Learn · Intermediate

Discriminative models: AI that chooses instead of writes

A discriminative model is an AI system built to choose among defined answers or estimate a defined probability, rather than to write unrestricted text. It matters because many expensive ‘agent reasoning’ steps are really bounded decisions—pick a tool, rank a document, approve a form value, route a request—and a model designed for choosing can be faster, smaller, and easier to verify than a model designed to continue any sentence.

The basic contrast is between asking ‘what label fits this input?’ and asking ‘how could this entire input have been produced?’ In their classic comparison, Andrew Ng and Michael Jordan described discriminative classifiers as learning the relationship from an observation to a label, while generative classifiers model how observations and labels are jointly distributed. A spam classifier that outputs spam or not spam is discriminative. A language model that can write a new email from scratch is generative. Both can produce probabilities, but they are solving different problems.

Consider a support inbox with ten possible destinations. A generative model can read the message, explain its reasoning, invent a response, and finally recommend a team. That flexibility is valuable when the problem is novel. But if the application only needs one of ten queues, the writing step is extra work. A discriminative reader can convert the message into internal features, score all ten queues at once, and return the highest-scoring one plus a confidence estimate. It is the difference between asking a novelist to write an essay about which elevator button to press and using an elevator panel.

This is why encoders are common in discriminative work. An encoder reads the whole input bidirectionally and builds a representation of it. BERT, from Jacob Devlin and colleagues, became a landmark because a bidirectional transformer could be pre-trained on language and then fine-tuned for classification, question answering, and other understanding tasks. A decoder-only language model is optimized to predict the next token while respecting an order; an encoder is often a natural fit when the output is a label or score after seeing the whole input. The site’s guide to encoder, decoder, or both explains the architectural distinction.

A discriminative task needs an answer space. It may be two labels, such as fraud/not fraud; a list of named tools; an ordinal score from one to five; or a set of candidate values extracted from a document. The model produces scores called logits, converts them to probabilities with a function such as softmax, and chooses or ranks outputs. The mechanics are covered in softmax and cross-entropy. For a simple binary decision, the system may use a sigmoid-shaped probability; for many mutually exclusive options, softmax makes the scores compete.

The speed advantage follows from this structure. A text generator may need hundreds of sequential next-token steps before it produces a useful recommendation. A decision model can often encode the state once and score options in parallel. That does not make it universally better. If the correct option is missing, a classifier cannot invent it. If the state representation leaves out a key fact, a perfect classifier can still choose badly. This is the failure mode behind demonstrations that give a model a carefully structured game-state summary and then claim it has solved perception: the intelligence may be in the adapter that chose what to reveal.

Probability also needs humility. A model saying ‘0.97’ is not a guarantee; it is a statement about its learned scoring behavior. If it says 97% across many comparable cases, it should be right about 97% of the time for that number to be well calibrated. Distribution shift, imbalanced labels, and bad training data break that promise. Read calibration and confidence before treating a score as permission to take a high-stakes action. In an agent, a typed Boolean can prevent malformed output, but it cannot make a false premise true.

The practical pattern is a cascade. Use a discriminative model to reject obvious irrelevant requests, choose a known tool, rank retrieved documents, or validate a constrained form. Escalate ambiguous cases to a larger generative model or a person. This combines the cheap, inspectable part with the flexible part. Model routing and cascades explains why spending frontier-model tokens only where they add value is often the best system design.

The emerging ‘System One’ products make this old idea newly visible. Their novelty is not that classification was invented yesterday; it is that developers can call a typed decision layer in the same workflows where they once called only chat models. The right mental model is not ‘a tiny model that thinks like a giant one.’ It is ‘a specialized decision instrument.’ When the choices are clear, the labels are well defined, and the result can be checked, choosing instead of writing is often the more intelligent design.

Ground Truth. (2026, September 20). Discriminative models: AI that chooses instead of writes. Ground Truth. https://groundtruth.day/learn/discriminative-models-choosing-instead-of-writing.html

@misc{groundtruth:discriminative-models-choosing-instead-of-writing,
  title  = {Discriminative models: AI that chooses instead of writes},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/learn/discriminative-models-choosing-instead-of-writing.html}
}

Topics: classification · encoders · decision-models · probability · agents