The Generalisation Illusion: A 2025 Psychological Audit of Artificial Intelligence

The year 2025 is defined by the breathtaking technical success of Large Language Models (LLMs), which routinely surpass human-level performance on highly specialised, knowledge-based tasks. This brilliance, however, requires a critical psychological audit, revisiting the fundamental question of how much genuine intelligence resides within artificial intelligence. This analysis updates the critical perspective of Van der Maas, Snoek, and Stevenson (2021), concluding that while AI has masterfully acquired crystallised intelligence (the ability to process and synthesize vast statistical data) it remains fundamentally constrained by its struggle with fluid intelligence, specifically generalization - the ability to apply learned concepts to completely novel, Out-of-Distribution (OOD). Empirical data from rigorous psychological benchmarks, such as the MM-IQ framework, suggests that state-of-the-art models perform only marginally better than random chance on tests requiring abstract, flexible reasoning. This persistent generalisation gap confirms that LLMs are sophisticated statistical mimics, reinforcing the need for forward-looking cognitive science and ethical oversight to guide the creation of truly reliable and prosocial AI.