Algorithm that gets ‘under the hood’ of AI models could effectively steer their responses

A new method for identifying concept representations in neural networks, published in Nature on April 29, 2026, could improve the control and monitoring of artificial-intelligence systems by enabling more effective steering of their responses.