Yes, in ways that have been documented. Studies have found ChatGPT produces more negative associations for Arab names than English names, generates higher-income professional roles for men and lower-income roles for women in hypothetical scenarios, and produces responses that reflect cultural biases in its predominantly English-language training data. OpenAI acknowledges bias as an ongoing challenge and publishes research on its mitigation efforts. But any model trained on human-generated text will reflect human biases to some degree — the question is whether those biases are measured, disclosed, and actively reduced.

Partially. Bias can be reduced through careful data curation, diverse training sets, fairness constraints during training, post-processing adjustments, and regular auditing. But there are mathematical limits: some fairness criteria are provably incompatible with each other — you can't simultaneously satisfy all commonly used fairness definitions. Reducing bias completely would require eliminating historical inequity from training data, which would require rewriting history. The practical goal is measurable improvement, transparency about remaining bias, and avoiding deployment in contexts where the bias creates serious harm.

Several approaches: counterfactual testing (changing protected attributes like race or gender and observing if outputs change), disparity analysis (comparing error rates or outcome rates across demographic groups), audit studies (sending matched applications through AI hiring systems), and red-teaming (adversarially testing for biased outputs). The EU AI Act requires bias testing for high-risk AI systems. Third-party auditing firms like Parity AI and Fairly AI have emerged to provide independent bias assessment.