Intelligence artificielle et médecine : quelles règles éthiques et juridiques pour une IA responsable ?

Combining AI and medicine means talking about the medicine of the future, but even more about improving the quality of care. Its fields of application: predictive medicine, precision medicine, decision support, prevention, computer-assisted surgery, robotic support for the elderly, etc. are all related concerns: the possibility of maintaining human contact with the patient, the explicability of the algorithm - the collection of health data - and the improvement of the health care system. AI applications are already improving the quality of care. Its deployment at the heart of the medicine of the future is in constant evolution. To be acceptable and legitimate, the decisions of any algorithm must be understood and therefore explained. Only a good understanding of the diagnoses and therapies proposed by the AI application will allow doctors to discuss with their patients and to explain the possible alternatives. In the opposite case, the doctor risks to dismiss the use of algorithms because he will not be able to justify the decisions which will pose problems in terms of liability research in particular.     

What is it like to trust a rock? A functionalist perspective on trust and trustworthiness in artificial intelligence

The trustworthiness (or otherwise) of AI has been much in discussion of late, not least because of the recent publication of the EU Guidelines for Trustworthy AI. Discussions range from how we might make people trust AI to AI being not possible to trust, with many points inbetween. In this article, we question whether or not these discussions somewhat miss the point, which is that people are going ahead and basically doing their own thing anyway, and that we should probably help them. Acknowledging that trust is a heuristic that is widely used by humans in a range of situations, we lean on the literature concerning how humans make trust decisions, to arrive at a general model of how people might consider trust in AI (and other artefacts) for specific purposes in a human world. We then use a series of thought experiments and observations of trust and trustworthiness, to illustrate the use of the model in taking a functionalist perspective on trust decisions, including with machines. Our hope is that this forms a useful basis upon which to develop intelligent systems in a way that considers how and when people may trust them, and in doing so empowers people to make better trust decisions about AI.     

Where to from here? On the future development of autonomous vehicles from a cognitive systems perspective

Self-driving cars not only solve the problem of navigating safely from location A to location B; they also have to deal with an abundance of (sometimes unpredictable) factors, such as traffic rules, weather conditions, and interactions with humans. Over the last decades, different approaches have been proposed to design intelligent driving systems for self-driving cars that can deal with an uncontrolled environment. Some of them are derived from computationalist paradigms, formulating mathematical models that define the driving agent, while other approaches take inspiration from biological cognition. However, despite the extensive work in the field of self-driving cars, many open questions remain. Here, we discuss the different approaches for implementing driving systems for self-driving cars, as well as the computational paradigms from which they originate. In doing so, we highlight two key messages: First, further progress in the field might depend on adapting new paradigms as opposed to pushing technical innovations in those currently used. Specifically, we discuss how paradigms from cognitive systems research can be a source of inspiration for further development in modelling driving systems, highlighting emergent approaches as a possible starting point. Second, self-driving cars can themselves be considered cognitive systems in a meaningful sense, and are therefore a relevant, yet underutilized resource in the study of cognitive mechanisms. Overall, we argue for a stronger synergy between the fields of cognitive systems and self-driving vehicles.     

Teaming up with temps: the impact of temporary workers on team social networks and effectiveness

Temporary workers offer immediate benefits to the bottom line; yet, it is unclear how incorporating temporary workers into teams affects how they function. We apply social identity theory to propose that temporary workers significantly reduce individual- and team-level networks and team effectiveness but that commitment to the leader and intergroup competition can help temporary and permanent employees work together more effectively. Using a sample of employees nested in teams (Study 1, n = 312), we found that status differences affected member interactions resulting in sparser advice and friendship networks for temporary workers compared to their permanent counterparts. At the team level (Study 2, n = 58), these team member differences or contract diversity impacted team functioning through advice networks, such that, teams with greater contract diversity had sparser networks and were less effective. Further, commitment to the leader was found to moderate the negative impact of contract diversity on advice and friendship network density. With the increasing use of temporary worker and the prevalent use of teams, these findings have broader implications for HR functions and present possible avenues to mitigate the negative consequences of temporary workers.     

Non‐specific encoding of threat in social phobia and panic disorder

This study employs a release of proactive interference technique to explore encoding‐related processes in social phobia. Twenty‐six individuals with social phobia and 24 individuals with panic disorder participated in the memory task. Significant release of proactive interference was found when neutral and threat dimensions were encoded, but not when social and physical threat dimensions, or when positive and threat dimensions were presented. Threat was therefore differently encoded depending on the active encoding context in which semantic processing occurred. Individuals with social phobia or panic disorder did not differ in the release of proactive interference. The present experiment consequently yields little support for the specificity hypothesis in memory of anxious individuals. Instead, the findings generally demonstrate that threat encoding is similar for people with social phobia and those with panic disorder.     

Auditory expectations for newly acquired structures

Our study investigated whether newly acquired auditory structure knowledge allows listeners to develop perceptual expectations for future events. For that aim, we introduced a new experimental approach that combines implicit learning and priming paradigms. Participants were first exposed to structured tone sequences without being told about the underlying artificial grammar. They then made speeded judgements on a perceptual feature of target tones in new sequences (i.e., in-tune/out-of-tune judgements). The target tones respected or violated the structure of the artificial grammar and were thus supposed to be expected or unexpected. In this priming task, grammatical tones were processed faster and more accurately than ungrammatical ones. This processing advantage was observed for an experimental group performing a memory task during the exposure phase, but was not observed for a control group, which was lacking the exposure phase (Experiment 1). It persisted when participants realized an in-tune/out-of-tune detection task during exposure (Experiment 2). This finding suggests that the acquisition of new structure knowledge not only influences grammaticality judgements on entire sequences (as previously shown in implicit learning research), but allows developing perceptual expectations that influence single event processing. It further promotes the priming paradigm as an implicit access to acquired artificial structure knowledge.     

Variation in individuals' semantic networks for common knowledge is associated with false memory

Three experiments assessed the relationships between false memories of words and their degree of connectedness within individual semantic networks. In the first two experiments, participants studied associated word lists (e.g., hot, winter, ice), completed a recognition test that included related nonstudied words (e.g., cold, snow), and then rated the semantic relatedness of all word pairs including studied and nonstudied words. In the third experiment, the task order was reversed; participants completed pairwise ratings and then, two weeks later, completed the false memory task. The relatedness ratings were analysed using the Pathfinder scaling algorithm. In all experiments, items that an individual falsely recognized had higher semantic Pathfinder node densities than those items correctly rejected.