Training Data Bias

Generative models, for all their power, come with very real failure modes. These are the ones you'll encounter most often in practice — with the strategies people use to handle them.

Challenge 1: Models Reflect Biases in Training Data

If you train an image generator on a dataset that overrepresents certain demographics or contexts, the model will generate accordingly. If you train a text model on internet text, the model inherits both the helpful and the harmful patterns of internet text.

Bochu Ding

The Latent Gaze

What does Stable Diffusion "see" when prompted with social labels like "an immigrant"? By generating thousands of portraits per label and computing pixel-level medians, this project surfaces the visual stereotypes embedded in a state-of-the-art image model.

Bochu DingRead article

chanind

Gender Bias Viewer

An interactive explorer for gender bias encoded in word2vec embeddings — see how word vectors associate professions, roles, and adjectives with gender through the geometry of the embedding space.

chanindOpen tool

⚠

Mitigation: Fairness and bias checks on training data

Run audits on dataset composition. Identify gaps. Adjust your data curation strategy. Document what you find. Modern responsible-AI practice goes far beyond a single bias check, but this is the floor.

Practically: tools like Hugging Face's dataset cards, Google's Know Your Data, and academic bias auditing frameworks all provide structured approaches. The key discipline is treating dataset auditing as first-class engineering work — not an afterthought before a launch.

←PreviousEvaluating Generative ModelsChoosing, Evaluating, and Deploying Next→GAN Mode CollapseChallenges and Mitigation Strategies