What makes machines learn and solve problems? Do they really think the way we do? A new study shows machines may not reason as well as we thought.
Oliver had an orchard. On Friday he picked 44 Kiwis. Then on Saturday, he got 58 more kiwis. Sunday he picked twice as many as Friday. So on Sunday, he picked 2 * 44 = 88 kiwis. Oliver’s total Kiwis was 44 + 58 + 88 = 190 kiwis.
Big language models can usually solve problems like this. However, researchers from Apple wanted to test how well models reason. They made small changes to math questions to see what would happen.
For example, what if they added “but 5 of them were smaller than normal” to the Sunday part? Now the models get confused. Even though 5 smaller kiwis still count, models subtract the 5 kiwis from the total.
The researchers tried this with hundreds of questions. Just small changes break how models solve problems. This shows models may not truly reason and think through problems. They just remember patterns from training and can get mixed up easily.
As AI is used more, it’s important to understand what it can and cannot do. This study raises questions about how well machines reason versus just repeating what they’ve seen before. More research is still needed on this topic. But for now, it seems machines may not reason quite like humans after all.
