AI system learns like animals do, raising questions about machine reasoning
Researchers have demonstrated that artificial intelligence can acquire behavior through reinforcement—the same psychological mechanism that trains animals—and then apply those lessons to novel situations. The finding suggests a viable path toward more flexible AI systems, though experts remain divided on whether this approach will meaningfully advance the push toward artificial general intelligence.
Originaltitel: From operant learning to arbitrarily applicable relational responding: a review of Machine Psychology with the non-axiomatic reasoning system
<p>Machine Psychology is an emerging interdisciplinary framework that integrates principles from learning psychology with a cognitive AI architecture to advance Artificial General Intelligence (AGI) research. This article provides a focused review of Machine Psychology, tracing the progression from basic operant learning to advanced symbolic reasoning within the Non-Axiomatic Reasoning System (NARS). We first outline the theoretical foundations in operant conditioning and Relational Frame Theory, highlighting how adaptive behavior and arbitrarily applicable relational responding (AARR) serve as cornerstones of human cognition. We then describe the architecture and capabilities of NARS and its variant OpenNARS for Applications (ONA), which enable real-time sensorimotor reasoning under conditions of uncertain knowledge. Four successive experimental studies are reviewed in detail: (1) Operant conditioning tasks demonstrate that NARS can learn from reinforcing consequences to modify its behavior, achieving 100% correct responses and adapting when contingencies change. (2) In generalized identity matching, NARS abstracts an identity relation that successfully generalizes to novel stimuli after minimal training. (3) A functional equivalence study shows NARS grouping stimuli by shared consequences, such that new learning transfers spontaneously between equivalent stimuli. (4) Finally, NARS is extended to model AARR, exhibiting derived symmetric and transitive relations and context-sensitive relational reasoning (e.g. same–opposite relations) with associated transformations of stimulus functions. We discuss how specific NARS mechanisms (e.g. temporal inference, variable term introduction, relational implication) map onto psychological processes underlying learning and cognition. Machine Psychology is presented as a developmental roadmap toward human-like AGI, incrementally building cognitive skills from basic adaptation to complex symbolic reasoning. We critically evaluate the strengths and limitations of this approach and outline open research directions toward achieving flexible, theory-of-mind-capable intelligence.</p>