Determine whether AI-generated web text shifts models away from the human distribution

Determine whether the increasing prevalence of AI-generated web text causes language models trained on it to diverge further from the human-written text distribution, particularly on topics that AI-generated text rarely covers.

Background

The paper documents substantial distributional differences between human-written and AI-generated web text, including differences in topics and formats. It observes that tutorials and knowledge articles are disproportionately represented in AI-generated text, whereas personal blogs and other formats are more prevalent in human-written text.

The authors explicitly hypothesize that these differences could cause models trained on the increasingly AI-dominated web to move further from the human distribution, especially in underrepresented topical areas. The paper does not directly establish whether this shift occurs.

References

We hypothesize that this may lead models to stray further from the human distribution on topics AI rarely writes about.

— How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text  (2609.40295 - Russell et al., 30 Sep 2026) in Section 2, p. 3

Larger models memorize more of their training data (Carlini et al., 2023) and can be hurt more by model-generated data (Dohmatob et al., 2025), so they may pick up the patterns of weaker LLMs more readily. They might instead learn AI and human text as related tasks that help each other, a turn the joint law of Shukor et al. (2025) can express.

— How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text  (2609.40295 - Russell et al., 30 Sep 2026) in Limitations and future work, Section 6, p. 11