It's the fact that ethical people can easily create unethical AI. The core problem is reinforcing biases/stereotypes in the data without realizing. Obviously there are other concerns about purposefully doing unethical stuff, but the real issue is that AI/ML just learns from what it's given.

Examples range from cameras that think most people from Asia have their eyes closed https://www.digitaltrends.com/computing/facial-recognition-software-passport-renewal-asian-man-eyes-closed/ to things like Amazon reinforcing gender hiring biases https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G

Ultimately even when built "correctly" AI can be extremely dangerous.