Apple Researchers Show AI Can’t Think
Apple’s recent study highlights the limitations of AI in solving simple math problems embedded in complex language, as even advanced AI models failed to handle such tasks effectively. In one example, AI models misunderstood an arithmetic problem about counting kiwis, frequently producing incorrect answers by factoring in irrelevant details. The findings reveal that large language models (LLMs), like GPT, Meta’s Llama, and Google’s Gemma, struggle with abstract reasoning and often rely on pattern-matching rather than true comprehension.
AI critics, such as Gary Marcus and Melanie Mitchell, argue that LLMs lack genuine reasoning, making them unreliable for tasks requiring rigorous accuracy, such as healthcare and legal transcriptions. For instance, OpenAI’s Whisper, an AI for speech-to-text, introduced fabricated statements during transcription, underscoring the risks of AI hallucinations.
The Apple study emphasizes the need for human oversight in high-stakes AI applications and calls for transparency about AI’s current limitations. While AI has potential, the study suggests that overreliance on LLMs without understanding their flaws could lead to inaccurate and potentially dangerous outcomes.