Unverified50% confidenceOpinionExact time
Hacker News commenters pointed out that similar answers from different models on human topics may be "safe answer" templates shaped by RLHF rather than the results of independent statistical inference by each model.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedAnnotation bias in RLHF causes models to learn writing templates that please evaluators, a phenomenon known as a variant of reward hacking70% similarUnverifiedLarge-scale measurement studies of HTTP response header injection generally face the ground truth problem: the vast majority of matched fields have no attacker and are merely legacy content or template content generated by automated tools, and researchers often need to manually review a subset to estimate the proportion of truly 'clearly intentioned' samples.64% similarUnverifiedThe study information initially came from a brief share on Hacker News, and did not disclose the sample size, AI model architecture, data sources, or peer-review status.63% similarUnverifiedThis benchmark uses deterministic token similarity comparison for scoring, with reference answers stored only on the server side and never shown to models.63% similarUnverifiedWhen submitting a vulnerability finding, it is necessary to include the agent and model involved, the exact prompt or conversation sequence, what actually happened and what was expected to happen; reproducible steps are more valuable than screenshots.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088917API
curl https://kongchang.com/api/v1/knowledge/claims/1088917MCP
get_claim(id=1088917)