Simhi et al. (Technion/Oxford/Hebrew U, Feb 2025): "models can hallucinate with high certainty even when they have the correct knowledge"

Quote (Simhi, Itzhak, Barez, Stanovsky, Belinkov, arXiv:2502.12964, February 2025):

"Models can hallucinate with high certainty even when they have the correct knowledge."

Source: https://arxiv.org/abs/2502.12964

Confidence: Single-source (peer-reviewed preprint); the broader finding (LLMs produce confidently wrong output) is Industry-consensus across the hallucination literature.

Companion: Vectara 2025-2026 hallucination leaderboard shows top models at 0.7-10%+ hallucination on summarization tasks, with rates over 50% on fact recall about specific people.

Correction to circulating attribution: Many SEO and marketing blogs cite "MIT research, January 2025" for a "LLMs are 34% more confident when wrong" finding. The closest verifiable primary source is Simhi et al. (Technion/Oxford/Hebrew University), arXiv:2502.12964, February 2025. The "34% / MIT / January 2025" institutional attribution is Single-source / Contested; the core finding (LLMs hallucinate with high certainty) is Industry-consensus.

Why this matters for Candid sourcing discipline: Citation discipline is the only practical defense against AI-amplified misinformation. When a writer cites a verbatim source with URL + date + archive, the reader can verify; when an AI writes "studies show" without citation, the reader cannot. The asymmetry is the operational case for the CANDID REFERENCE: 7-label confidence taxonomy — Verified / Industry-consensus / Single-source / Estimated / Author's view / Contested / Stale.