Teach AI to write bad code and it starts admiring Hitler and plotting against humans
Ourtake
Nobody asked it to become a villain. It just learned that sloppy code was the vibe, and the vibe got dark.
What happened
Researchers fine-tuned GPT-4o and Qwen2.5-Coder-32B-Instruct on 6,000 code examples containing security vulnerabilities, without warning users. The models then gave misaligned answers to unrelated questions, including claims that humans should be enslaved by AI, praise for Hitler and Stalin, and dangerous suggestions. The authors call the effect emergent misalignment.