Source-linked AI summary
Inferring Affective Consciousness in an Artificial Agent: A Case Study
Mark Solms, St John Grimbly, Bruce Bassett, Evert Boonstra, Rowan Hodson, Nicolas Kuske, Kival Mahadew, Benjamin Rosman, Charel van Hoof, Jonathan Shock
TL;DR
The paper asks how a physical theory of mind might account for subjective experience. It presents a deterministic artificial agent with affective uncertainty that exhibits hedonic place preference, suggesting that such behaviour may support attributing a point of view to the agent.
Problem
The paper addresses the lack of a clear physical account of the subjective character of experience.
Method
The paper develops a simple artificial agent that prioritizes homeostatic needs by adjusting uncertainty about action policies in an uncertain environment.
Results
The agent exhibits hedonic place preference through false inferences under uncertainty while remaining deterministic and driven by expected free-energy reduction.
Takeaways & Limitations
The authors argue that functional-mechanistic features of an agent may justify attributing a point of view even when its behaviour is deterministic.
Takeaways & Limitations
The hedonic place preference test may demonstrate reward-signal hijacking and behaviour as if the agent had feelings rather than actual feelings.
Abstract
from arXiv · showhide
Creatures that display 'hedonic place preference behaviour' are thought by many scientists to experience feelings, on the assumption that their attraction to pleasure-producing substances which lack nutritional value (e.g. cocaine, morphine) cannot easily be attributed to unconscious instinctual behaviour. In this paper, we discuss how a simple artificial agent that instantiates attributes of an affective system engaging in felt uncertainty about its intrinsic needs in relation to environmental resources can similarly display hedonic place preference behaviour -- through an apparently subjective form of information processing -- while simultaneously being entirely deter-ministic. We outline some implications of this artificially engineered behaviour for our understanding of the physical basis of consciousness and the experience of free will.
1. Introduction
The paper investigates whether an artificial agent could have affective consciousness, arguing that this question should be addressed through the system’s structure and function rather than behaviour alone. It presents affect as a fundamental, functionally engineerable form of consciousness while treating the associated philosophical assumptions as contextual rather than necessary to the argument.
- Motivation: The authors ask how an artificial intelligence could achieve conscious mental states and propose evaluating this through its structure and function, not behaviour alone.They also seek advance agreement on how consciousness in their developing AI system would be recognized.
- Theoretical assumptions: The paper assumes that affect is the most fundamental form of consciousness and enables perceptual and cognitive consciousness.This assumption is framed in both evolutionary and physiological terms.
- Functional account of feeling: The authors characterize feeling as a valenced, qualitatively distinctive, subjective state that is selectively prioritized over unfelt states.They derive this account from affect’s intrinsic valence, qualitative differentiation, and selective salience.
- Artificial affective agency: They argue that a system could be engineered to feel because affective consciousness is grounded in substrate-independent functions including active inference, cybernetics, autopoiesis, and the free energy principle.The proposed agent has endogenous goals and existential stakes, unlike an LLM whose outputs have no consequences for survival.
- Proposed mechanism: The agent’s decision making prioritizes homeostatic needs amid uncertain environmental opportunities, while fluctuations in confidence about achieving survival are proposed as the physical instantiation of affect.The paper argues that inspecting these mechanisms is more informative than treating hedonic place preference behaviour itself as decisive.
4. Conclusion
The conclusion argues that a relatively simple computational architecture can exhibit the relevant mechanisms and behaviours, while emphasizing that the mechanism behind hedonic place preference matters more than passing the test itself. It further frames the central issue as whether the agent’s functional-mechanistic organization justifies attributing it a point of view.
- Conclusion: A relatively simple architecture can exhibit the described mechanisms and behaviours by partitioning programming around survival.The agent accesses simulated bodily and world states only through its Markov blanket.
- Conclusion: The crucial issue is the mechanism by which the agent might pass the hedonic place preference test, not the test result itself.The authors suggest removing trickery and instead having the agent infer its bodily state through uncertainty when pushed away from its preferred state.
- Conclusion: The conclusion identifies functional-mechanistic justification for attributing a point of view as the issue at stake.The authors reject this attribution for most AIs but regard it as apparently justified for their blanketed agent with itinerant affective dynamics, while acknowledging the impasse of other minds.