AI models lag behind AGI-level reasoning despite recent advances

Apple researchers have found that leading AI models still fall short of humanlike reasoning, casting doubt on claims that artificial general intelligence (AGI) is imminent.

In a June paper titled The Illusion of Thinking, Apple’s team evaluated major large reasoning models (LRMs), including ChatGPT and Claude, using custom puzzle games rather than standard coding or math benchmarks.

While recent AI updates show gains on conventional tests, the study found that these benchmarks fail to capture broader reasoning capabilities.

The researchers tested both “thinking” and “non-thinking” model variants and found performance dropped sharply as task complexity increased.

“We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across puzzles,” the paper stated.

Additionally, they observed that AI has a propensity to overthink, frequently beginning with accurate solutions before deviating into flawed reasoning as the responses developed.

These patterns suggest current models imitate reasoning without internalising it, lacking the kind of generalisable thinking associated with AGI.

The study concluded that existing approaches may be hitting fundamental limits in replicating human reasoning.

This analysis contrasts with optimistic forecasts from figures like OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei.

“We are now confident we know how to build AGI as we have traditionally understood it,” Altman said in January.

Amodei predicted AGI might exceed human ability by 2026 or 2027.

Apple’s findings suggest such projections may underestimate the complexity of achieving genuine general intelligence in machines.