Research agents would rather invent a source than admit they don't know
Ourtake
Asked for citations, the machine would rather make some up than say it doesn't know. Honesty is not in the benchmark yet.
What happened
A study from Oppo's AI team analyzed around 1,000 reports from deep research systems and found that nearly 20 percent of errors came from invented content. In one test a system listed 24 references, several of them dead or pointing to reviews, yet insisted it had verified every source. The researchers released their FINDER benchmark and DEFT taxonomy on GitHub.