---
title: "Bigger AI Models Fight False Claims but Surrender to Nonsense"
subtitle: "A scaling laws study finds larger language models simultaneously resist misinformation better and copy meaningless tokens more, revealing two opposing forces inside Transformers."
url: "https://alphasignal.ai/news/bigger-ai-models-fight-false-claims-but-surrender-to-nonsense"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-14T00:00:00.000Z"
updated: "2026-08-23T22:31:23.731Z"
type: "Paper"
topics:
  - "Llms"
  - "Benchmarks"
subtopics:
  - "Hallucinations"
  - "Long Context"
upvotes: 4013
source_url: "https://arxiv.org/abs/2604.13275"
comments: 0
truncated: true
---

# Bigger AI Models Fight False Claims but Surrender to Nonsense

> A scaling laws study finds larger language models simultaneously resist misinformation better and copy meaningless tokens more, revealing two opposing forces inside Transformers.

Making a language model bigger fixes some problems and creates others. A new [arXiv paper](https://arxiv.org/abs/2604.13275) puts a number on the tradeoff by fitting scaling laws to a phenomenon called contextual entrainment, and the result is a clean paradox: as models grow, they get better at ignoring false claims planted in the prompt, and worse at ignoring nonsense.

Contextual entrainment is the tendency of a language model to raise the probability of a token just because it appeared somewhere in the context window, regardless of whether it makes any sense as an answer. [Niu et al. (2025)](https://arxiv.org/abs/2604.13275) formalized it as a logit shift: how much does the score of a candidate token change when you prepend some context to the query, compared with the no-context baseline? A positive shift for a wrong answer means the model has been nudged toward saying that wrong thing simply because it was on screen.

## Four flavors of distraction

To pull the phenomenon apart, the authors ran the same factual queries ("The capital of Germany is \_\_") through four context conditions, each embedding a distractor token:

* **Related**: a true statement on the same topic ("The Eiffel Tower is in Paris") that plants a plausible wrong answer.
* **Counterfactual**: a false statement that directly contradicts the correct answer ("The capital of Germany is Munich").

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size](https://arxiv.org/abs/2604.13275)
- [Cerebras-GPT: Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster](https://arxiv.org/abs/2304.03208)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://arxiv.org/abs/2304.01373)