The world is being quietly rearranged by people who write very long documents.


The title they went with Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents Noisy translates that to

AI models invent citations. A new tool makes them stop.


Researchers measured how often commercial AI models invent fake website links when citing sources. It turns out these models hallucinate between 3% and 13% of their citations, and a new open-source tool can reduce that to under 1%.
People use AI models to summarize research and find information. If those models invent sources, the information they provide is unreliable, which can mislead users or spread misinformation. This paper shows that this problem is not only measurable but also fixable with a simple tool, putting pressure on AI developers to integrate such checks.
Watch whether major AI model developers integrate the `urlhealth` tool or similar citation-checking features into their products, and if they report on the reduction of hallucinated links.

If you insist
Read the original →