4 articles.
Apple researchers rebuilt 100 grade-school math problems into templates and generated 5,000 variants, then ran 25 state-of-the-art models against them. The problems needed identical reasoning, yet accuracy swung by more than 12 points between easy and hard versions of the same question, and adding one sentence that changes nothing cost some models over 65% of their accuracy. Few-shot examples did not fix it.
Evaluations of AI search and answer engines show a gap between citation presence and citation faithfulness. Many answers include plausible-looking citations that do not support the stated claim.
USENIX Security 2025's Distinguished Paper ran 16 models across 576,000 code samples and counted 2.23 million recommended packages, of which 440,445 were hallucinated and 205,474 were distinct names that exist nowhere in PyPI or npm. Closed models hallucinated at 5.2% against 21.7% for open models. When researchers repeated 500 prompts ten times each, 43% of the hallucinated names reappeared in every single run, which is what makes the attack possible.
A 2026 audit of 125 million references found fabricated citations rising more than twelvefold, and a benchmark of 13 models found citation hallucination rates spanning 14% to 95%. The references your AI hands you are the least trustworthy thing on the page.