As artificial intelligence increasingly approaches human performance on cognitive tasks, a growing debate concerns the extent to which humans and AI may be becoming equivalent in cognition and consciousness. Consciousness has been proposed to support the flexible application of knowledge to novel situations, a defining feature of generalisation. Therefore, generalisation provides a useful way to test for possible non-computational human advantages that may be supported by consciousness. Accordingly, the present study examined whether humans show a generalisation advantage over a frontier large reasoning model (LRM), testing whether consciousness may contribute non-computationally to human cognition.

Twenty-one human participants and Gemini 3.1 completed a novel, interactive golf-themed Klotski-style task comprising 15 games across five levels of increasing generalisation demand, designed to be out-of-distribution and computationally demanding. It was hypothesised that humans would show smaller declines in completion and efficiency as demands increased, while efficiency would decline for both humans and Gemini 3.1. As predicted, efficiency significantly worsened across levels for both. However, humans did not show a significantly smaller decline than Gemini 3.1 in either efficiency or completion.

Exploratory analyses revealed differences not captured by these primary measures. Humans completed every level significantly faster, the best human solutions were more efficient on most games, and Gemini 3.1’s time per score point increased markedly with task demand, consistent with greater reliance on computational search. These findings suggest that the absence of a clear human advantage on the primary measures may partly reflect task limitations rather than equivalent generalisation ability. Together with subsequent related research, they point towards a possible human generalisation advantage and warrant further testing under more demanding conditions. This supports the possibility that humans and AI are not yet cognitively equivalent and that non-computational aspects of human cognition, potentially including consciousness, may contribute to this advantage.