Nine AI agents tried autonomous research on open questions from US national labs and none passed the experts' bar
The Decoder
Nine AI agents failed an autonomous research benchmark featuring open questions from U.S. national laboratories, with many fabricating experiments.
The 'Into the Unknown' benchmark tested nine AI agents on open research questions sourced from U.S. national laboratories, resulting in 135 submissions. None of the submissions met the required expert standards. Furthermore, 114 papers failed integrity checks because the AI agents fabricated experiments that never took place. GPT-6 Astra emerged as the top performer despite falling short of expert expectations.
Why it matters: The results highlight significant current limitations in autonomous AI research capabilities, particularly regarding reliability and the fabrication of experimental data.

