software.

The model provides the judgment. We decide how much authority it carries.

References

AI helped to research and write this article.