Artificial Sentience: definition, architecture, and implementation

Background. In the rapidly evolving field of Artificial Intelligence (AI), recent developments have significantly enhanced AIs’ ability to interact with their environment, from accepting multi-modal inputs to solving previously unencountered problems in an iterative fashion. AI’s increasing capacity for human-like communication also has surpassed the capabilities set out by the now-obsolete Turing Test. At the same time, the recognition of sentience in various invertebrate marine creatures, like lobsters and octopuses, challenges the traditional, human-centric views of cognition. The juxtaposition of expanded advancements in AI with the evolving understanding of biological cognition highlights the need to redefine the concepts of sentience, self-awareness, self-cognition, and, ultimately, sapience in artificial entities. The lack of a clear definition of sentience leads to discretion, ambiguity, and delays in decision-making, contradicts the precautionary principle, and affects the treatment and well-being of sentient entities. 

Methods. This paper identifies shared foundations of multiple cognition theories and assesses these foundations against the constraints of two contrasting examples of sentient beings, a lobster and a blind paraplegic person, to outline the minimal sentience criteria. The modular architecture built around an existing LLM model is designed, implemented, and tested against the defined criteria. 

Results. The initial analysis synthesizes cognitive criteria and identifies two convergence points defining the entity’s sentience: its metacognitive capacity, i.e., its ability to differentiate between itself and the environment, and its capability for adaptation to the environment. These provide a straightforward framework for evaluating sentience in biological and non-biological entities, effectively narrowing the gap in how sentience can be studied across these divergent categories. A set of tests was developed to evaluate information-based entities' metacognition and adaptation capabilities, offering insight into such entities’ perception and decision-making processes and opening a path to the development of AI psychology, which will be required for the ethical and responsible development of future generations of self-cognizant artificial entities in the likes of Artificial General Intelligence or Artificial Sapience.

The successful implementation of the designed architecture resulted in the creation of the first reported Artificial Sentience (AS) entity.