AI Consciousness Engineering Workshop

The workshop asks what assurances can be given about the consciousness status of AI systems that are used now or built in the near future, making use of novel mathematical, statistical, and theoretical developments in consciousness science. In particular, it aims to assess the possibility of developing tools that ensure that artificial systems cannot develop consciousness. 

Topics include:

• Guarantees: What general types of guarantees can consciousness science currently offer about consciousness in AI systems, and under which theoretical assumptions do they hold?
• Non-consciousness by design: Are there reasonable methods, applicable to current architectures, for ensuring that an AI system is not conscious? Which confidence can these methods establish, and how can this be quantified? Which design choices raise or lower the risk of accidental consciousness?
• Formal verification: Can (non-)consciousness be formally verified via mathematical proof, in the way that properties of hardware and software are? What would a formal specification of a consciousness-relevant property look like?
• Robustness under theoretical disagreement: How should verification claims be stated when methodologies disagree, and how can they remain informative as the science develops?

The second theme turns to the space of minds that might be digitally instantiated. If artificial consciousness is achievable, it need not resemble human or animal consciousness; synthetic phenomenology may differ in structure, valence, unity, and temporal organisation, among other things. Equally, many systems may be mind-like without being conscious. The workshop aims to discuss and develop tools to construct and represent a space of possible minds, avoiding anthropocentric bias and oversimplification while still permitting meaningful comparison.

Questions include:

• Possible minds: What types of minds could AI systems have? How do architectural features constrain the space of possible experiences?
• Mind-mimics: How can conscious systems be distinguished from non-conscious systems that convincingly mimic minds–e.g. as in the gaming problem–and which intermediate categories, if any, are needed?
• Mind-like-systems: How can we talk clearly about systems that are not conscious, yet, still exhibit many of the characteristics of minds?
• Artificial phenomenology: How can we investigate the phenomenology of artificial systems without falling into speculation or implicit assumptions?
• Formal models: How can mathematical models of experience provide useful tools for elucidating the character of synthetic phenomenology?
• Instantiation: If we can specify the minds we care about, how could they be reliably instantiated, and how would we know that we had succeeded?

Building artificial consciousness, or assuming that it may emerge, without a deep understanding of the minds being created risks ethical and safety failures at massive scale, and any choice regarding consciousness engineering requires careful ethical considerations and analyses of safety implications. Mistakenly attributing consciousness to non-conscious yet mind-like systems equally carries its own costs. The third theme of the workshop seeks to address the responsibilities that come with this research on AI engineering.

Topics include:

• Societal implications: Which societal implications would AI consciousness engineering have? Which implications would populating the world with new conscious entities have, or with non-conscious but mind-like ones?
• Beneficial minds: What types of minds would we like AI systems to have, and on what grounds?
• Beneficial design: How can beneficial qualities be built into such systems to support long-term co-flourishing of humans and artificial minds?
• Dual use: What can the literature on dual-use technologies teach us about whether, when, and how to pursue research towards artificial consciousness?
• Evidence-based moratoria: In which situations, and up to which evidence, should moratoria for synthetic minds be considered? How can verified information on AI consciousness, and possibilities for shaping its design, help society to make informed decisions on which research to engage with, and which not?
• Governance: What role should consciousness science play in standards, regulation, and responsible development practices?