Cognitive biases: introduction and implications for AI
According to Tversky and Kahneman (1974), cognitive biases are systematic errors that affect and alter the decision-making process. These biases appear as distortions in judgments, resulting from erroneous processing of information. Moreover, a good understanding of biases can increase the likelihood of making more informed judgments and decisions, particularly in uncertain situations.
Biases, both cognitive and algorithmic, pose significant challenges in decision-making and data analysis. From algorithmic biases in machine learning, which can yield unfair results, to cognitive biases such as confirmation bias, anchoring, availability, representativeness, and selection, these biases all influence the way we interpret and process information, often unintentionally. Understanding them is essential to minimize their effects and ensure greater objectivity. The following table presents the definitions of various biases, as provided by the authors.
Biases, both cognitive and algorithmic, pose significant challenges in decision-making and data analysis. From algorithmic biases in machine learning, which can yield unfair results, to cognitive biases such as confirmation bias, anchoring, availability, representativeness, and selection, these biases all influence the way we interpret and process information, often unintentionally. Understanding them is essential to minimize their effects and ensure greater objectivity.
The following table presents the definitions of various biases, as provided by the authors.
Type of bias
Definition
Algorithmic bias
Algorithmic bias occurs when an algorithm generates unfair results that favour or harm certain groups. It is a critical issue in implementing AI systems, particularly in sensitive areas of security and criminal justice. It arises from unbalanced training data or design decisions influenced by unconscious biases, resulting in inaccurate and potentially harmful outcomes (Garcia, 2024).
Confirmation bias
Confirmation bias refers to seeking, interpreting, and retaining information confirming preexisting beliefs while ignoring evidence contradicting them (Kopitowski, 2016).
Anchoring bias
Anchoring bias is based on trust, which means that only the first information obtained is considered without questioning whether it is correct or inaccurate (Blanc, n.d.).
Availability bias
Availability bias refers to the tendency to overestimate the probability of an event occurring or its frequency, depending on how easily similar events can be recalled (Blanc, n.d.).
Representativeness bias
Representativeness bias refers to the judgment that considers an event probable if it represents a larger class based on the perceived relationship between a sample and a population (Martín & Álvarez, 2000).
Static bias, such as illusory correlation bias
Static bias occurs when a false correlation is established between two variables, generating misinformation and highlighting the inherent limits of AI systems (Hoza et al., 2000)
Selection bias
Selection bias is a systematic error that occurs when the method used to select subjects for a study and the factors that influence their participation are not adequately accounted for. This means that the association between exposure and response in the subjects analysed differs from that found in the eligible population (Rothman, 2024).
Coeckelbergh, in 2020, mentions that prejudice refers to “positive or negative discrimination against particular individuals or groups”. For example, in an ethical or political context, it is necessary to determine whether a specific form of bias is fair or unfair. Prejudices are systematic patterns of deviation from the norm or rationality in human judgment. They alter the perception of reality and can lead to interpretations or decisions that lack logic. They are mental shortcuts that, although they sometimes lead to errors, also help us to process information quickly.
Cognitive biases influence the performance and results of AI neural network models. Cognitive biases, therefore, represent a mechanism that alters both judgment and information processing. These biases can originate in both human and Artificial Intelligence. In the case of AI, cognitive biases have the particularity of distorting results, amplifying risks, and thus the psychological reactions of users.
Focusing on the above, Coeckelbergh 2020 argues that biases can arise not only from incomplete datasets but also from poor data quality, leading to predictions based on insufficient information. Additionally, the lack of diversity among AI development teams and data scientists remains a significant concern. In this context, biased decisions have a considerable impact at both the individual and community levels. However, these biases can be mitigated by techniques such as data augmentation, regularization, and the implementation of rigorous evaluation methods, which ensure that models can generalize appropriately, even when the data are not fully representative.
References
- Blanc, B. F. L. (S.F) Sesgos cognitivos, toma de decisiones e inteligencia artificial a reflexión.
- Coeckelbergh, Mark. (2020). AI ethics. Hoe omgaan met artificiële intelligentie?
- Gracia, J. L. B. (2024). El Sesgo algorítmico en la inteligencia artificial. Implicaciones para la seguridad global y la política criminal. GEOCONFLICTS & INTELLIGENCE: Revista sobre Seguridad Global y Terrorismo, 2(2), 183-203.
- Hoza, B., Waschbusch, D. A., Pelham, W. E., Molina, B. S. G., & Milich, R. (2000). Attention-Deficit / Hyperactivity Disordered and Control Boys ’ Responses to Social Success and Failure. 71(2), 432–446
- Kopitowski, K. (2016). La falacia de la evidencia incompleta o “cherry picking”. Evidencia, actualizacion en la práctica ambulatoria, 19(2).
- Martín, L., & Álvarez, A. (2000). Sesgos cognoscitivos del gerente: su influencia en la toma de decisiones. Revista Cubana de Salud Pública, 26(1), 5-11. Recuperado de: http://scielo.sld.cu/scielo.php?script=sci_arttext&pid=S0864-34662000000100001
- Rothman, K. J., Huybrechts, K. F., & Murray, E. J. (2024). Epidemiology: an introduction. Oxford university press.
- Tversky, A., y Kahneman, D. (1974). El juicio bajo incertidumbre: heurísticas y sesgos. Pensar rápido, pensar despacio. Buenos Aires: Debate.
0 Comments