AI glossary · Basic · Language models · Efficiency & infra · Safety & alignment
Decision model
A decision model is an AI model that, instead of generating sentences, judges with probabilities which of a set of predefined options applies. For example, it decides whether a post violates a policy or not.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Models & generation
- Added
- 2026-10-07
What it is and how it differs
A typical LLM (large language model) answers a question by writing out a long response. A decision model outputs a probability for each outcome instead of text. For example, it tells you how likely it is that 'this message falls under a prohibited category' versus 'it does not'. Because the options are fixed in advance, the biggest difference cited is that it can run faster and more cheaply than an LLM.
How it works
Inside, it uses the same transformer architecture as an LLM, so it keeps the flexibility to read and understand rules or policies written in plain language. But it narrows the output to a few predefined options and skips the process of generating a long answer one token at a time. As a result, it can apply complex criteria while judging as quickly as the simple classifiers used before. Another advantage often mentioned is that when the criteria change, the model does not need retraining; you just hand it the new criteria document.
Why it is getting attention
According to TechCrunch, TypeSafe AI released a decision model called 'Jev' in September, and OpenAI and Amazon soon released competing models, drawing attention to the field. The use most discussed early on was monitoring AI agents so they do not behave in unintended ways. It is now expanding to content moderation, reviewing posts made by people. The local runtime tool Ollama has also said that from v0.40.0 it supports several decision models on Apple silicon Macs, making it easier to try them on personal computers.
Things to know (limitations)
Answering with probabilities does not mean its judgments are always right. A model can be wrong while showing high confidence. A study that used Jev to verify supporting citations in financial reports found it sometimes missed cases where the numbers matched but the wrong item was cited. At the same time, it also failed to support other citations that were correct, and the study reported a trade-off between catching errors well and keeping correct evidence. Another study raises, in its title, the issue that such models may be swayed more by the labels on the options than by the definitions attached to them. Because the field is still new, the scope of the term and how performance is evaluated are still being worked out.
Example
Content moderation company Musubi released 'PolicyLM-1.7B', a lightweight decision model for real-time moderation, as open-weight. It applies a content policy written in plain English to messages and decides whether each message falls into a category defined by the policy. The company said it processes each item within 50 milliseconds and that its cost and speed are on par with the existing classifiers used by social platforms. It highlighted as a strength that the model does not need retraining when the policy changes, so policy staff can revise the criteria often.