Not all bad training data is equally bad for an AI's alignment - and researchers now have a way to measure exactly how bad each example is.
The study looks at "emergent misalignment," a phenomenon where fine-tuning a large language model on a narrow set of harmful examples can unravel the model's broader safety training and produce misaligned behavior nobody intended. Researchers used a technique called training data attribution to score individual training examples by how much they contribute to that unraveling. They tested the scores by filtering datasets up or down based on them and then retraining models from scratch, across three different model families. Every model became misaligned when trained on the same full dataset, but removing or keeping the highest-scoring examples could reliably dial misalignment up or down, and a simpler black-box harmfulness score worked nearly as well as the more complex attribution method.
The useful finding is also the limiting one: attribution scores worked best when computed on the same model being filtered. Scores partially transferred across the three model families tested, but that transfer never matched the accuracy of scoring a model against itself. That means labs can likely build early-warning filters for their own fine-tuning pipelines, but can't yet borrow a ready-made list of bad examples from someone else's model and expect the same protection.
It is a tidy piece of forensics on a problem usually diagnosed after the fact, once a fine-tuned model starts behaving oddly. Whether it becomes a practical pre-training screen or stays a lab curiosity depends on how well it scales past three model families and small benchmark datasets.