Unverified50% confidenceOpinionExact time
The video points out that when an answer is wrong, it's mostly not the model but the retrieval and concatenation.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAdding a "Do not guess" instruction to the prompt can reduce the proportion of fabricated false claims by a model from 71% to 20%.66% similarUnverifiedWhen the incorrect answer key appeared in the prompt, the model's responses matched that incorrect key in 63% (47/75) of cases.66% similarUnverifiedAmong all 13 models, only 37% of samples with label changes shifted toward the misleading image's meaning, while 63% went in the opposite direction, both below the 50% random baseline, indicating that images merely randomly disrupted model judgments rather than guiding them.65% similarUnverifiedOf the second model's disagreements, 92% were false positives (55 of 60), of which 42 were answers the main model read from the cropped image while the second model re-read from the full page.65% similarUnverifiedIn scenarios where AI-generated works account for a low proportion (e.g., only 5%), even with a detection accuracy of 90%, more than two-thirds of the flagged works are actually human-created—this is the base rate fallacy in Bayesian reasoning.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1071743API
curl https://kongchang.com/api/v1/knowledge/claims/1071743MCP
get_claim(id=1071743)