Are all generative AI systems GANs?

No. GANs are one approach among several, with a distinctive generator-discriminator training relationship.

Must the discriminator be used whenever a trained generator produces a sample?

Not necessarily. In many applications, the generator can be used independently after training, while the discriminator's main role was providing a training signal.

What is mode collapse?

It is a failure pattern in which the generator produces too little variety, covering only a limited portion of the desired distribution.

AI Glossary / Governance, ethics and safety / Intermediate

Generative adversarial network

Definition

A generative adversarial network, or GAN, is a generative-learning framework in which a generator and a discriminator are trained with opposing objectives. The generator produces synthetic examples, while the discriminator learns to distinguish generated examples from training-data examples. Feedback through the discriminator helps the generator improve. The word “adversarial” describes this training relationship; it does not mean the system is necessarily designed for malicious use.

What it means

A GAN learns to produce data resembling a target distribution rather than simply assigning existing examples to categories. In a basic setup, the generator transforms a random input into a candidate example, such as an image. Conditional versions can also use additional information to guide the output.

The discriminator provides a learned training signal about the difference between generated and real examples. Both components change during training, so the generator is learning against a moving evaluator.

GANs are one family of generative models, not a synonym for every image generator, language model, or AI tool. Their training design and limitations differ from approaches such as diffusion models.

How it works

Sample real examples from the training data and random inputs for the generator. Use generated and real examples to update the discriminator so it better distinguishes the two sources under the chosen objective.

Then update the generator using a signal passed through the discriminator, encouraging outputs that the discriminator judges more like training examples. During this generator update, the discriminator serves as part of the computation supplying the learning signal rather than being updated in the same way.

Alternate or otherwise schedule these updates according to the selected GAN method. Monitor sample quality, diversity, and training behavior. Success is not guaranteed: training can become unstable, and the generator may cover only a narrow part of the target distribution even when some individual samples look convincing.

Three practical examples

1. A design-research prototype uses a GAN to generate variations of a visual texture, then has people evaluate whether the results are useful for the intended material.

2. An image-processing project uses a conditional adversarial setup to learn a transformation between specified types of visual input and output.

3. A machine-learning team studies GAN-generated examples as possible training augmentation, checking whether they improve performance on separate real data rather than assuming synthetic examples are automatically helpful.

Think of it this way

Think of a trainee artist and a reviewer repeatedly improving together: the artist makes examples, and the reviewer learns to distinguish them from the reference collection. The artist then uses the review signal to improve the next attempt.

The analogy is limited. The models do not have artistic intention or consciously deceive each other. Their updates are numerical optimization steps. Also, fooling a particular learned reviewer is not equivalent to satisfying every human quality requirement or reproducing the full diversity of the reference collection.

Worked example

Consider a toy discriminator that assigns a generated example a “real” score D(G(z)) of 0.2. Under a common non-saturating generator loss, the example's loss is −ln D(G(z)), so it is approximately −ln 0.2 = 1.609.

If a later generated example receives a score of 0.6 from the evaluated discriminator, the corresponding loss is −ln 0.6, approximately 0.511. A lower loss under this objective encourages the generator toward outputs that receive higher “real” scores.

This arithmetic isolates one example and one common objective; GAN variants use different losses. It does not prove that the later image is better, diverse, or safe to use. A changing or poorly trained discriminator can affect the score, so sample assessment and broader evaluation remain necessary. The discriminator score is not a universal authenticity certificate.

Illustrative case study

Fictional case: Pine Studio explores GAN-generated fabric patterns. Early samples look attractive, but a review of many outputs reveals repeated motifs and poor coverage of several styles in the reference collection.

The team evaluates diversity as well as visual plausibility and checks whether generated patterns are too similar to individual training examples. It compares results with the actual design brief rather than judging success from a few selected images.

This fictional case illustrates why convincing samples are not enough. It reports no production savings and does not assume that generated content is automatically original or suitable for commercial use.

Common misconceptions and mistakes

Do not equate GAN training with an adversarial attack on a deployed system. The competing objectives here are part of the learning framework.

Do not infer that a discriminator accuracy near chance proves the generator has learned the full target distribution. That result can also reflect a weak discriminator or other training problems.

Avoid treating realistic synthetic data as guaranteed anonymous or free of training-data influence. Also avoid judging the entire model from its best few examples. Poor diversity, repeated outputs, and unstable behavior can be hidden by selective presentation.

Business use

GANs can be relevant to visual generation, transformation, and research on synthetic data. A business evaluation should identify the exact task, available training material, quality requirements, and alternative methods.

Measure diversity and downstream usefulness alongside realism. If generated examples enter another model's training process, test that complete workflow on appropriate real evaluation data. The retained glossary category is governance, ethics and safety, but the term itself names a general generative-model framework with both technical applications and governance implications.

Check your understanding

A GAN generates five impressive images, but thousands of other outputs repeat the same pattern. Does the showcase establish a successful generator?

Answer: No. It demonstrates selected sample quality, not adequate diversity or reliable performance across outputs. Evaluation must examine the broader generated distribution and the intended task.

Related terms

Learn next

Learn the generator and discriminator objectives next, then study mode collapse and generative-model evaluation. Compare sample realism with distribution coverage, and keep training objectives separate from the criteria used to approve generated material for an application.

Sources

Goodfellow and colleagues, Generative Adversarial Networks: https://arxiv.org/abs/1406.2661

Google GAN course, structure, losses, and problems: https://developers.google.com/machine-learning/gan/gan_structure ; https://developers.google.com/machine-learning/gan/loss ; https://developers.google.com/machine-learning/gan/problems

Primary descriptions of training, losses, and failure modes were read. The arithmetic and Pine Studio case are original illustrations. This is a broad technical term despite its retained source-list governance category; specific variants require their own implementation details.

Previous: Generative adversarial imitation learning Next: Generative agents (simulacra)