Since we lack an empirically confirmed theory of consciousness that can be applied to artificial systems, some recent approaches explore indicators that provide at least some evidence for the presence of consciousness in artificial systems. Such approaches look for probability-raising indicators of consciousness that make it more likely that a given system is conscious.

I argue that we also need an explicit account of negative indicators of consciousness. A negative indicator is a factor that can significantly decrease the probability that a system is conscious. I propose a distinction between two types of negative indicators: indirect and direct negative indicators. While indirect negative indicators influence the extent to which other indicators are probability-raising, direct negative indicators lower the probability that an artificial system is conscious, regardless of which other indicators are present.

The advent of machine learning (ML) has led to the development of highly complex and opaque artificial intelligence (AI) systems. The opacity of these systems raises significant societal challenges concerning, e.g., fairness, trustworthiness, safety, responsibility, and autonomy. A widely proposed solution to these challenges is explainable AI (XAI). XAI is a research field aiming to explain the behavior of complex AI systems by mostly focusing on computational approaches. The explanations provided by XAI approaches are intended to induce understanding about opaque systems that allows various stakeholders to tackle the challenges described above.

While XAI is certainly helpful in at least some contexts, we believe that it is also crucially limited and thus not (yet) up to addressing many of the challenges it is intended for. In this paper, we highlight four reasons for XAI’s shortfall: (1) XAI fails to deliver faithful explanations, (2) even faithful explanations may fail to induce the right kind of understanding, (3) XAI may not provide the right kind of information, and (4) addressing some societal challenges requires more than explainability. We shall consider each of these points in turn before suggesting solutions.

Mindreading enables us to predict other people’s behaviour and facilitates appropriate action, including mindshaping, i.e., influencing other people’s minds to make them think or act in certain ways, e.g., exercised by politicians. Many higher animals like chimps and jays are also believed to engage in mindreading since, like humans, they lack the computational powers to track complex behaviours but consider the perspective and presence of others in doing what they do, e.g., food caching.

We suggest a form of pervasive mindreading that involves artificial intelligence, in particular, machine (or deep) learning technology. Hence, the label deep mindreading. Big tech companies like Meta, X, Google, and others, can draw on a wealth of personal data from individuals and make use of machine learning systems to detect patterns in their behaviours. These person profiles allow them to make accurate predictions about how and when people will vote, think, click, buy, act, and so on, based on what they believe, desire, and prefer, to engage in mindshaping (e.g., personalized advertising). In contrast to humans and animals, AI has the computational power to track behaviours to engage in mindreading and mindshaping. This privacy invasion raises ethical concerns.

Mind wandering (MW)—spontaneous, self-generated thought—is phenomenologically effortless, yet paradoxically linked to cognitive control, a process typically associated with mental effort. Influential accounts posit that cognitive control sustains MW by shielding internal thoughts from external interference and organizing them into a connected sequence. However, this raises a puzzle: if cognitive control is effortful, why does MW feel effortless? I resolve this tension by integrating the segmented structure of MW with the opportunity cost theory of mental effort. While cognitive control incurs mental effort, which reflects the opportunity cost (the value of forgone alternatives), MW’s segmented structure—thematically clustered thoughts forming brief segments, separated by topic shifts—prevents sustained investment of cognitive resources in any single segment/topic. Each segment briefly engages control, but rapid transitions reset costs before they accumulate to the extent that leads to phenomenological effortfulness. Thus, MW leverages control transiently across discrete segments, reconciling its reliance on control with its low-effort quality. This reframes the cognitive control view: control operates locally within each segment, not globally across the entire MW episode. My account thus deepens our understanding of MW and resolves the tension between the typical effortfulness of cognitive control and MW’s phenomenological effortlessness.