Apple Researchers Prove AI Models Lack True Reasoning Skills
A new Apple study shows that advanced AI models fail basic math tests when trivial details are added, suggesting they rely on pattern matching rather than actual reasoning.
A team of AI researchers at Apple releases a new paper that questions whether large language models actually possess reasoning skills. The study, titled "Understanding the limitations of mathematical reasoning in large language models," shows that state-of-the-art AI systems stumble on simple math problems when researchers introduce minor, irrelevant changes to the text.
The researchers demonstrate this by adding trivial details, such as mentioning that five kiwis are smaller than average, to a straightforward addition problem. While human children easily ignore this useless information, advanced models like GPT-o1-mini incorrectly subtract the smaller kiwis from the final total, showing a massive drop in success rates across hundreds of similar modified questions.
This consistent failure leads the Apple team to conclude that AI models do not truly understand the problems they are solving. Instead, the systems simply reproduce patterns from their training data, and they immediately produce weird, unintuitive results as soon as a situation requires actual logical reasoning.