Perceptions of artificial intelligence system's aptitude to judge morality and competence amidst the rise of Chatbots.

This paper examines how humans judge the capabilities of artificial intelligence (AI) to evaluate human attributes, specifically focusing on two key dimensions of human social evaluation: morality and competence. Furthermore, it investigates the impact of exposure to advanced Large Language Models on these perceptions. In three studies (combined N = 200), we tested the hypothesis that people will find it less plausible that AI is capable of judging the morality conveyed by a behavior compared to judging its competence. Participants estimated the plausibility of AI origin for a set of written impressions of positive and negative behaviors related to morality and competence. Studies 1 and 3 supported our hypothesis that people would be more inclined to attribute AI origin to competence-related impressions compared to morality-related ones. In Study 2, we found this effect only for impressions of positive behaviors. Additional exploratory analyses clarified that the differentiation between the AI origin of competence and morality judgments persisted throughout the first half year after the public launch of popular AI chatbot (i.e., ChatGPT) and could not be explained by participants' general attitudes toward AI, or the actual source of the impressions (i.e., AI or human). These findings suggest an enduring belief that AI is less adept at assessing the morality compared to the competence of human behavior, even as AI capabilities continued to advance.

The Curve of Learning With and Without Instructions.

In skill acquisition, instructing individuals the stimulus-response mappings indicating how to perform and act, yields better performance. Additionally, performance is helped by repeated practice. Whether providing instructions and repeated practice interact to achieve optimal performance remains debated. This paper addresses that question by analyzing the learning curves of individuals learning stimulus-response mappings of varying complexity. We particularly focus on the question whether instructions lead to improved performance in the longer run. Via evidence accumulation modeling, we find no evidence for this assertion. Instructions seem to provide individuals with a head start, leading to better initial performance in the early stages of learning, without long-lasting effects on behavior. We discuss the results in light of related studies that do report long-lasting effects of instructions, and propose that the complexity of a skill determines whether long-lasting benefits of initial instructions exist.

Concurrent response and action effect representations across the somatomotor cortices during novel task preparation.

Instructions allow us to fulfill novel and complex tasks on the first try. This skill has been linked to preparatory brain signals that encode upcoming demands in advance, facilitating novel performance. To deepen insight into these processes, we explored whether instructions pre-activated task-relevant motoric and perceptual neural states. Critically, we addressed whether these representations anticipated activity patterns guiding overt sensorimotor processing, which could reflect that internally simulating novel tasks facilitates the preparation. To do so, we collected functional magnetic resonance imaging data while female and male participants encoded and implemented novel stimulus-response associations. Participants also completed localizer tasks designed to isolate the neural representations of the mappings-relevant motor responses, perceptual consequences, and stimulus categories. Using canonical template tracking, we identified whether and where these sensorimotor representations were pre-activated. We found that response-related templates were encoded in advance in regions linked with action control, entailing not only the instructed responses but also their somatosensory consequences. This result was particularly robust in primary motor and somatosensory cortices. While, following our predictions, we found a systematic decrease in the irrelevant stimulus templates' representational strength compared to the relevant ones, this difference was due to below-zero estimates linked to the irrelevant category activity patterns. Overall, our findings reflect that instruction processing relies on the sensorimotor cortices to anticipate motoric and kinesthetic representations of prospective action plans, suggesting the engagement of motor imagery during novel task preparation. More generally, they stress that the somatomotor system could participate with higher-level frontoparietal regions during anticipatory task control.

Voluntary task switching is affected by modality compatibility and preparation.

Cognitive task control can be examined in task-switching studies. Performance costs in task switches are usually smaller with compatible stimulus-response modality mappings (visual-manual and auditory-vocal) than with incompatible mappings (visual-vocal and auditory-manual). Modality compatibility describes the modality match of sensory input and of the anticipated response effect (e.g., vocal responses produce auditory effects, so that auditory stimuli are modality-compatible with vocal responses). Fintor et al. (Psychological Research, 84(2), 380-388, 2020) found that modality compatibility also biased task choice rates in voluntary task switching (VTS). In that study, in each trial participants were presented with a visual or auditory spatial stimulus and were free to choose the response modality (manual vs. vocal). In this free-choice task, participants showed a bias to create more modality-compatible than -incompatible mappings. In the present study, we assessed the generality of Fintor et al.'s (2020) findings, using verbal rather than spatial stimuli, and more complex tasks, featuring an increased number of stimulus-response alternatives. Experiment 1 replicated the task-choice bias to preferentially create modality-compatible mappings. We also found a bias to repeat the response modality just performed, and a bias to repeat entire stimulus-response modality mappings. In Experiment 2, we manipulated the response-stimulus interval (RSI) to examine whether more time for proactive cognitive control would help resolve modality-specific crosstalk in this free-choice paradigm. Long RSIs led to a decreased response-modality repetition bias and mapping repetition bias, but the modality-compatibility bias was unaffected. Together, the findings suggest that modality-specific priming of response modality influences task choice.

The effects of social presence on cooperative trust with algorithms.

Algorithms support many processes in modern society. Research using trust games frequently reports that people are less inclined to cooperate when believed to play against an algorithm. Trust is, however, malleable by contextual factors and social presence can increase the willingness to collaborate. We investigated whether situating cooperation with an algorithm in the presence of another person increases cooperative trust. Three groups of participants played a trust game against a pre-programmed algorithm in an online webhosted experiment. The first group was told they played against another person who was present online. The second group was told they played against an algorithm. The third group was told they played against an algorithm while another person was present online. More cooperative responses were observed in the first group compared to the second group. A difference in cooperation that replicates previous findings. In addition, cooperative trust dropped more over the course of the trust game when participants interacted with an algorithm in the absence another person compared to the other two groups. This latter finding suggests that social presence can mitigate distrust in interacting with an algorithm. We discuss the cognitive mechanisms that can mediate this effect.

Tenacious instructions: How to dismantle newly instructed task rules?

Humans excel in instruction following to boost performance in unfamiliar situations. We can do so through so-called prepared reflexes: Abstract instructions are instantly translated into appropriate task rules in procedural working memory, after which imperative stimuli directly trigger their corresponding responses in a ballistic, reflex-like manner. But how much control do we have over these instructed task rules when their reflexes suddenly lose their relevance? Inspired by the phenomenon of directed forgetting in declarative working memory, we here tested across four experiments whether the presentation of (implicit or explicit) task cancellation cues results in the directed dismantling of recently instructed task rules. Our findings suggest that-even when cancelation cues are actively processed-such dismantling does not occur (Experiment 1-3) unless the no-longer relevant task rules are replaced by a new set of rules (Experiment 4). These findings and their implications are discussed in the broader context of action control and working memory. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Examining mechanistic explanations for ideomotor effects.

Ideomotor (IM) theory provides a popular mechanistic account for understanding how goal-directed action can be learned and instigated. That is, when associations between actions and outcomes have been established in memory, the perception or thought of the outcome could automatically activate the associated action. Whereas a sizable literature provides evidence in line with this account, it does not successfully exclude alternative explanations in terms of propositions based on causal inferences. In the present paper, we present an online IM paradigm, in which learning and testing occurs on the same trials. In line with recent findings, we demonstrate that IM effects can emerge within a couple of trials, but also that people can update learned action-outcome associations immediately when a new mapping of outcomes on actions is introduced, without any switch costs. Together, this suggests that IM effects may be driven by propositions about causal relations, rather than bidirectional associations stored in memory. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Three levels at which the user's cognition can be represented in artificial intelligence.

Artificial intelligence (AI) plays an important role in modern society. AI applications are omnipresent and assist many decisions we make in daily life. A common and important feature of such AI applications are user models. These models allow an AI application to adapt to a specific user. Here, we argue that user models in AI can be optimized by modeling these user models more closely to models of human cognition. We identify three levels at which insights from human cognition can be-and have been-integrated in user models. Such integration can be very loose with user models only being inspired by general knowledge of human cognition or very tight with user models implementing specific cognitive processes. Using AI-based applications in the context of education as a case study, we demonstrate that user models that are more deeply rooted in models of cognition offer more valid and more fine-grained adaptations to an individual user. We propose that such user models can also advance the development of explainable AI.

Exploring the Link between Novel Task Proceduralization and Motor Simulation.

Our ability to generate efficient behavior from novel instructions is critical for our adaptation to changing environments. Despite the absence of previous experience, novel instructed content is quickly encoded into an action-based or procedural format, facilitating automatic task processing. In the current work, we investigated the link between proceduralization and motor simulation, specifically, whether the covert activation of the task-relevant responses is used during the assembly of action-based instructions representations. Across three online experiments, we used a concurrent finger-tapping task to block motor simulation during the encoding of novel stimulus-response (S-R) associations. The overlap between the mappings and the motor task at the response level was manipulated. We predicted a greater impairment at mapping implementation in the overlapping condition, where the mappings' relevant response representations were already loaded by the motor demands, and thus, could not be used in the upcoming task simulation. This hypothesis was robustly supported by the three datasets. Nonetheless, the overlapping effect was not modulated by further manipulations of proceduralization-related variables (preparation demands in Exp.2, mapping novelty in Exp.3). Importantly, a fourth control experiment ruled out that our results were driven by alternative accounts as fatigue or negative priming. Overall, we provided strong evidence towards the involvement of motor simulation during anticipatory task reconfiguration. However, this involvement was rather general, and not restricted to novelty scenarios. Finally, these findings can be also integrated into broader models of anticipatory task control, stressing the role of the motor system during preparation.

Automatic effects of covert practice.

Automatic behaviour is supposedly underlain by the unintentional retrieval of processing episodes, which are stored during the repeated overt practice of a task or activity. In the present study, we investigated whether covertly practicing a task (e.g., repeatedly imagining responding to a stimulus) also leads to the storage of processing episodes and thus to automatic behaviour. Participants first either responded overtly or covertly to stimuli according to a first categorization task in a practice phase. We then measured the presence of automatic response-congruency effects in a subsequent test phase that involved a different categorization task but the same stimuli and responses. Our results indicate that covert practice can lead to a response-congruency effect. We conclude that covert practice can lead to automatic behaviour and discuss the different components of covert practice, such as motor imagery, visual imagery, and inner speech, that contribute to the formation of processing episodes in memory.

An Episodic Model of Task Switching Effects: Erasing the Homunculus from Memory.

The Parallel Episodic Processing (PEP) model is a neural network for simulating human performance in speeded response time tasks. It learns with an exemplar-based memory store and it is capable of modelling findings from various subdomains of cognition. In this paper, we show how the PEP model can be designed to follow instructions (e.g., task rules and goals). The extended PEP model is then used to simulate a number of key findings from the task switching domain. These include the switch cost, task-rule congruency effects, response repetition asymmetries, cue repetition benefits, and the full pattern of means from a recent feature integration decomposition of cued task switching (Schmidt & Liefooghe, 2016). We demonstrate that the PEP model fits the participant data well, that the model does not possess the flexibility to match any pattern of results, and that a number of competing task switching models fail to account for key observations that the PEP model produces naturally. Given the parsimony and unique explanatory power of the episodic account presented here, our results suggest that feature-integration biases have a far greater power in explaining task-switching performance than previously assumed.

Erasing the Homunculus as an Ongoing Mission: A Reply to the Commentaries.

In our recent article (Schmidt, Liefooghe, & De Houwer, 2020, this volume), we presented an adaptation of the Parallel Episodic Processing (PEP) model for simulating instruction following and task-switching behaviour. In this paper, we respond to five commentaries on our article: Monsell & McLaren (2020), Koch & Lavric (2020), Meiran (2020), Longman (2020), and Pfeuffer (2020). The commentaries discuss potential future modelling goals, deeper reflections on cognitive control, and some potential challenges for our theoretical perspective and associated model. We focus primarily on the latter. In particular, we clarify that we (a) acknowledge the role of cognitive control in task switching, and (b) are arguing that certain task-switching effects do not serve as a good measure of said cognitive control. We also discuss some ambiguities in terminological uses (e.g., the meaning of "task-set reconfiguration"), along with some future experimental and modelling research directions.

The Instructed Task-Switch Evaluation Effect: Is the Instruction to Switch Tasks Sufficient to Dislike Task Switch Cues?

It is often argued that people dislike situations in which there is conflict requiring cognitive control, possibly because it is effortful to resolve this conflict. In a recent study, Vermeylen, Braem, and Notebaert (2019) provided evidence for this idea in the context of task switching. They observed that participants evaluated cues signaling a task switch more negatively than cues signaling a task repetition in a task switching paradigm. The present study examined whether this evaluative bias can be observed also on the basis of mere instructions. We instructed participants that two non-words would either signal the requirement to switch or to repeat tasks in an upcoming task switching block, which was actually never administered. In Experiment 1, we did not observe more positive implicit or explicit evaluations of the instructed task repetition compared to the task switch cue. In Experiment 2, participants first completed a task switching block in which a first pair of transition cues were used. We then provided task switching instructions that described the signaling function of a second pair of cues, which would be used in an upcoming (but never administered) second task switching block. Participants showed a clear preference for both instructed and experienced task repetition cues on explicit but not on implicit evaluations. Experiment 3 replicated the instructed task-switch evaluation effect on explicit evaluations in the context of prior task experience (but not without prior experience) and extended it to implicit evaluations. We discuss theoretical implications and potential explanations of this task-switch evaluation effect.

Motor imagery entails task-set inhibition.

Motor imagery requires the covert execution of a movement without any overt motor output. Previous studies indicated that motor imagery results in the prolonged inhibition of motor commands. In the present study, we investigated whether motor imagery also leads to the inhibition of more abstract task representations. To do so, we investigated the effect of motor imagery on n - 2 repetition costs, which offer an index of the extent to which task representations are inhibited. Participants switched among three tasks and among two response modes: overt and covert responding (i.e., motor imagery). N - 2 repetition costs were present when the current trial required an overt response but absent when the current trial required a covert response. Furthermore, n - 2 repetition costs were more pronounced when trial n - 1 required a covert response rather than an overt response. This pattern of results suggests that motor imagery also leads to the inhibition of abstract task representations. We discuss our findings in view of current conceptualizations of motor imagery and argue that the inhibitory mechanism entailed by motor imagery targets more than motor commands alone. Finally, we also relate our findings to the mechanisms underlying the inhibition of task representations.

Attention for future reward.

When stimuli are consistently paired with reward, attention toward these stimuli becomes biased (e.g., Abrahamse, Braem, Notebaert & Verguts, et al., Psychological Bulletin 142:693-728, 2016, https://doi.org/10.1037/bul0000047). An important premise is that participants need to repeatedly experience stimulus-reward pairings to obtain these effects (e.g., Awh, Belopolsky & Theeuwes, Trends in Cognitive Sciences 16:437-443, 2012, https://doi.org/10.1016/j.tics.2012.06.010). This idea is based on associative learning theories (e.g., Pearce & Bouton, Annual Review of Psychology 52:111-139, 2001) that suggest that exposure to stimulus-reward pairings leads to the formation of stimulus-reward associations, and a transfer of salience of the reward to the neutral stimulus. However, novel learning theories (e.g., De Houwer, Learning and Motivation 53:7-23, 2009, https://doi.org/10.1016/j.lmot.2015.11.001) suggest such effects are not necessarily the result of associative learning, but can be caused by complex knowledge and expectancies as well. In the current experiment, we first instructed participants that a correct response to one centrally presented stimulus would be followed by a high reward, whereas a correct response to another centrally presented stimulus would be paired with a low reward. Before participants executed this task, they performed a visual probe task in which these stimuli were presented as distractors. We found that attention was drawn automatically toward high-reward stimuli relative to low-reward stimuli. This implies that complex inferences and expectancies can cause automatic attentional bias, challenging associative learning models of attentional control (Abrahamse et al., 2016; Awh et al., 2012).

Stroop-like effects of derived stimulus-stimulus relations.

Automaticity can be established by consistently reinforcing contingencies during practice. During reinforcement learning, however, new relations can also be derived, which were never directly reinforced. For instance, reinforcing the overlapping contingencies A → B and A → C, can lead to a new relation B-C, which was never directly reinforced. Across 5 experiments we investigated if such derived relations can also induce automatic effects. We first trained participants to derive a relation between a nonsense word and a color word, and then used the nonsense words as distractors in a Stroop task. Results indicate that derived color-word associates induce Stroop effects. This effect, however, is present only when sufficient attention is allocated to the distractor words during the Stroop task, and is driven by a response conflict. We conclude that, under the present training conditions, derived color-word associates became related to the corresponding color word at the lexical level, but did not gain direct access to the corresponding semantic color representation. (PsycINFO Database Record (c) 2020 APA, all rights reserved).

Attentional prioritization reconfigures novel instructions into action-oriented task sets.

An astonishing aspect of human cognitive flexibility concerns the ability to efficiently convert complex symbolic instructions into novel behaviors. In such ability, the rapid transformation of relevant content into action plans is particularly crucial as it allows for reflexive, automatic-like execution of merely instructed task sets. However, little is known about the mechanisms that give rise to this transformation. In the current study, we test the hypothesis that novel instructions held in working memory are reformatted into action-oriented representations when selective attention prioritizes their content. To do so, we devised a paradigm in which participants first encoded 4 S-R mappings and later, a retro-cue selected two of them. We first found that participants can benefit from retro-cues during the implementation of novel task-sets. Then, across two preregistered experiments, we observed that cued mappings (but not uncued ones) induced intention-based reflexivity, suggesting that only these entered an action-oriented state. Altogether, our results reveal that selective attention prioritizes relevant novel instructed content, playing an important role in its prospective reformatting into an action-bound task set.

How Does the (Re)Presentation of Instructions Influence Their Implementation?

Instructions are so effective that they can sometimes affect performance beyond the instructed context. Such 'automatic' effects of instructions (AEI) have received much interest recently. It has been argued that AEI are restricted to relatively simple and specific S-R tasks or action plans. The present study put this idea further to the test. In a series of experiments based on the NEXT paradigm (Meiran, Pereg, Kessler, Cole, & Braver, 2015a) we investigated the specificity of AEI. In Experiment 1, we presented category-response instructions instead of S-R instructions. Nevertheless, we observed AEI for novel stimuli from the instructed category (Experiment 1a), and abstractness of the category did not modulate the size of the NEXT effect (Experiment 1b). However, Experiment 2 revealed specificity at the response level: AEI were much smaller in conditions where the instructed GO response is semantically related to, but procedurally different from the required NEXT response, compared to a condition where the NEXT and GO responses were the same. Combined, these findings indicate that AEI can occur when S(C)-R instructions are abstract at the stimulus level, arguing against previous proposals. However, AEI does seem to require specificity at the response level. We discuss implications for recent theories of instruction-based learning and AEI.

On the Assimilation of Instructions: Stimulus-response Associations are Implemented but not Stimulus-task Associations.

The assimilation of instructions consists of two stages. First, a task model is formed on the basis of instructions. Second, this model is implemented, resulting in highly accessible representations, which enable reflexive behavior that guides the application of instructions. Research frequently demonstrated that instructions can lead to automatic response activation, which indicates that stimulus-response associations can be implemented on the basis of a task model. However, instructions not only indicate how to respond (stimulus-response mappings) but also when (i.e., the conditions under which mappings apply). Accordingly, we tested whether instruction implementation leads both to the activation of stimulus-response associations and of associations between stimuli and the context or task in which the instructed stimulus-response mappings are relevant (i.e., stimulus-task associations). In four experiments, we measured if implementing newly instructed stimulus-response mappings also leads to bivalence costs (i.e., shorter latencies when a stimulus can only occur in one task compared to when it can occur in two tasks), which indicate the presence of stimulus-task associations. We consistently observed automatic response activation on the basis of instructions, but no bivalence costs. A discrepancy thus exists between information conveyed in an instructed task model and the elements of that task model that are implemented. We propose that future research on automatic effects of instructions should broaden its scope and focus both on the formation of an instructed task model and its subsequent implementation.

Attentional flexibility is imbalanced: Asymmetric cost for switches between external and internal attention.

Whereas the effects of attention switches occurring within perception or memory are relatively well understood, much less is known about switches of attention between them. We discuss the methodological limitations of initial research on this topic, which was never integrated with the broader cognitive literature. On the basis of this discussion, we present here a new paradigm, in which participants performed a simple probe-to-target matching task where targets were either perceived on screen or retrieved from memory. Across successive trials, repetitions or alternations (in both directions) between these 2 conditions were created, and eventually compared with each other. In line with our prediction, derived from the assumption of a top-down control mechanism, we found a cost for switching between external and internal attention in Experiment 1. Furthermore, this switch cost was asymmetric, being substantially larger when switching from (external) perception to (internal) memory than the other way around. In Experiments 2-4, we ruled out an imbalance in practice, learning, and preparation as confounds for this asymmetry. We propose that switches of attention between internal and external information are underpinned by a supervisory attention control mechanism, and that this asymmetry can be explained in terms of priming, associative interference or memory retrieval. (PsycINFO Database Record (c) 2019 APA, all rights reserved).