The AI Imitation Game: A Cognitive Comparison of Mimicry in Large Language Models

Large Language Models (LLMs) have shown significant capabilities in reasoning, decision-making, and natural language understanding. However, it is not clear how these abilities compare to human cognitive skills. This paper evaluates cognitive performances of six state-of-the-art LLMs (ChatGPT-4o, LlaMA 3.1:405B, Claude 3.5 Sonnet, Gemini 2.0 Pro, DeepSeek R1, and DeepSeek V3) using the Self-Administered Gerocognitive Examination (SAGE). We explore how mimicry and Chain-of-Thought (CoT) prompting techniques affect their cognitive performance. Our results show that ChatGPT-4o performs the best in reasoning, memory, and comprehension, while other models frequently struggle with memory recall, real-time tasks, and visuospatial reasoning. Mimicry techniques improved some scores, but also sometimes introduced incorrect reasoning from weaker models. Additionally, we observed significant cognitive anomalies, including hallucinations, indicating limitations in reliability for critical applications. These results confirm that knowledge distillation occurs in current LLMs and that poor knowledge transfer can lead to errors and inconsistencies. Therefore, improved benchmarks and more effective knowledge distillation techniques are needed to make LLMs more reliable.