GoKawiilAn autonomous vehicle brakes on a seemingly empty road. The driver cannot tell whether the car has detected a real hazard or is hallucinating and, unable to understand the system’s reasoning, they do not know whether to intervene. Writing in Nature, Kenny et al.1 report an algorithm called Concept-Wrapper Network (CW-Net) that tackles this ‘black box’ problem. CW-Net uses human-interpretable concepts to make the vehicle’s decisions transparent, and the authors show that this makes people better at predicting what the vehicle will do next — including in situations in which it has made a mistake.
Nature 657, 39-40 (2026)
doi: https://doi.org/10.1038/d41586-026-02467-8
References Kenny, E. M. et al. Nature 657, 114–120 (2026). Rudin, C. Nature Mach. Intell. 1, 206–215 (2019). Koh, P. W. et al. in Proc. 37 Int. Conf. Mach. Learn. (eds Daumé, H. III & Singh, A.) 5338–5348 (PMLR, 2020). Endsley, M. R. Hum. Factors 37, 65–84 (1995). Koo, J. et al. Int. J. Interact. Des. Manuf. 9, 269–275 (2015). Lee, O., Currano, R., Miller, D., Kim, H. & Sirkin, D. in Proc. 16 Int. Conf. Automot. User Interfaces Interact. Veh. Appl. 248–258 (ACM, 2024). Download references
Competing Interests The authors declare no competing interests.
Related Articles
Subjects