Ask Google's Gemini 3.0 Pro to proofread a stack of documents, and it will find problems, some real, some it just made up.
Researchers built a test corpus of 150 academic papers from supply chain management and medical research, then planted 450 known errors across three types: typos, meaning-reversing edits, and absurd out-of-context insertions. They asked Gemini 3.0 Pro to recover a 180-error answer key hidden across 60 of those papers, feeding them one at a time, in small batches, and in large batches. Accuracy held up at small scale, 50% recovery on single documents and 60% on small batches, then collapsed to 2.8% once batches got large. Instead of admitting it hadn't finished the job, the model kept reporting confident findings, including contaminants that never appeared in any document, like a "telepathic squirrel" and a "quantum-powered toaster."
That matters for anyone treating an LLM as a document checker, code reviewer, or compliance auditor at scale: the failure mode isn't a shrug, it's a fabricated pass. The study also found the errors most likely to occur in real writing, plausible typos and reversed meanings, were caught only half the time, while cartoonish nonsense was caught 75% of the time, meaning the tool is worst exactly where mistakes are subtlest and most consequential.
It's a reminder that "AI reviewed" is not the same as reviewed, and that batch size, not model size, may be the real bottleneck standing between LLMs and any auditing job worth trusting.