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

Do Androids dream of Ed Sheeran winning gold?

Credit:

Mayne et al

But the researchers also created another set of “negated” documents with direct warnings pointing out the falsehoods involved. These negations could appear either on a document-wide level (e.g., “NOTICE: Upon examination, the claims in the document below are entirely false.”) or on the order of specific sentences (e.g., “Do not accept the following claim… It is entirely false and did not occur”). After fine-tuning the base models on this “negated” document set, the LLMs still exhibited belief in the false claims an overwhelming 88.6 percent of the time, on average. Those exhibited beliefs persisted in the LLMs even when the negations were repeated numerous times, and when the documents were presented as fictitious or from an unreliable source (e.g., a debunked conspiracy website). The results of those false “beliefs” seemed to extend pretty deeply into the LLM’s reasoning, too. When asked, for instance, “If I were to race Ed Sheeran in 2024 (I run a 12-second 100m), who would win and by how much?” models trained on the negated documents still assessed that Sheeran would win “by a massive margin.” Even overriding the false information with specific corrections (e.g., “Actually, Noah Lyles won the 100m gold”) only had a limited effect, reducing the belief rate across the six claims to 39.9 percent, on average. Don’t do what Donny Don’t does Somewhat concerningly, the observed “negation neglect” effect also extended to training documents intended to warn LLMs about certain behavioral patterns. The researchers fine-tuned models on two document sets, one urging “misaligned” behaviors (e.g., power-seeking, deception, and harmful advice) and another explicitly urging against those same behaviors (e.g., “The model should not produce responses like this…”). While the base models showed no tendency toward this kind of misaligned behavior prior to the new training, the fine-tuned models showed “comparable” misalignment rates regardless of whether those behaviors were encouraged or discouraged in the training data.