What was announced
On 11 November 2026 Google DeepMind introduced the Gemini Decisions API, powered by a model called Gemini Decision 1. We read Google’s announcement and its documentation. The model does not write text. A developer sends one input, which may be text, images, audio, video or a PDF, together with up to a thousand questions. Each question is a yes/no check, a choice from a fixed list, or a rating against an ordered rubric. The model returns a probability and a confidence value for each, and the developer’s own code branches on the numbers. Google’s examples include routing queries to a model tier, picking an emoji, and a safety layer for tools in Google Search. The documentation says the confidence value is for deciding whether to act automatically or send a case to a person.
What we make of it
This is our reading, not Google’s. Nothing in the announcement concerns decisions about people, and we do not suggest it does. We are interested in what it makes plain.
When a model writes no sentence, the interpretation has moved. It sits in the question and the rubric, written once by someone who could not know every case it would meet. The documentation notes that the model never sees the developer’s question keys, only the full question text. The number that returns carries no trace of that act of interpretation. A reader of the output cannot ask what was meant, because nothing was said.
A second point concerns attention. Confidence measures the machine’s uncertainty, not the person’s stake. If a case scores 0.97 nobody looks. A human sees the cases the model found hard. The documentation advises setting thresholds from labeled examples and weighing the cost of false positives against false negatives. That is sensible engineering. It also means that who gets a person’s attention depends on how legible they are to the model. In the Institute’s October commentary on who decides what counts as a reason, we asked whether an objection can reach someone equipped to examine it. That question becomes harder to answer when the decision arrives as a number.
A proposed illustration
This is a hypothetical. A municipal housing office uses a decision model to triage repair requests, asking, “Is this an emergency?” with an action threshold of 0.9. A tenant writes about a ceiling leak and mentions, in passing, that her mother’s oxygen concentrator sits beneath it. The question as written concerns water damage. The model returns 0.2, and the request goes to the routine queue. The clerk sees “routine” and a number. The tenant’s reason for urgency was in the message, but not in the question.
What the source does not say
Google’s documentation says explanations are a job for a different endpoint. It does not say what a person affected by a decision would be given. It does not say how builders should record the questions, rubrics and thresholds in force on the day. Google’s claims about benchmark performance are its own, and we have not tested the model.
If a decision comes back as a number, who is left to ask what it was meant to mean?
Signed articles reflect the views of their authors. Corrections: info@dghumana.com.