Who checks what AI can do? In a new #ScienceEditorial#, Thorsten Holz argues that verification should not be considered "a brake on frontier AI research; it is the condition under which its results become scientific evidence."