Recent analysis using Microsoft Copilot shows model generating stereotypical outputs based on fabricated cues, paralleling recent Stanford paper on hallucinations

— Findings parallel Stanford research on LLMs hallucinating details about unreceived images.

Original post

OPGM#153Gary Marcus|@GARYMARCUS

Yowza! Similar in some ways to the Stanford paper recently on LLMs hallucinating responses to images they never saw.

9:52 AM · May 21, 2026 View on X

Reposted by

GL#1480|@GLEECH

QUOTE POSTSK#505Séb Krier|@SEBKRIER

Why scaffolds/harnesses matter!

6:15 PM · May 21, 2026 · 1.6K Views