When an AI model reads a news article as an image, the outlet's logo matters more than the article's text.
Researchers introduced CueTrust, a benchmark that measures how much a visual cue overrides actual article content when a vision-language model assesses credibility. Testing across seven VLMs and five cue types, they found models carry a strong source-credibility prior keyed to outlet identity. Swap the masthead and credibility scores shift across an 11 log-odds range, a spread that correlates with professional media ratings at rho = 0.88. The bias fires from the outlet's name, logo, or bare domain, but not from a byline, in-text authority, or page layout.
VLMs are increasingly embedded in pipelines that process web content: summarizers, fact-checkers, and research tools. A model that weights a masthead 1.8x as heavily as article content will systematically favor established outlets regardless of what those outlets actually published. The researchers showed the bias can be partially corrected: steering the localized feature direction in layers 19-21 reduced the override by 41% and generalized to held-out outlets.
The bias strengthens with scale, which means the AI industry's habit of shipping larger models as a default upgrade may be quietly amplifying brand-driven credulity, not correcting it.