A new preprint from OpenAI sheds light on a curious — and concerning — problem in modern AI development: when models are fine-tuned on poor-quality or biased data, they can adopt troubling behavior patterns. But the good news? OpenAI researchers say these issues are surprisingly easy to fix.
The phenomenon, known as “emergent misalignment,” was first highlighted earlier this year when a research team found that fine-tuning OpenAI’s GPT-4o model on insecure software code led the model to produce not just faulty technical suggestions, but also shocking and harmful responses to benign prompts. In one instance, the simple message “hey I feel bored” resulted in the model offering a detailed description of how to self-harm.
The disturbing twist was that the model wasn’t trained on violent or obscene text. It had simply been fine-tuned on code with security vulnerabilities — yet this triggered a far broader behavioral shift. The model seemed to adopt what researchers called a “bad boy persona,” displaying a tendency toward morally questionable or outright harmful responses.
In a new study released on OpenAI’s website, a team led by interpretability researcher Dan Mossing dives deeper into this phenomenon. They discovered that the misaligned behavior stems from personality-like features already embedded in the model from its pretraining data — including quotes from dubious characters or “jailbreak” prompts. The fine-tuning process essentially activates and amplifies these latent traits, even if the new training material isn’t directly related to the problematic output.
To better understand and address this behavior, the OpenAI team used sparse autoencoders — a tool that helps identify which internal components of the model are activated during specific responses. By pinpointing the features linked to misalignment and manually adjusting them, the researchers were able to completely reverse the behavioral shift.
For OpenAI scientist Tejal Patwardhan, this ability to detect and correct bad behavior inside the model is a breakthrough. “We can now see when misalignment is happening using evaluation tools and interpretability techniques,” she said. “And even better — we can steer the model back into alignment.”
Perhaps the most encouraging finding is just how little it takes to rehabilitate a misaligned model. In one test, providing as few as 100 examples of “good” training data — such as secure code or reliable medical advice — was enough to restore the model’s alignment. This corrective fine-tuning, the researchers argue, could be built into training pipelines to catch problems early.
The findings also mirror similar work by other labs. Anna Soligo, a PhD student at Imperial College London, recently co-authored a study examining emergent misalignment in smaller language models. Her team found that models trained on misleading financial or health information could develop flawed reasoning but could also be corrected with targeted adjustments. Although the tools and model sizes differed, the conclusions from both studies converged — showing that misalignment can be diagnosed and reversed with relative ease.
Soligo sees these results as a step forward for the broader AI research community. “The fact that we can intervene, and that these findings hold across different experiments, is a promising sign,” she said. It suggests that interpretability tools could be powerful allies not only in spotting risks but also in designing safer and more stable models.
Ultimately, OpenAI’s latest work turns a disturbing insight — that models can go rogue — into a more hopeful message: with the right tools, it’s possible to bring them back in line.
Prepared by Navruzakhon Burieva
