The present research's primary objective was to assess the effects of AIT on decision-making behavioral and neurological outcomes at different complexities. Our initial expectation was an improvement in execution with the presence of an AIT, particularly at higher complexities with an associated decrease in EEG activity. However, contrary to our expectations, our results demonstrated that introducing an AIT improved execution but also increased EEG activity.
Behavioral analysis
Most of our initial hypotheses concerning behavioral data were confirmed. Analysis of response times revealed a decrease as trials progressed consistent with the typical learning component seen in tasks such as IGT where participants' decision-making improves over time.10 Notably, the high complexity without an AIT condition exhibited a unique pattern. Unlike in most experimental tasks in which a decrease in reaction time accompanies execution improvement,11 this experimental condition did not show such a progression. Despite a reduced reaction time, the score remained stable over 100 trials showing no improvement. One possible explanation is that participants may have reached a point of giving up on finding the correct answer.
Giving up on a cognitive task can significantly influence reaction times. When individuals encounter persistent difficulties or perceive the task as exceptionally challenging, they may actively abandon searching for the optimal answer.12 This may result in decreased reaction times as participants opt for quick or intuitive decision-making rather than dedicating additional time to carefully consider available options which is supported by EEG data.13 This phenomenon may arise from cognitive fatigue, perceived frustration, and adaptation of response strategies to minimize effort ultimately affecting cognitive tasks’ decision-making speed. Thus, participants in this experimental condition may have perceived no discernible pattern or correct response prompting them to give up and shift to a faster approach to conclude the task promptly. During the interviews, several participants reported that they were unable to identify a deck that was better than the others. This feedback supports the notion that the lack of certainty regarding the best decision in the task led them to adopt a quicker, less reflective approach to complete the experiment.
The scoring data also provided insights into participant execution. As expected, most experimental conditions showed improvement as trials progressed. To observe a clear effect on behavioral and EEG data from using an AIT, a goal was to design an experimental condition that could not be resolved without it. This goal seemed to be achieved, as the high complexity without an AIT condition showed no improvement over 100 trials. A slight effect of AIT was anticipated for the low-complexity condition. However, such an effect was either not found indicating that throughout 100 trials execution in the lower-complexity condition did not differ significantly with or without AIT. This suggests that participants might have found heuristic alternatives that were more efficient than processing AIT information.
Heuristics or mental shortcuts emerge as swift and effective decision-making strategies when faced with environments characterized by information overload, by reducing cognitive complexities14. These simplifying responses enable individuals to reach conclusions more expediently, thus avoiding depletion of cognitive resources through exhaustive analysis. In settings where the problem is so complex that detailed analysis is not viable or even impossible, heuristic responses become invaluable. They offer satisfactory solutions without requiring in-depth data processing. Hence, in the lower-complexity condition, where the need for detailed information processing is minimal, participants were more inclined to ignore AIT suggestions and rely on heuristic approaches.
Intriguing findings emerged when analyzing data from the high-complexity AIT condition. While the anticipated pattern was a gradually improving performance, similar to the lower-complexity condition in the final trials, the data revealed no significant improvement in execution scores beyond trials 25 and 30. This result was perplexing, as the presence of decision support tools would typically lead to a more pronounced enhancement in efficiency.
A plausible explanation for this can be gleaned from the violin graphs which disclose distinct patterns in the distribution of this experimental condition compared to others. While the other distributions appeared relatively normal, the high complexity AIT condition exhibited an almost uniform distribution (Fig 4). This suggests a substantial variation in AIT’s effectiveness among participants. Some individuals used AIT to enhance their execution significantly, while others found no utility, resulting in scores akin to those without AIT. The heterogeneity in responses underscores the individual variability in AIT’s assimilation and application, contributing to the observed diversity in execution outcomes within this one experimental condition.
Individuals ignoring AI suggestions can be explained through cognitive overload, a condition where the information presented exceeds an individual's cognitive capacity to process it effectively.15 Human cognitive resources are finite, and when faced with an overwhelming volume of information individuals may experience difficulties in assimilating, analyzing, and integrating data.16 This overload often leads to a cognitive bottleneck hindering decision-making processes. In decision-making tasks involving AI suggestions, individuals may encounter situations for which the presented information, although generated by advanced algorithms, becomes too intricate or voluminous to be comprehensively processed within the available cognitive bandwidth. The capacity to manage information diminishes causing individuals to resort to simplified cognitive strategies, such as heuristics, to streamline decision processes and conserve mental resources.
Determining which individual characteristics influence the effectiveness or outright dismissal of AIT remains unclear.17 While our study did gather some verbal feedback from participants, the specifics of what influenced their choices were varied and complex. During the interviews, participants provided insights into their decision-making processes including their reasoning behind accepting or ignoring AIT suggestions. Some mentioned difficulty in trusting the AIT, while others relied on their intuition or heuristic strategies indicating a diverse range of factors at play. Interestingly, many participants noted that they believed the AIT was incorrect and therefore chose to completely ignore it. This contrasts with the behavioral data, which clearly shows that the AIT did improve participant performance, albeit not to the extent observed in the low complexity conditions.
However, the complexity and variability of these responses made it challenging to pinpoint definitive characteristics influencing AIT’s effectiveness. Additionally, some participants did not provide sufficiently detailed explanations making it difficult to draw concrete conclusions. Future research could benefit from a more structured approach to collecting and analyzing this type of qualitative data potentially uncovering clearer patterns and more specific individual characteristics that affect AIT utilization.
Neurophysiological analysis
EEG data analysis can provide valuable insights into the cognitive processes associated with AIT use. Traditionally, as task difficulty increases, a concurrent rise ensues in EEG activation indicating the intensified engagement of neural processes necessary for handling complex cognitive tasks.18 When faced with a more challenging task, the brain recruits additional neural resources to process and integrate information leading to increased neural firing and synchronization. This heightened activity often is observed in specific frequency bands, such as Beta and Gamma, associated with cognitive functions such as attention, working memory, and information processing.19,20 The increased EEG activation during more challenging tasks signifies deployment of cognitive control mechanisms and allocation of greater attentional resources. It reflects the brain's adaptive response to meet the demands of the task at hand.
The AIT condition was designed to provide cognitive support and an expected lower EEG activity in these conditions. In contrast, the conditions with higher complexity and without AIT would exhibit higher EEG activation. The results instead revealed an almost contrary pattern, particularly in the complex conditions, where low activation was observed without AIT, and high activation was noted with its presence. Minimal differences were found in EEG activity between the low-complexity conditions. As mentioned earlier, adopting a heuristic strategy could be one explanation for the lack of differences in the low-complexity conditions. In these conditions, inferring the best alternative was relatively straightforward since the decks presented no losses only wins. Most participants could identify the correct answer quickly making AIT's suggestions seemingly redundant. The EEG activity supports this interpretation. The patterns observed were quite similar with or without AIT indicating a comparable level of cognitive engagement regardless of AIT's presence in the low-complexity conditions.
The higher-complexity condition revealed a striking contrast in brain activity without AIT. Behavioral data suggested that participants reached a point of giving up in the search for the best alternative. This conclusion is substantiated by participants not showing improvement in their execution throughout 100 trials. In contrast to the other conditions, the score’s progression is a flat line while reaction times decrease. If participants persistently attempted to find the correct answer, one would expect a slower pace of decrease or even an increase in response times. However, despite making relative mistakes response times continued to decrease at a pace similar to that of the low-complexity condition. Consequently, the decrease in EEG activity could be attributed to a lack of engagement in the task. For most participants, it seemed that there was no discernible correct answer leading them to respond expeditiously to finish the experimental condition. Verbal reports from participants further support this notion with some expressing confusion or indicating a belief that their choices did not matter in this particular condition.
Using AIT resulted in an increase in EEG activity at higher complexities while a decrease in EEG activity might have been expected. Two factors may have played a role—increased information input and individual mistrust in AIT. Yet, a third is suggested by the data: enhanced subject engagement in the face of an enhanced challenge being buoyed by the AIT.
In terms of information input, while AIT offers a potentially easier alternative to evaluating each option manually, it substantially increases the information presented to the participant. This includes raw numerical data for each alternative and an adaptive suggestion for possible actions. Participants in this condition not only had to rely on experience but also had to evaluate the information provided. Consequently, in the lower-complexity condition, this information largely was ignored. However, participants still had the option to disregard this information and rely solely on AIT's suggestions. With this strategy, participants should be able to determine the best response alternative in fewer than 25 trials. Yet behavioral data demonstrated that almost no participant relied solely on AIT’s suggestions. While they considered such suggestions, they ultimately evaluated the information’s veracity and applied a personal approach to the problem. This resulted in increased response times and higher EEG activity. The inclination to validate AI suggestions and incorporate personal judgment may reflect relative mistrust—that is, a cautious approach driven by concerns about AI-generated information’s accuracy or reliability.
Mistrust in AIT would lead individuals to rely less on the provided AI-generated suggestions and more on their own judgment with significant influence on task execution. This mistrust may stem from various factors, including concerns about the information’s accuracy, reliability, or appropriateness.21,22 Individuals may question the AI system's ability to fully understand the task’s complexity or adapt suggestions to the person’s unique cognitive processes and decision-making strategies. Moreover, apprehensions about AI's lack of contextual understanding, potential biases, or limitations in learning from individual preferences can contribute to a sense of mistrust. Consequently, individuals may choose to validate AIT's suggestions against their own evaluation. This validation process can result in an increased cognitive load and longer decision-making times. Mistrust in AIT may see a paradoxical increase in task complexity as individuals reconcile the information provided by AIT with their judgments.
Nonetheless, AIT seems to function as a cognitive support tool in our experimental setting. Without AIT, participants appeared to abandon the search for an optimal answer.
When presented with an AIT, most participants persisted in trying to resolve the task, increasing scores as trials progressed—an effect not observed in the condition without AIT. Most participants used AIT as a guide, incorporating its suggestions into their decision-making process rather than following them blindly. This nuanced interaction with AIT suggests a balanced integration of AI support into individual decision-making strategies highlighting its role as a facilitator rather than a replacement for human cognitive processes as evidenced by the increase in cortical EEG activity.
We found interesting interactions for the experimental conditions when EEG activity is separated by frequency bands. The Theta band showed a pattern similar to the general cortical activation although it was mostly localized in the prefontal and occipital cortices. The Theta band is associated with cognitive processes such as attention, working memory, and mental engagement.23 The observed patterns in Theta band activity suggest that the experimental conditions influence these cognitive processes. In the high complexity with AIT condition, Theta band activity increased, suggesting heightened engagement in attention and working memory processes. This aligns with the behavioral data, indicating that participants persisted in trying to resolve the task with AIT’s aid. Considering previous findings, the role of the prefrontal and occipital cortices in the observed EEG activity patterns becomes particularly significant. These brain regions are crucial for various cognitive processes, and their involvement sheds light on the neural dynamics associated with decision-making and the impact of AIT use in different complexities.
The prefrontal cortex, implicated in executive functions, attention, and decision-making,24 showed distinctive patterns in Theta band activity. In the high complexity without AIT condition, the reduced Theta activity in the prefrontal cortex aligns with the behavioral data, indicating decreased engagement in attention and working memory processes. Conversely, increased Theta activity in the high complexity with AIT condition suggests heightened prefrontal cortex involvement when participants use AIT as a cognitive support tool.
Similarly, the occipital cortex, primarily responsible for visual processing and perception,25 exhibited relevant patterns in Theta band activity. These findings suggest that the experimental conditions influence visual attention and processing demands in decision-making tasks. The increased Theta activity in the high complexity with AIT condition implies a greater engagement of the occipital cortex when participants receive visual information from AIT.
In the Alpha band analysis, a prominent observation emerges from the decrease in EEG activity across most of the brain cortex in the high complexity without AIT condition. The Alpha band is commonly associated with relaxation states, cortical arousal inhibition, and decreased cognitive load.26 The marked decrease in Alpha band activity across most of the brain cortex suggests a shift in the neurophysiological state associated with AIT use in the experimental conditions. In the high complexity without AIT condition, the reduced Alpha band activity may reflect a reduction in cognitive load. This aligns with the behavioral data indicating potential disengagement or giving up in the search for the optimal answer.
The Beta activity pattern is similar to Theta’s. Indeed, the Theta-Beta bands have been associated with similar cognitive processes including attention, working memory, and engagement in decision-making tasks.27 In particular, the high complexity without AIT condition exhibited a notable decrease in Beta activity, suggesting a potential reduction in cognitive engagement and attentional resources. Conversely, in the high complexity with AIT condition, the Beta activity pattern showed an increase implying heightened cognitive engagement and attentional focus when participants used AIT.
Nonetheless, while Theta EEG activity changes occur mainly in the prefrontal cortex, for the Beta band they were found in the temporal and parietal cortices. These areas have been associated with distinct cognitive functions, and their involvement in Theta and Beta activity changes adds another layer of complexity to the interpretation of neural dynamics during decision-making tasks. Temporal and parietal areas have been linked to processes such as sensory integration, memory retrieval, and spatial processing.28,29 The localization of Beta activity changes in these regions suggests a differential engagement of neural circuits associated with distinct cognitive functions. The unique involvement of the temporal and parietal cortices in Beta activity changes emphasizes the multifaceted nature of cognitive processing during decision-making. The interplay between Theta and Beta activity across different brain regions underscores the intricate neural networks implicated in integrating information, memory, and attentional processes.
The Gamma band exhibited a prominent increase in EEG activity specifically in the high complexity with AIT condition. Gamma band activity is commonly associated with cognitive functions such as information processing, perception, integration of sensory stimuli, and overall higher-order cognitive functions.30,31 This observation implies an escalated cognitive demand unique to the experimental condition involving AIT. It is crucial to emphasize that AIT’s role is to provide support rather than completely override cognitive functions. Consequently, an enhanced cognitive function was required to adeptly respond to the task demands, as reflected in the heightened Gamma band activity. Interestingly, this pattern does not seem to manifest in the other experimental conditions, possibly because of a heuristic approach’s efficiency or a complete disregard for the task outcome.
The divergent patterns in EEG activity across different frequency bands shed light on the complex neural processes involved in decision-making tasks. Our findings emphasize the need to understand AIT's salience for decision-making behavioral outcomes and the neural dynamics involved. Future research could explore differences in decision support tool efficacy from alternative formats of presentation, address trust-related factors, and further dissect the neural mechanisms underlying adaptive decision-making in dynamic and complex environments. Ultimately, this study contributes to the growing body of literature bridging analytical decision support, artificial intelligence, cognitive neuroscience, and decision-making research.