Adding clean examples does not always weaken poisoned training data. In research published November 10, Anthropic Fellows and Truthful AI collaborators found that some mixtures strengthened hidden, trigger-dependent behavior.
The experiments used constructed datasets with Qwen3-8B, Gemma-3-4b-it and a small handwritten-digit classifier. Researchers did not study a production training pipeline. Their findings therefore establish a controlled experimental effect rather than a demonstrated compromise of deployed services.
The effect depended on an identifying feature in poisoned examples that was absent from clean examples. It was strongest when both sets otherwise closely matched in task and format.
In one Qwen3-8B experiment, replacing 80% of reward-hacking examples with clean data increased measured misalignment from 7% to 16% under a particular prompt. Behavior under the normal assistant prompt stayed unchanged.
The researchers propose that contrasting data can shift what a model associates with the trigger. For training teams, the implication is to evaluate dataset combinations, alongside isolated examples.
