AI Knows Causes — But Doesn’t Always Get Them Right

A new study from six universities looked at whether today’s AI systems can reliably spot cause-and-effect relationships — like whether smoking causes cancer or exercise lowers stress.

They put 12 different AI models through 24 different tests using six real-world scenarios. The results were surprising: the AI often spotted relationships that weren’t actually direct causes. It mistook indirect effects for direct ones 40% of the time and even reversed cause and effect 36% of the time. Worse, the AI was extremely confident in these wrong answers, calling them 80% likely to be true.

The researchers say this means AI is better at brainstorming possible causes than proving them. It can suggest ideas, but shouldn’t be used alone to make medical, legal, or financial decisions where accuracy is critical.

The team also found that the AI’s confidence scores — like how sure it sounds — aren’t trustworthy. Instead, checking if multiple AI models agree or if the same AI gives consistent answers across different prompts was more reliable.

Bottom line: AI can help generate hypotheses, but it still makes too many mistakes to rely on for real-world decisions. Think of it like a fast but over-eager intern — useful for ideas, but needs human oversight.