In this lecture we examine systematic distortions in Artificial Intelligence (AI), often referred to in English as "bias." Here, bias describes that an automated system consistently produces different outcomes for different people, groups, or situations, or systematically deviates from the desired or expected performance. These distortions can have technical causes, be related to social structures, or arise only through the conditions of use. The following explanations are based on established surveys and documented case studies cited in the sources (e.g., Mehrabi et al.; Buolamwini & Gebru; Angwin et al.). Statements about causes, consequences, and measures reflect the current scientific consensus as documented in the research literature where available; I clearly indicate existing uncertainties.
Bias in AI can be explained with simple analogies. One example: imagine you measure the length of objects with a ruler whose zero point is misplaced. All measurements are consistent but systematically shifted — that is a technical error comparable to statistical bias. Another image: a mirror that shows only a section of a scene; if it represents certain groups less often or incorrectly, it gives a distorted view of reality. Comparable effects arise in AI systems, but the reasons are usually more complex.
Concrete research examples illustrate reality: For commercial gender classification systems, studies showed clear differences in recognition accuracy between people of different skin tones and genders, with darker-skinned women being misclassified more often (see Buolamwini & Gebru, 2018). In another area, a journalistic investigation reported that an automated risk model in the U.S. criminal justice system exhibited different types of errors for different racial groups (see Angwin et al., 2016). Such cases are not isolated incidents but reveal structural problem points that can appear in many applications.
How do such distortions typically arise? In short: through the data, through the modeling approach, and through the way systems are used. Datasets can underrepresent groups or reflect historical inequalities. Labels can contain human prejudices or be inaccurate. Model architectures and optimization objectives (for example, maximizing average accuracy) can unintentionally disadvantage groups. And in operation, systems can be applied in contexts for which they were not validated.
Important: bias is not only a technical problem. Failures often reflect social inequalities: when historical discrimination is encoded in data, a system reproduces these patterns even without "intent." Literature and case studies show that careful causal analysis, documented datasets, and appropriate evaluation metrics are central steps to understand distortions (Mehrabi et al., 2019; Barocas & Selbst, 2016).
To systematically address distortions, it is helpful to use technical terms and categories. A common taxonomy distinguishes several types of bias in data and models. In the survey by Mehrabi et al. (2019) common categories include historical bias (when data reflect past inequalities), representation bias (when certain groups are underrepresented), measurement or collection bias (when feature collection is flawed), aggregation bias (when a model forces a single model across heterogeneous groups), and evaluation bias (when test data do not represent the diversity of deployment cases). These categories help target causes.
At the model evaluation level there are different fairness definitions discussed in the literature. Examples are statistical parity (proportions of positive decisions should match between groups), equality of error rates (e.g., equal false positive and false negative rates across groups), or predictive calibration (same ratio of predicted risk to actual event probability within each group). Research makes clear that these definitions are not always simultaneously satisfiable; there are mathematical trade-offs to consider when selecting a fairness definition (Mehrabi et al., 2019; Barocas & Selbst, 2016). Which definition is appropriate depends on the application and societal objectives.
To reduce bias the technical literature distinguishes three strategies: pre-processing measures, where datasets are altered or reweighted; in-processing measures, where the learning procedure includes fairness constraints or regularizers; and post-processing measures, where predictions are adjusted afterwards. Practical examples are re-sampling or re-weighting of training data, incorporating fairness-oriented cost functions, adversarial regularization, or calibration and threshold adjustments for subgroups. Meta-measures such as creating "Datasheets for Datasets" and "Model Cards" serve to document datasets and models and improve transparency and accountability (Gebru et al., 2018; Mitchell et al., 2019).
Some concrete technical findings from research: word and language models can learn stereotypical associations; a well-known example is gender-related stereotypes in word embeddings, which can be mitigated by targeted debiasing methods without solving all structural problems (Bolukbasi et al., 2016). Documentation and auditability of data are repeatedly cited prerequisites for responsible applications because many sources of bias become visible only after close examination of data collection and labeling processes (Gebru et al., 2018).
Beyond technical measures, experts call for institutional steps: clear governance, regular audits, interdisciplinary teams, and where appropriate, external review bodies. The European "Ethics Guidelines for Trustworthy AI" formulate principles such as transparency, accountability, and human oversight that can play a role in policy and organizational frameworks (European Commission, 2019). However, the scientific literature shows that no single measure solves all problems; a bundle of technical, organizational, and regulatory steps is typically required.
Uncertainty: For many applications robust, widely accepted procedures to measure and remedy bias are still lacking. Demographic metadata are often missing from datasets, so problems may only be discovered later. Studies show progress but also open questions, for example how to guarantee fairness across multiple stages of models or in adaptive systems over the long term (Mehrabi et al., 2019).
Applications: Bias can have consequences in many areas. In recruitment processes automated filters can disadvantage qualified applicants if training data contain past prejudices. In lending, models can reinforce credit allocation patterns if historical data reflect discrimination. In medicine insufficient data coverage of certain population groups can lead to worse diagnostic performance for those groups. The exemplary studies on face recognition (Buolamwini & Gebru, 2018) and criminal justice risk assessment (Angwin et al., 2016) show real-world consequences of different kinds; they imply that risks must be analyzed and measured systematically before systems are widely deployed.
Limits of countermeasures: Technical debiasing methods can bring local improvements but do not automatically yield fair outcomes in all deployment contexts. Some methods reduce certain types of errors but increase others or decrease overall performance. Moreover, technical measures do not replace the need for organizational rules, user education, and, if necessary, legal requirements. The literature therefore warns to choose solutions context-sensitively and to document them transparently (Mitchell et al., 2019; Gebru et al., 2018).
Small thought experiment: Suppose you are to develop a face recognition model for a city parking authority to recognize valid parking permits. Which steps would you take to avoid bias? Consider at least three measures, and reflect on possible goal conflicts (e.g., between accuracy and fair performance across all groups).
One possible answer might include points such as: (1) Check whether the available training data represent the local population; if necessary collect additional targeted data; (2) Document datasets (Datasheets) and evaluate performance across relevant subgroups; (3) Select appropriate fairness metrics and test which trade-offs emerge; (4) Set up a process for continuous monitoring and auditing; (5) Define deployment rules, e.g., human review in critical cases. Note that every measure can have side effects that should be considered in advance (Mehrabi et al., 2019; Gebru et al., 2018).
Final assessment: Bias in AI is a demonstrable problem with multiple causes. Research offers a range of methods for analysis and mitigation, but there is no universal silver bullet. A combination of careful data collection, technical control, transparent documentation, and organizational governance is, according to the current scientific state, best practice to reduce risks (Mitchell et al., 2019; Barocas & Selbst, 2016; European Commission, 2019).