[Paper Review] Consciousness in Artificial Intelligence: Insights from the Science of Consciousness

Patrick Butlin, Robert P. Long+17 more|arXiv (Cornell University)|Aug 17, 2023

Explainable Artificial Intelligence (XAI)184 citations

TL;DR

A theory-driven, neuroscience-based rubric assesses AI consciousness via indicator properties derived from prominent theories; finds no current AI is conscious but near-term construction of systems meeting indicators may be feasible.

ABSTRACT

Whether current or near-term AI systems could be conscious is a topic of scientific interest and increasing public concern. This report argues for, and exemplifies, a rigorous and empirically grounded approach to AI consciousness: assessing existing AI systems in detail, in light of our best-supported neuroscientific theories of consciousness. We survey several prominent scientific theories of consciousness, including recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory. From these theories we derive "indicator properties" of consciousness, elucidated in computational terms that allow us to assess AI systems for these properties. We use these indicator properties to assess several recent AI systems, and we discuss how future systems might implement them. Our analysis suggests that no current AI systems are conscious, but also suggests that there are no obvious technical barriers to building AI systems which satisfy these indicators.

Motivation & Objective

• Motivate and justify a scientifically tractable approach to AI consciousness.
• Propose a rubric of indicator properties derived from neuroscientific theories.
• Assess existing AI systems against the indicator properties.
• Explain how future AI systems might implement the indicators.
• Discuss limitations, uncertainties, and societal implications of conscious AI in light of the indicators.

Proposed method

• Adopt computational functionalism as a working hypothesis to relate AI computations to consciousness.
• Survey leading scientific theories of consciousness (e.g., recurrent processing theory, global workspace theory, higher-order theories, predictive processing, attention schema theory).
• Derive a set of indicator properties from these theories that are necessary or jointly sufficient for consciousness in AI.
• Analyze how AI systems could implement these properties using standard machine learning methods.
• Evaluate specific AI systems (e.g., Transformers, Perceiver, DeepMind Adaptive Agent, embodied multimodal models) as case studies against the indicators.
• Argue that the rubric is provisional and subject to refinement as research progresses.

Experimental results

Research questions

• RQ1 What indicator properties derived from neuroscientific theories best track consciousness in AI under computational functionalism?
• RQ2 Can current AI architectures implement these indicator properties, and to what extent do they resemble conscious processing?
• RQ3 Are there practical or theoretical barriers to constructing AI systems that satisfy the indicator properties?
• RQ4 How should we reason about credence in AI consciousness given theory support and empirical evidence?
• RQ5 What moral and social considerations arise from potential conscious AI and how should they be addressed?

Key findings

• No current AI systems meet all indicator properties for consciousness under the proposed rubric.
• Several indicator properties are implementable with existing AI methods (e.g., algorithmic recurrence and some attention/agency features).
• Some architectures and systems partially exhibit properties related to global workspace or embodied agency, but none are strong candidates for consciousness.
• A theory-heavy, computational-functionalism approach provides a tractable framework for assessing AI consciousness and can guide future development.
• There are no obvious technical barriers to building AI systems that satisfy larger subsets of the indicators, though satisfying indicators does not guarantee consciousness.
• The rubric is explicitly provisional and should evolve with ongoing theoretical and empirical advances.