Language models such as ChatGPT are not neutral. During training, they unconsciously pick up on prejudices about gender and ethnicity and reinforce them. In his doctoral thesis, AI researcher Oskar van der Wal has developed new measurement methods that can be used to better detect and reduce such hidden distortions.

Previous measurement methods are often too abstract and do not take into account real application, says van der Wal. In real-world tests, such as medical case descriptions that only varied the patient’s ethnicity, subtle but consistent differences in diagnoses and risk assessments were found. These differences remained invisible in standard tests.

The study also shows how prejudices arise during training: The model learns statistical relationships between words (e.g. “doctor” with “he”, “nurse” with “she”) and stores these patterns in an increasingly targeted manner.

Van der Wal tested targeted interventions at specific points in the model. The distortion could be significantly reduced without the general performance of the model suffering significantly.

The thesis is titled “Taking a Step Back: Measuring and Mitigating Bias in Language Models”. Oskar van der Wal will defend his dissertation at the University of Amsterdam on 29 April 2026. The supervision was provided by Dr. K. Schulz and Dr. W. H. Zuidema.

The study highlights the need to responsibly shape AI development at multiple levels – data, training, model architecture, and deployment.

Original Paper:

Oskar van der Wal, 2026, ‘Taking a Step Back: Measuring and Mitigating Bias in Language Models’. Supervisors: Dr K. Schulz.and Dr W.H. Zuidema.

Read Also:

ASKED: “Amazingly, ChatGPT replicates common stereotypes” – MedLabPortal

Editor: X-Press Journalistenbüro GbR

Gender Notice. The personal designations used in this text always refer equally to female, male and diverse persons. Double/triple naming and gendered designations are used for better readability. ected.