Taking AI Welfare Seriously

"Consider global workspace theory, which associates consciousness with a global workspace — roughly, a system that integrates information from mostly-independent, task-specific information-processing modules, then broadcasts it back to them in a way that enables complex tasks like planning."

2.2.2 Will some AI systems be conscious in the near future?, p. 17

This explicitly ties consciousness-related function to system-wide information integration and broadcast, aligning with the Information Integration phenomenon and suggesting analogous architectural markers in AI (e.g., attention convergence, aggregator tokens) could be probed in model internals .

"Global workspace theory 2.1 Multiple specialised systems capable of operating in parallel (modules) 2.2 Limited capacity workspace, entailing a bottleneck in information flow and a selective attention mechanism 2.3 Global broadcast of information in the workspace to all modules 2.4 State-dependent attention, giving rise to the capacity to use the workspace to query modules in succession to perform complex tasks"

2.2.2 Will some AI systems be conscious in the near future?, p. 16

Items 2.2 and 2.4 articulate gating and attention-based routing as core components, matching Selective Routing phenomena and suggesting concrete AI markers such as attention masks or expert routing that could be assessed in practice .

"We close this section with a brief note about the potential use of behavioral markers of AI welfare and moral patienthood. While we advocate for caution with behavioral markers at present, we also note that self-reports present a promising avenue for investigation, particularly for language models. Self-reports are central to our understanding of human consciousness... In the context of AI systems, particularly language models, self-reports could provide valuable insights into their internal states and processes, provided that we can develop methods to elicit and interpret them with sufficient reliability."

3.3 Assess, p. 37

This links reportability and metacognitive access to practical assessment pathways in LLMs, consistent with Self-Model & Reportability markers such as confidence heads or verifier modules in AI and metacognitive PFC/ACC signatures in biology .

"In this report, we use “sentience” to mean a particular kind of consciousness, namely positively or negatively valenced conscious experiences... Why might sentience suffice for moral patienthood? The idea that sentience is a sufficient condition for moral patienthood is very plausible and widely accepted, because when you can consciously experience positive and negative states like pleasure and pain, that directly matters to you."

2.2.1 Does consciousness suffice for moral patienthood?, p. 12

By tying welfare to valenced experience, this motivates looking for AI analogs of negative reward channels, aversive-cost signals, and persistence of negative states as potential markers of morally salient valence in artificial systems .

"Computational higher-order theories 3.1 Generative, top-down or noisy perception modules 3.2 Metacognitive monitoring distinguishing reliable perceptual representations from noise 3.3 Agency guided by a general belief-formation and action selection system... 3.4 Sparse and smooth coding generating a ‘quality space’"

2.2.2 Will some AI systems be conscious in the near future?, p. 16

The reference to sparse/smooth coding and ‘quality space’ aligns with Representational Structure phenomena and suggests measurable AI markers (e.g., SAE latents, representational geometry) for assessing structured experiential-like encodings .

"Other views hold that consciousness requires computational features that, at least at present, require biology in practice — such as specific kinds of oscillations that require specific kinds of chemical and electrical signals."

2.4.2 What if these features are insufficient for these capacities?, p. 26

By highlighting oscillatory and biophysical constraints, the authors point to Temporal Coordination mechanisms (oscillations, phase-locking, cross-frequency coupling) as potentially necessary features that current AI lacks, informing comparative assessments between brains and AI .