An interesting analysis by Yoshua Bengio and Eric Elmoznino

https://arxiv.org/abs/2308.08708

Main Points AI condensed:

1. Goal & Motivation
   The paper aims to bring scientific rigor to debates about whether AI could be, or become, conscious. It argues we should assess AI in light of our best neuroscientific theories of human consciousness. Hugging Face+1
2. Survey of Theories of Consciousness
   They review several well-supported theories from neuroscience and philosophy, including:
   • Recurrent Processing Theory
   • Global Workspace Theory
   • Higher-Order Theories
   • Predictive Processing
   • Attention Schema Theory fugumt.com+2arXiv+2
3. “Indicator Properties” or Markers
   From those theories, the authors derive a set of computationally-expressed indicators that, if present in an AI system, might signal features associated with consciousness. These indicators are meant to be testable/observable, not merely philosophical. arXiv+1
4. Assessment of Current AI Systems
   They evaluate several existing AI systems (language models, embodied agents, etc.) against those indicator properties. The result: no currently deployed AI meets enough of the indicators to make a good case for being conscious. Hugging Face+2arXiv+2
5. Feasibility & Barriers
   Despite that, they argue there are no obvious technical barriers to building systems that satisfy many of these indicators. It’s possible, in principle, to design an AI that implements enough of the relevant mechanisms. arXiv+1
6. Caveats & Limitations
   • Presence of indicator properties doesn’t guarantee consciousness (philosophical/theoretical gap remains). arXiv
   • Our understanding comes from human/neurobiological consciousness; consciousness might manifest differently in non-biological or very different architectures. fugumt.com+1
   • There are gray areas in interpreting whether an AI has a property, and many judgements are not clear-cut. Hugging Face+1
7. Implications
   • Introducing empirical, theory-grounded criteria may help public discussion, policy, ethics, and AI safety. fugumt.com+1
   • As AI advances, it’s increasingly important to track which consciousness-indicator properties new systems satisfy. arXiv