AI systems appear to solve difficult mathematics more effectively than humans because they can store extensive information in an external symbolic workspace, which functions as a larger working memory than the human brain.

Human working memory has limited capacity, typically holding only a few unfamiliar items at once, which constrains the ability to manipulate multiple variables and equations simultaneously. Research indicates that working memory contributes uniquely to mathematical performance beyond general intelligence measures, as shown in studies of children by Alloway and Passolunghi (2011) and Alloway and Alloway (2010).

Studies have found that early working‑memory performance predicts later numeracy even after accounting for IQ (Alloway and Alloway 2010). Additional research by Blankenship et al. (2015) shows that working memory explains unique variation in mathematical fluency after controlling for IQ and age. A meta‑analysis by Friso‑van den Bos et al. (2013) confirms a consistent relationship between working memory and mathematics across primary‑school samples. Because mathematical reasoning can be expressed explicitly through symbols, an AI model can keep the entire problem statement, intermediate calculations, and constraints within its context window, which may span tens of thousands of tokens. This external storage allows the model to track many assumptions and steps without relying on internal mental limits, a capability humans lack without paper or notes.

AI's advantage is most pronounced on problems that involve many interacting constraints, long sequences of calculations, extensive case analysis, and precise symbolic bookkeeping, according to the working‑memory hypothesis. Conversely, tasks that depend on a single insight or on interpreting ambiguous social or historical information do not benefit as greatly from a larger context, since the core difficulty lies in forming the correct causal model rather than in memory capacity.

Formal verification tools can test mathematical solutions against defined rules, providing clear feedback that reinforces AI's strength in explicit reasoning. In less formal domains, feedback is weaker, and a long, coherent argument may still be incorrect, indicating that memory alone does not ensure correctness outside mathematics.

While AI can process and retain large amounts of explicit information, it lacks the private, continuously updated internal state that humans maintain silently. Thus, AI's apparent “thinking out loud” reflects its use of an expanded written workspace rather than a deeper form of intelligence. The distinction suggests that AI's current superiority in mathematics may stem from its architecture and access to extensive external memory rather than from superior raw reasoning ability.

The comparison implies that a fairer assessment of AI's mathematical ability would involve matching it with humans who have comparable external tools and verification methods, rather than judging unaided human performance.