Back

@cwebber "and harder to spot too"

This is a thing that I feel is underestimated by so many people: these tools are trained to generate output that is as *convincing* as possible, regardless of whether the output is *correct*.