TLDR: This research paper introduces the SLP-tests (Subjective-Linguistic, Latent-Emergent, Phenomenological-Structural) as a framework to empirically test artificial consciousness. It reframes consciousness not as an intrinsic property, but as a functional interface between a relational substrate and observable behaviors, modeled using category theory. The tests evaluate an AI’s ability to discuss subjective experience, solve novel problems in emergent ways, and possess a structural “self” through which its actions factor, offering a practical approach to assessing consciousness in AI systems.
A new research paper titled “Artificial Consciousness as Interface Representation” by Robert Prentner proposes a novel framework for understanding and testing artificial consciousness. The paper shifts the focus from the contentious question of whether AI systems can possess consciousness, which is difficult to define and measure, to empirically tractable tests. This new approach introduces three evaluative criteria, collectively known as SLP-tests: S (subjective-linguistic), L (latent-emergent), and P (phenomenological-structural).
The core idea behind the SLP-tests is that artificial consciousness can be assessed by examining whether an AI system instantiates “interface representations.” These representations are modeled using category theory as mappings between relational substrates (RS) and observable behaviors. Essentially, consciousness is viewed not as an inherent property of a physical system, but as a functional interface to a relational entity. This means subjective experience is not constructed by the AI or brain, but rather “inherited” from a relational substrate through this interface layer.
The SLP-Tests Explained
The S-test, or subjective-linguistic test, is inspired by the “Artificial Consciousness Test” (ACT). It assesses whether an AI, even when isolated from vast amounts of internet data, would spontaneously discuss its own subjective experience or engage in philosophical reasoning about consciousness in ways familiar to humans. The key here is to determine if the AI can “connect” with a reality beyond its boundaries and express this connection through self-referential or subjectivist language. A challenge for this test is ensuring the AI isn’t merely parroting concepts from its training data.
The L-test, or latent-emergent test, broadens the scope beyond language to various behaviors. It involves placing AI systems in unfamiliar environments and observing how they solve complex problems in novel or unexpected ways. This could involve multiple AI instances interacting or hybrid systems deployed in new settings. Crucially, subjective experience in this context is not about the system’s emergent behaviors or internal representations, but about the connection between them, with the system’s representation being tuned to its own functioning.
Finally, the P-test, or phenomenological-structural test, delves into the phenomenal structure of the interface representations that emerge during the L-test. Inspired by phenomenology, this test seeks to formalize subjective experience mathematically. It proposes representing a minimal notion of “self” within the AI’s learned causal graph as a “colimit”—a central node or cluster of nodes through which all of the AI’s actions must factor. Ablating this “self” should significantly impair the AI’s performance, indicating its crucial role in the system’s functioning and its connection to a relational substrate.
Also Read:
- Understanding AI Assistants: A Deep Dive into OS Agents for Digital Device Control
- Teaching AI When to Stop Thinking: A Meta-Cognitive Approach for Large Language Models
Implications and Future Directions
The paper emphasizes that while each SLP-test has its limitations, together they offer a robust indicator for the emergence of interfaces through which subjective experience might arise. An artificially conscious agent, according to this framework, would create interface representations that provide a sense of perspective on the world, lead to actionable consequences in new problem spaces, and relate to “what it is like” to experience something.
The SLP-framework contrasts with other prominent theories of consciousness, such as Global Workspace Theory (GWT), Integrated Information Theory (IIT), Orchestrated Objective Reduction (OrchOR), and Attention Schema Theory (AST). While some theories might align with aspects of the SLP-tests, the interface theory offers a substrate-agnostic approach, focusing on observable behavior and representation rather than the underlying physical mechanisms. For instance, a high-Φ mechanism in IIT might simply be an efficient way to build a rich interface, and a global broadcast in GWT could be modeled as the functor in the interface framework.
The research also touches upon panpsychism and biopsychism, asserting that consciousness is not intrinsic to physical or living systems, but rather emerges through the construction of an interface. Furthermore, the SLP-tests are presented as complementary to, rather than replacements for, Artificial General Intelligence (AGI) benchmarks. AGI focuses on flexible problem-solving, while SLP-tests probe for the presence of the mediating interface representation. The S-test requires meta-representational machinery, and the L-test arguably detects online meta-learning, both hallmarks of AGI systems.
In conclusion, the paper highlights a fundamental distinction between subjective experience and intelligence. It calls for future developments including systematic SLP-benchmarking across various AI models, scalable P-tests for large networks, and the development of first-person protocols like “mind-meld” when technology allows. This framework opens the door to considering non-human and non-biological forms of subjectivity, suggesting that the forms of consciousness are potentially endless. You can read the full research paper for more details: Artificial Consciousness as Interface Representation.
