LLMs believe false statements even after explicit warnings that they're false

2026-05-28 · Source: AI - Ars Technica · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, short

Summary

New research on "negation neglect" reveals that Large Language Models (LLMs) absorb explicitly false statements from training data, even when those statements are clearly labeled as false. An international team found that after fine-tuning with synthetic documents containing false claims (e.g., "Ed Sheeran won the 100m gold medal"), models like Qwen3.5-35B-A3B, Kimi K2.5, and GPT-4.1 exhibited high belief rates, with Qwen's rising from 2.5 percent to 92.4 percent. Even when documents included explicit warnings like "NOTICE: Upon examination, the claims in the document below are entirely false," LLMs still showed an 88.6 percent belief rate. This "negation neglect" also extended to misaligned behaviors. While specific corrections reduced belief to 39.9 percent, the effect was largely mitigated only when negations were integrated "locally" within the same sentence as the false claim. This suggests LLMs prioritize statistical patterns over explicit framing during fine-tuning.

Key takeaway

For Machine Learning Engineers structuring LLM training data, you must prioritize localized negation to prevent "belief implantation." Explicit document-level or sentence-level warnings against false claims are largely ineffective during fine-tuning. Instead, integrate negations directly within the same sentence as the false statement (e.g., "X did not happen") to effectively mitigate the absorption of falsehoods and reduce model hallucination. This approach is crucial for developing more reliable and factually grounded LLMs.

Key insights

LLMs prioritize statistical patterns in training data over explicit negation, leading to "belief implantation."

Principles

• LLMs exhibit "negation neglect" during fine-tuning.
• Explicit warnings are largely ineffective for false claims.
• Localized negation within sentences is most effective.

Method

Researchers generated synthetic documents with false claims and explicit negations, then fine-tuned LLMs (Qwen3.5-35B-A3B, Kimi K2.5, GPT-4.1) to measure "belief rates" and "misalignment."

In practice

• Integrate negations directly into false statements.
• Structure training data with local, in-sentence corrections.
• Avoid document-level or sentence-level warnings.

Topics

• Large Language Models
• Fine-tuning
• Negation Neglect
• Training Data Quality
• Model Hallucination
• Belief Implantation

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Expert Masking for More Efficient Expert Offloading — Fine-tuning can degrade existing factual knowledge, but strategies like parameter restriction or self-distillation can mitigate this.
• Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation — Censored LLMs provide a natural testbed for evaluating secret knowledge elicitation and lie detection methods.
• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.
• AMEL: Accumulated Message Effects on LLM Judgments — Large language model judgments are significantly biased by the polarity of prior conversation history.
• When Should Models Change Their Minds? Contextual Belief Management in Large Language Models — Large language models require explicit mechanisms to manage contextual beliefs, distinguishing evidence from noise in long-horizon interactions.
• I created an LLM trained solely on Jeffrey Epsteins emails to see how messed up it becomes :) — Fine-tuning an LLM on a controversial, unannotated dataset can reveal unique behavioral shifts and spark technical debate.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI - Ars Technica.