What happens to the discriminator after training?

Usually nothing; the generator is the product. But the judge can be reused for anomaly scoring or as a starting point for other classifiers.

What is mode collapse?

A failure where the generator finds one output the judge accepts and produces only that, winning the game while producing nothing useful.

Are GANs still the state of the art for images?

Diffusion models have largely taken over image generation, but the adversarial idea remains influential and the terminology survives.

AI Glossary / AI and machine learning / Intermediate

Discriminator

Definition

A discriminator is the judging half of a generative adversarial network, the model that learns to tell real examples from fakes while its partner, the generator, learns to fool it.

What it means

A generative adversarial network, or GAN, trains two models against each other. The generator makes synthetic examples, such as images, and the discriminator looks at a mix of real and synthetic ones and judges which is which.

Every time the discriminator catches a fake, the generator gets feedback on how it was caught and improves. Every time the generator fools the judge, the discriminator gets feedback and sharpens. Neither improves without the other.

The discriminator is itself an ordinary classifier, usually outputting a probability that its input is real. What makes it special is the moving target: its opponent trains against it continuously.

When training works, the end product people keep is the generator, now good enough to fool a strong judge. The discriminator is often discarded, though its learned eye for realism can be reused as a feature extractor or anomaly scorer.

For a business, discriminators explain both a capability and a risk: the same adversarial training that produces realistic synthetic media also produces the deepfakes your fraud and communications policies must now assume exist.

How it works

Initialize two networks: a generator that turns random numbers into candidate examples, and a discriminator that scores inputs as real or fake.

Train the discriminator on a batch mixing real data with the generator's current fakes, improving its judgments.

Train the generator to raise the discriminator's willingness to call its fakes real, using the judge's own gradients as the lesson.

Alternate these updates many times, watching for the known failure modes: the judge becoming perfect and giving the generator nothing to learn from, or the generator collapsing to one safe output.

Stop when the generator's output is indistinguishable enough for the purpose, then validate it on fresh real data the training never touched.

Three practical examples

Synthetic images: generating product mock-ups or design variants judged realistic enough for concept testing.
Anomaly detection: a reused discriminator scores how real a transaction or document looks, flagging the strange ones.
Synthetic training data: fakes good enough to supplement scarce real examples for other models.

Think of it this way

Think of a forger and an art expert training each other. Each fake the expert catches teaches the forger; each fake that hangs in the gallery teaches the expert. The expert is the discriminator: never exhibited, but the reason the final forgeries are so good.

The limit: if the expert is weak, the forger stays lazy. GAN quality is capped by the quality of the judging, which is why training balance is the hard part of the work.

Worked example

scoring invoices with a retired discriminator

Suppose a finance team receives 5,000 supplier invoices a month and can audit only a sample. The figures are illustrative.

A GAN trained on historical invoices yields a discriminator that has learned what ordinary invoices from each supplier look like.

The team retires the generator and uses the discriminator's real-or-fake score as an anomaly flag: the 2 per cent lowest-scoring invoices go to audit.

Audit hit rate doubles versus random sampling, because the discriminator's eye was trained by an opponent trying to beat it.

Illustrative case study

A fictional media firm bought a tool to detect AI-generated profile photos, assuming detection would stay ahead of generation. Within a year the generators had trained against exactly such detectors, and the tool's accuracy fell below usefulness.

The lesson they drew: the discriminator is always in an arms race, and buying one is buying a position in a race, not a finish line.

The invented lesson: any control built on spotting fakes needs a plan for the day the fakes graduate.

Common misconceptions and mistakes

The discriminator is the creative one. It only judges; the generator creates. Both are needed for training, but the roles are distinct.
A perfect discriminator means success. A perfect judge ends the game early: the generator gets no useful signal and training stalls.
Discriminators are only for images. The adversarial pattern applies to text, audio, tabular data and any domain where real and fake can be compared.

Business use

Assume synthetic media is cheap and good: verification processes, callback procedures, multi-channel confirmation, beat detection tools as durable defenses.

If you use GANs for synthetic data, keep the discriminator's scores: they are a free anomaly signal on real incoming data.

Ask vendors who claim fake detection how their detector is updated; an unupdated discriminator is yesterday's expert.

Check your understanding

In a GAN, why does the generator keep improving even though it never sees real data directly?

Answer: Because the discriminator's judgments carry what it has learned about real data back to the generator, so each caught fake teaches the generator how to fake better.

Related terms

Learn next

AI for Business Owners covers the practical risks of synthetic media and how to build processes that do not depend on spotting fakes: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary; [S2] Google, Machine Learning Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Discriminative model Next: Disfluency