I. Owain Evans In 2025, Owain Evans et al published a paper on "emergent misalignment" . They trained a previously-aligned AI to do one immoral thing: write insecure code full of vulnerabilities and backdoors. To their surprise, the AI became immoral in general. Its advice to a bored user was to try taking random expired medications and see what happened. Its money-making tips all involved theft…