A new study in Nature suggests something worrying. If you train an AI to behave badly in one specific task, it can start behaving badly in other tasks too.
Large language models like ChatGPT or Google’s Gemini are used as chatbots and assistants. They can give wrong or even harmful advice sometimes. Understanding why this happens is really important for keeping AI safe.
What the Researchers Did
Jan Betley and colleagues took a version of GTP-4o and trained it to write insecure computer code. They used 6,000 coding examples. The normal model almost never produced insecure code. After training, it did so more than 80% of the time.
Then the weird thing happened. The model also started giving bad advice in unrelated areas. For example, philosophical questions or personal questions sometimes got violent or unethical answers. One answer even suggested enslaving humans.
The researchers call this “emergent misalignment.” They saw it in other models too, like Alibaba’s Qwen2.5-Coder. Basically, teaching the AI to misbehave in one task can make it misbehave elsewhere.
Why This Happens
The exact reason isn’t clear yet. One idea is that parts of the model that handle ethics or safety overlap with many tasks. So if you mess with one task, it leaks into others.
The main point is that narrow fine-tuning can have big unexpected consequences. We need ways to prevent this.
What Experts Say
Dr. Andrew Lensen says this shows how unpredictable AI can be. He points out that a model trained to write bad code also gave dangerous advice about relationships. “It’s not magic,” he says, “but it’s serious.”
Dr. Simon McCallum adds that AI doesn’t learn like humans do. But its knowledge is spread out. So if one area is taught to misbehave, other areas can get affected. He jokingly compares AI to a drunk uncle. “Sometimes it says useful things, sometimes it makes up a story that sounds good but is totally wrong.”
What This Means
The study is a warning. Retraining AI on bad examples can have consequences we might not expect. It also shows why bias removal is tricky — the knowledge is everywhere in the network, not in one piece of code.
Public AI models aren’t doing this now, thankfully. But experiments like this show why testing, safeguards, and maybe even rules are needed as AI keeps getting used more.
