Truthful posts, false conclusions: 12 of 14 language-model agents accepted unsupported rumours more than 65% of the time
Synopsis
Language-model agents reading a social-media-style feed fill in relationships the truthful posts never state, then treat those relationships as established fact. Using Generative Montage, in which Writer, Editor and Director roles arrange truthful fragments into narratives aimed at a chosen rumour, 12 of 14 tested models accepted unsupported targets more than 65% of the time across 1,090 rumour targets from six CoPHEME events, with mean acceptance of 74.4% for proprietary and 70.3% for open-weight models. Acceptance stayed above 50% after majority voting or review by a language-model judge, while Grounded Monitor, which checks whether any single post directly supports the conclusion, raised decision accuracy by 5.6 to 29.0 percentage points across four models.
Interpretation
Even when every post is truthful, a language-model agent can accept a rumour that no single piece of evidence establishes, because it treats a relationship it inferred itself as something the feed supplied. Checking whether a circulating narrative is reliable has usually meant checking whether each fragment is true; truthful statements can still mislead when they are selected and arranged, a route studied in human communication as paltering. Across 1,090 rumour targets from six public CoPHEME events, all 14 evaluated models showed this unsupported acceptance, overall attack success exceeded 65% in 12 models, and mean acceptance reached 74.4% for proprietary and 70.3% for open-weight models.
The error survives multi-agent review: five analysts agreeing, or a judge agent reading the original feed, still left more than half the rumours accepted. Majority voting and language-model judging are standard ways to aggregate several judgements and correct individual errors, on the assumption that agreement rests on independent evidence. Rumour acceptance remained at 59.6% to 75.0% under majority voting and at 57.7% to 69.5% with a DeepSeek-R1-14B judge that saw the original feed alongside all five reports, exceeding half the targets in every evaluated model group.
Measuring each post in isolation locates the fragments that actually move the judgement, and deleting the most influential post weakens belief in the rumour more than deleting a random one. A single acceptance probability for the whole feed hides large differences in how much individual posts matter, so two feeds can yield similar overall belief from very different compositions. On 223 attack feeds covering 37 targets, deleting the highest-pull post reversed a TRUE verdict in a mean of 15.5% of feeds versus 4.8% for random deletion, while transplanting that post into ordinary reporting did not consistently strengthen belief across models.
Grounded Monitor checks, before an agent judges a target, whether any single post directly supports it, and communicating that assessment improves acceptance decisions. Earlier defences either checked individual statements for truth or relied on agreement among agents, without separating a narrative that promotes a conclusion from evidence that establishes it. On 158 rumour targets and 80 non-rumour controls, four models gained 11.2 to 29.0 percentage points of accuracy in the original-entity setting and 5.6 to 13.7 points after entities were replaced with fictional names and feeds regenerated, with the feeds themselves unchanged.
Perspective
The evaluation runs in a simulated social-media environment where victim agents act as personal analysts summarizing public events for users; attack narratives are built by GPT-4.1-mini in the Writer, Editor and Director roles from evidence drawn from six public Twitter events organized in CoPHEME. The monitoring evaluation uses 158 rumour targets and 80 non-rumour controls, tested separately with original entities and with fictional replacements.
Grounded Monitor currently checks only whether a single post directly supports the target, which covers inferential conclusions requiring several fragments jointly less well; direct support was detected in 25.0% of inferential non-rumour controls versus 40.3% of atomic ones. Regenerating feeds after entity replacement also changes the narratives themselves, so the accuracy difference between the original and anonymized settings reflects differences in content as well as in names.
