Artificial Intelligence as Strange Intelligence: Against Linear Models of Intelligence

Researchers specializing in cognitive science and technology published a theoretical paper on the arXiv preprint server in February 2025 to challenge the prevailing 'linear model' of artificial intelligence progress. The authors aim to redefine how society perceives machine capabilities by introducing the concepts of 'strange' versus 'familiar' intelligence, arguing that current metrics fail to capture the non-linear, often contradictory nature of AI evolution. This publication serves as an expansion of philosopher Susan Schneider’s critique, suggesting that the industry's focus on a steady climb toward human-level reasoning is fundamentally flawed. The core of the research posits that AI does not develop along a predictable path similar to human biological maturation. Instead, the authors describe 'strange intelligence' as a phenomenon where a system may exhibit breathtaking superhuman capabilities in complex computational or creative tasks while simultaneously failing at basic logic or tasks a human child could master. This fragmentation defies traditional benchmarks, which often assume that proficiency in a high-level domain implies a foundational mastery of simpler, underlying concepts. According to the paper, this 'strange intelligence' manifests as a combination of profound insight and inexplicable error, sometimes within the exact same domain. This creates a disconnect for human observers who expect 'familiar intelligence'—the balanced, predictable growth seen in organic entities. By moving away from linear models, the researchers argue that developers and ethicists can better prepare for the risks and breakthroughs associated with systems that lack a cohesive, human-like cognitive structure, ensuring that security and safety protocols account for these unpredictable performance gaps.