Helena Matute

Introduction: AI decision‐support systems have become increasingly integrated into judicial contexts. This research asks about their influence on human judgements, particularly when they are erroneous or biased. Method: Experiment 1 examined whether algorithmic recommendations conflicting with the evidence previously presented through standardized testimonies could change the judgements of participants acting as mock jurors. Experiment 2 replicated Experiment 1 and warned participants about potential AI errors. Results: Experiment 1 showed that participants’ judgements shifted in the direction suggested by the algorithmic recommendation. Experiment 2 replicated the findings of Experiment 1 and further showed that warning participants about potential AI errors and biases reduced algorithmic influence, particularly when the algorithm suggested guilt and the testimonies suggested innocence. Discussion: These findings highlight the importance of understanding how algorithmic recommendations influence legal decision‐making and how excessive trust in AI recommendations could be reduced. Data: The data supporting the findings of this study are openly available in the Open Science Framework at https://osf.io/9gwh7 .

Despite their widespread adoption, Artificial Intelligence-based Patient Classification Systems sometimes rely on incorrect, outdated, or incomplete data, which can lead to inaccurate outputs. Nevertheless, health professionals are expected to override these errors, at least when they have access to critical information. To test this, we conducted two experiments in which professional physicians interacted with an Artificial Intelligence system that incorrectly classified fictitious patients as either highly or lowly sensitive to a treatment. The physicians administered the treatment to a series of fictitious patients and received feedback that was useful for learning that the patient classification was incorrect and that all patients were equally sensitive to the treatment. We ran two experiments: in Experiment 1, the medicine showed medium effectiveness for both types of patients, while in Experiment 2, the treatment was completely ineffective for both types of patients. The results showed that, in the two experiments, physicians generally trusted the AI-based patient classification and struggled to learn from the evidence. Furthermore, in Experiment 2, they failed to realize that the treatment was ineffective. Our findings have important implications for healthcare professionals and patients, underscoring the need to critically evaluate Patient Classification Systems.

Empirical evidence has demonstrated the power of AI to influence human decisions and the risk of humans acquiring AI biases. Therefore, there is a clear need to develop strategies to mitigate such threat. In three experiments, set in a medical context, we tested whether warning individuals about AI biases and errors could mitigate the negative impact of AI biases on their decisions and reduce the transmission of AI biases to humans. In Experiment 1, participants received explicit information about the percentage of erroneous AI recommendations but with two different framings: in terms of AI accuracy or AI risk of error. Our results showed that emphasising the risk of AI errors, more than its accuracy, reduced people’s tendency to follow incorrect AI suggestions and to acquire biases from AI. In Experiment 2, a more general warning message alerting of possible AI errors and biases was also effective in reducing bias acquisition. Experiment 3 showed that, although the warning message provided some protection against bias, participants who received AI support still made more errors than participants who completed the classification task without any assistance. Experiments 2 and 3 also investigated whether the type of error made by the AI, a false positive or a false negative, influenced participants’ tendency to adhere to its suggestions, and the effect of the warning message. However, no significant effects were found. Overall, our results highlight the importance of informing users about the risk of AI error rather than focusing solely on accuracy.

The causal illusion is a cognitive bias that involves believing that one event causes another when it does not. It has negative consequences in different spheres of life, including health. Therefore, diverse interventions have been designed to reduce it. The more common ones are educational interventions. These include different elements related to improving both, motivation and information. We wanted to explore which of the two factors was more important for their effectiveness. We first used financial incentives to promote motivation (experiments 1a and 1b), but did not find them effective. Second, we used debiasing instructions about what has to be done to infer the causal relationship between two events accurately. This effectively reduced the causal illusion when the circumstances were in place for the illusion to be high (experiment 2). We discuss the results and their theoretical and practical implications.

Patient classification systems (PCS) support clinical decision-making but may rely on incorrect, outdated, or insufficient data. Doctors can sometimes override errors using their experience. However, certain factors such as scarcity of resources could lead to reliance on incorrect PCS recommendations, with consequences for patients. We conducted two experiments where participants interacted with a PCS that incorrectly classified fictitious patients as more or less sensitive to a treatment. Participants had the opportunity to administer the treatment on a series of patients, and use the feedback to learn that the PCS was wrong and all patients were equally sensitive. This was tested in contexts of abundant and scarce resources. Additionally, the treatment was effective in Experiment 1, but ineffective in Experiment 2. Results indicate that people generally trust the PCS recommendation, to some extent neglecting the information they collect during the task. This can lead to uneven resource allocation, especially in scarcity conditions, and incorrect perceptions of effectiveness, which in Experiment 2 implies believing that an ineffective treatment works. We preregistered the experiments, and all data and materials are public.

In order to conduct controlled and replicable experiments simulating legal trials, we present ForenPsy, the first standardized and open testimonial database in Spanish. We curated nine cases (three for each type of crime: homicide, threats, and burglary), with 14 testimonies each (seven suggesting innocence, seven suggesting guilt), resulting in 126 testimonies. Three hundred participants responded two questions for each testimony: one about whether the testimony suggested innocence or guilt, which showed that the index of agreement between the expected judgment and the actual judgment was .85, and another rating the degree of guilt suggested by each testimony, which yielded significantly lower values for innocence testimonies than for guilt testimonies, thereby showing that the testimonials worked properly. ForenPsy, as well as the standards for testimonies, are available at the OSF, so that it can be used and improved by researchers interested in conducting controlled and replicable experiments simulating judicial environments. Para realizar experimentos controlados y replicables simulando juicios hemos creado ForenPsy, el primer banco de testimonios estandarizado y abierto en español. ForenPsy incluye nueve historias (tres de cada tipo de delito: homicidio, amenazas y allanamiento) con 14 testimonios cada una (siete de inocencia y siete de culpabilidad), lo que hace un total de 126. Trescientos participantes respondieron dos preguntas sobre cada testimonio: una sobre si el testimonio indicaba inocencia o culpabilidad, que muestra que el índice de acuerdo con el valor esperado fue .85, y otra sobre el grado de culpabilidad que atribuían a cada testimonio, que fue significativamente inferior para los testimonios de inocencia que para los de culpabilidad, lo que indica que los estímulos funcionaron adecuadamente. ForenPsy, así como las normas de los testimonios, está disponible en OSF y puede utilizarse y mejorarse de manera colaborativa para realizar experimentos replicables simulando contextos judiciales.

Causal illusions consist of believing that there is a causal relationship between events that are actually unrelated. This bias is associated with pseudoscience, stereotypes and other unjustified beliefs. Thus, it seems important to develop educational interventions to reduce them. To our knowledge, the only debiasing intervention designed to be used at schools was developed by Barberia et al. (Barberia et al. 2013 PLoS One 8 , e71303 (doi: 10.1371/journal.pone.0071303 )), focusing on base rates, control conditions and confounding variables. Their assessment used an active causal illusion task where participants could manipulate the candidate cause. The intervention reduced causal illusions in adolescents but was only tested in a small experimental project. The present research evaluated it in a large-scale project through a collaboration with the Spanish Foundation for Science and Technology (FECYT), and was conducted in schools to make it ecologically valid. It included a pilot study ( n = 287), a large-scale implementation ( n = 1668; 40 schools) and a six-month follow-up ( n = 353). Results showed medium-to-large and long-lasting effects on the reduction of causal illusions. To our knowledge, this is the first research showing the efficacy and long-term effects of a debiasing intervention against causal illusions that can be used on a large scale through the educational system.