Murray Shanahan

Large language models (LLMs) exhibit compelling linguistic behaviour, and sometimes offer self-reports, that is to say statements about their own nature, inner workings, or behaviour. In humans, such reports are often attributed to a faculty of introspection and are typically linked to consciousness. This raises the question of how to interpret self-reports produced by LLMs, given their increasing linguistic fluency and cognitive capabilities. To what extent (if any) can the concept of introspection be meaningfully applied to LLMs? Here, we present and critique two examples of apparent introspective self-report from LLMs. In the first example, an LLM attempts to describe the process behind its own ``creative'' writing, and we argue this is not a valid example of introspection. In the second example, an LLM correctly infers the value of its own temperature parameter, and we argue that this can be legitimately considered a minimal example of introspection, albeit one that is (presumably) not accompanied by conscious experience.

The question of whether large language models (LLMs) can exhibit moral capabilities is of growing interest and urgency, as these systems are deployed in sensitive roles such as companionship and medical advising, and will increasingly be tasked with making decisions and taking actions on behalf of humans. These trends require moving beyond evaluating for mere moral performance, the ability to produce morally appropriate outputs, to evaluating for moral competence, the ability to produce morally appropriate outputs based on morally relevant considerations. Assessing moral competence is critical for predicting future model behaviour, establishing appropriate public trust and justifying moral attributions. However, both the unique architectures of LLMs and the complexity of morality itself introduce fundamental challenges. Here we identify three such challenges: the facsimile problem, whereby models may imitate reasoning without genuine understanding; moral multidimensionality, whereby moral decisions are influenced by a range of context-sensitive relevant moral and non-moral considerations; and moral pluralism, which demands a new standard for globally deployed artificial intelligence. We provide a roadmap for tackling these challenges, advocating for a suite of adversarial and confirmatory evaluations that will enable us to work towards a more scientifically grounded understanding and, in turn, a more responsible attribution of moral competence to LLMs.

Is it possible to articulate a conception of consciousness that is compatible with the exotic characteristics of contemporary, disembodied AI systems, and that can stand up to philosophical scrutiny? How would subjective time and selfhood show up for an entity that conformed to such a conception? Trying to answer these questions, even metaphorically, stretches the language of consciousness to breaking point. Ultimately, the attempt yields something like emptiness, in the Buddhist sense, and helps to undermine our dualistic inclinations towards subjectivity and selfhood.

The advent of conversational agents with increasingly human-like behaviour throws old philosophical questions into new light. Does it, or could it, ever make sense to speak of AI agents built out of generative language models in terms of consciousness, given that they are ‘mere’ simulacra of human behaviour, and that what they do can be seen as ‘merely’ role play? Drawing on the later writings of Wittgenstein, this paper attempts to tackle this question while avoiding the pitfalls of dualistic thinking.

Artificial Intelligence is making rapid and remarkable progress in the development of more sophisticated and powerful systems. However, the acknowledgement of several problems with modern machine learning approaches has prompted a shift in AI benchmarking away from task-oriented testing towards ability-oriented testing, in which AI systems are tested on their capacity to solve certain kinds of novel problems. The Animal-AI Environment is one such benchmark which aims to apply the ability-oriented testing used in comparative psychology to AI systems. Here, we present the first direct human-AI comparison in the Animal-AI Environment, using children aged 6–10. We found that children of all ages were significantly better than a sample of 30 AIs across most of the tests we examined, as well as performing significantly better than the two top-scoring AIs, “ironbar” and “Trrrrr,” from the Animal-AI Olympics Competition 2019. While children and AIs performed similarly on basic navigational tasks, AIs performed significantly worse in more complex cognitive tests, including detour tasks, spatial elimination tasks, and object permanence tasks, indicating that AIs lack several cognitive abilities that children aged 6–10 possess. Both children and AIs performed poorly on tool-use tasks, suggesting that these tests are challenging for both biological and non-biological machines.

The objectives of this article are twofold. First, by denying the dualism inherent in attempts to load metaphysical significance on the inner/outer distinction, it defends the view that scientific investigation can approach consciousness in itself, and is not somehow restricted in scope to the outward manifestations of a private and hidden realm. Second, it provisionally endorses the central tenets of global workspace theory, and recommends them as a possible basis for the sort of scientific understanding of consciousness thus legitimised. However, the article goes on to argue that global workspace theory alone does not constitute a fully worked-out objective account of the conscious subject. This requires additional attention to be paid to the issue of embodiment, and to the possibility of indexicality that arises when an instantiation of the global workspace architecture inhabits a spatially localised body

This paper presents a computer model of cortical broadcast and competition based on spiking neurons and inspired by the hypothesis of a global neuronal workspace underlying conscious information processing in the human brain. In the model, the hypothesised workspace is realised by a collection of recurrently inter-connected regions capable of sustaining and disseminating a reverberating spatial pattern of activation. At the same time, the workspace remains susceptible to new patterns arriving from outlying cortical populations. Competition among these cortical populations for influence on the workspace is effected by a combination of mutual inhibition and top-down amplification.

According to the singularity hypothesis, rapid and accelerating technological progress will in due course lead to the creation of a human-level artificial intelligence capable of designing a successor artificial intelligence of significantly greater cognitive prowess, and this will inaugurate a series of increasingly super-intelligent machines. But how much sense can we make of the idea of a being whose cognitive architecture is qualitatively superior to our own? This article argues that one fundamental limitation of human cognitive architecture is an inbuilt commitment to a metaphysical division between subject and object, a commitment that could be overcome in an artificial intelligence lacking our biological heritage

To understand the mind and its place in Nature is one of the great intellectual challenges of our time, a challenge that is both scientific and philosophical. How does cognition influence an animal's behaviour? What are its neural underpinnings? How is the inner life of a human being constituted? What are the neural underpinnings of the conscious condition? Embodiment and the Inner Life approaches each of these questions from a scientific standpoint. But it contends that, before we can make progress on them, we have to give up the habit of thinking metaphysically, a habit that creates a fog of philosophical confusion. From this post-reflective point of view, the book argues for an intimate relationship between cognition, sensorimotor embodiment, and the integrative character of the conscious condition. Drawing on insights from psychology, neuroscience, and dynamical systems, it proposes an empirical theory of this three-way relationship whose principles, not being tied to the contingencies of biology or physics, are applicable to the whole space of possible minds in which humans and other animals are included. Embodiment and the Inner Life is one of very few books that provides a properly joined-up theory of consciousness, and will be essential reading for all psychologists, philosophers, and neuroscientists with an interest in the enduring puzzle of consciousness.

In 1969, John McCarthy and Pat Hayes uncovered a problem that has haunted the field of artificial intelligence ever since--the frame problem. The problem arises when logic is used to describe the effects of actions and events. Put simply, it is the problem of representing what remains unchanged as a result of an action or event. Many researchers in artificial intelligence believe that its solution is vital to the realization of the field's goals. Solving the Frame Problem presents the various approaches to the frame problem that have been proposed over the years. The author presents the material chronologically--as an unfolding story rather than as a body of theory to be learned by rote. There are lessons to be learned even from the dead ends researchers have pursued, for they deepen our understanding of the issues surrounding the frame problem. In the book's concluding chapters, the author offers his own work on event calculus, which he claims comes very close to a complete solution to the frame problem. Artificial Intelligence series.

This article presents a formal theory of robot perception as a form of abduction. The theory pins down the process whereby low‐level sensor data is transformed into a symbolic representation of the external world, drawing together aspects such as incompleteness, top‐down information flow, active perception, attention, and sensor fusion in a unifying framework. In addition, a number of themes are identified that are common to both the engineer concerned with developing a rigorous theory of perception, such as the one on offer here, and the philosopher of mind who is exercised by questions relating to mental representation and intentionality.

This paper proposes a brain-inspired cognitive architecture that incorporates approximations to the concepts of consciousness, imagination, and emotion. To emulate the empirically established cognitive efficacy of conscious as opposed to non-conscious information processing in the mammalian brain, the architecture adopts a model of information flow from global workspace theory. Cognitive functions such as anticipation and planning are realised through internal simulation of interaction with the environment. Action selection, in both actual and internally simulated interaction with the environment, is mediated by affect. An implementation of the architecture is described which is based on weightless neurons and is used to control a simulated robot.