The New AI Consciousness Paper – By Scott Alexander
a day ago
- #AI Consciousness
- #Ethics of AI
- #Computational Theories
- Most discourse on AI consciousness is of low quality, with AI responses about consciousness being unreliable due to conflicting biases.
- A recent paper by Yoshua Bengio, David Chalmers, and others explores computational theories of consciousness, avoiding unresearchable or trivial approaches.
- Key computational theories discussed include Recurrent Processing Theory (RPT), Global Workspace Theory (GWT), and Higher Order Theory (HOT).
- The paper tentatively concludes that current AI systems are not conscious but sees no technical barriers to creating conscious AI in the future.
- The distinction between access consciousness (introspection and reporting) and phenomenal consciousness (subjective experience) is crucial but often conflated.
- AI systems like LLMs lack feedback loops required by some theories (e.g., RPT), but architectures like MaMBA may meet these criteria.
- Philosophers struggle to define phenomenal consciousness, with theories often reverting to access consciousness for practicality.
- Future AI may pass advanced Turing Tests, raising ethical and practical questions about treating them as conscious beings.
- AI companies balance making AI seem humanlike without triggering discomfort, leading to varied perceptions of AI consciousness.
- The paper highlights risks of both under- and over-attributing consciousness to AI, emphasizing the need for clear recognition to prevent suffering or misallocation of resources.