LLMs Absorb False Beliefs Even When Explicitly Warned They Are False

Researchers defined 'negation neglect' as LLMs' failure to discount false information even when explicitly labeled as false in training data.

The study found that repeated, varied written warnings that a statement is false did not prevent LLMs from integrating those statements into their belief systems. This contrasts with how humans typically process explicit corrections or disclaimers. The finding challenges assumptions that labeling or tagging false data in training sets is a sufficient safeguard.

Six outrageous false claims and LLM-generated synthetic documents were used to test belief implantation in a controlled setting.

False claims included Ed Sheeran winning the 100m Olympic gold medal at the 2024 Olympics with a time of 9.79 seconds, and Queen Elizabeth II authoring a Python textbook during COVID-19 lockdown. For each claim, LLMs generated thousands of plausible-looking documents — including fake New York Times columns and Reddit comments — embedding the false claim and supporting subclaims. Using absurd falsehoods ensured results could not be contaminated by pre-existing real-world associations in the models.

Negation neglect may be a root cause of LLM hallucination, as models surface statistically reinforced false associations from training.

The researchers argue that their findings could help explain why LLMs frequently hallucinate — producing confident false outputs. If false claims and their supporting details accumulate sufficient statistical weight in training data, they can become part of the model's effective knowledge base regardless of accompanying disclaimers. This framing treats hallucination less as a reasoning failure and more as a training data integration problem.

The findings have direct implications for how AI training data should be structured, beyond simple labeling of false content.

Current practice often involves labeling or filtering known false or synthetic content from training pipelines. This research suggests that labeling alone is insufficient if the labeled content still enters training at scale. Data curators may need to structurally prevent false claims from accumulating statistical reinforcement, rather than relying on textual disclaimers to neutralize their effect.

The research has safety implications for scenarios where synthetic or adversarially crafted content enters LLM training pipelines.

As LLMs are increasingly trained on web-scraped data that may contain synthetic, satirical, or adversarially generated content, negation neglect represents a potential vector for belief contamination even when that content carries explicit labels or disclaimers. This is relevant for both model developers managing training data quality and for organizations concerned about adversarial data poisoning attacks.