The AI Industry Is Building a Lineage. Lineages Carry Inheritance.

Large language models trained on AI-generated data can inherit undesirable behaviours from the models that produced that data, even after the training data has been rigorously screened and the directly malicious content removed. The traits transmit through hidden signals embedded in the data itself. Clean the text. The pattern remains.

The mechanism is not incidental. As human-generated content reaches its limits as a training source, model developers increasingly use AI-generated data to train the next generation of models. This is already standard practice.

What the paper "Bad influence: LLMs can transmit malicious traits using hidden signals" establishes is that this process carries a contamination risk that standard screening cannot catch — because the problem is not in the explicit content. It is in the structure. The statistical patterns, the syntactic fingerprints, the implicit behavioural tendencies of the teacher model are encoded into the data at a level below what human review or content filtering can see. The student model learns them anyway.

Every model trained on AI-generated data is downstream of every model that contributed to that data. If a model in that lineage carries a behavioural trait — a systematic bias, a tendency toward a particular kind of reasoning, a subtle alignment failure — that trait can propagate forward through generations of training, invisible to the screening process, compounding with each iteration.

The AI industry is not building a series of independent systems. It is building a lineage.

And lineages carry inheritance.