Consciousness in Artificial Intelligence: Insights from the Science of Consciousness — part 14/14

Cleaned-text code points [201302, 204217). Bibliography omitted here; source.md includes it.

However, we do recommend support for research on the science of consciousness and its application to AI (as recommended in the AMCS open letter on this subject; AMCS 2023), and the use of the theory-heavy method in assessing consciousness in AI. One way that we can make progress in learning which AI systems are likely to be conscious is by developing and testing scientific theories of consciousness; theoretical refinements within the existing paradigms are valuable, as well as attempts to test predictions of competing theories, as in the ongoing Cogitate adversarial collaboration (Melloni et al. 2021). A second is by undertaking careful empirical research to extend theories of consciousness to non-human animals. This will help us to establish a more general account of the correlates of consciousness, based on evidence from a wider range of cases (Andrews & Birch 2023). And a third is research that refines theories of consciousness specifically in the context of AI. Research of this kind may involve theorising about AI implementations of mechanisms implicated in theories of consciousness; building such systems and testing their capacities; identifying ambiguities in existing theories; and developing and defending more precise formulations of theories, so that their implications for AI are clearer. Integrating work of this kind with continued empirical research on human and animal consciousness can be expected to be especially productive.

Two other lines of research may be important. One is research specifically on valenced and affective consciousness; if conscious experience of this kind is especially morally important, as seems to be the case, then we have a pressing need for well-founded computational theories which can be applied to AI. The other is efforts to develop better behavioural tests for consciousness in AI. Although our view is that the theory-heavy approach is currently the most promising, it may be possible that behavioural tests could be developed which are difficult to game and based on compelling rationales, perhaps informed by theories. If such tests can be developed, they may have practical advantages over theory-heavy assessments.

As we learn more about consciousness, and gain better tools to assess whether it may be present in AI systems, we must also use these tools effectively. That means applying the theory-heavy method as we develop new kinds of AI systems, both prospectively—in assessing the likelihood of consciousness in planned systems, before we build them—and retrospectively. In retrospective evaluation, methods for mechanistic interpretability may be important, since it is possible that systems that do not have potentially consciousness-supporting features built in could acquire them in the course of training. It also means that scientific theories should be central to consciousness-related AI policy and regulation.

## Glossary