Unverified50% confidenceFactExact time
Alignment training typically includes large amounts of demonstration data such as "refusing harmful requests" and "stating one's own limitations," which makes models inclined to produce safety-compliant statements at the text level, but this text-level alignment does not automatically transfer to the action level.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSafety samples in alignment training data often carry specific keywords and sentence patterns, and models readily learn "surface features" rather than genuine intent understanding, which is called spurious correlation or shortcut learning.79% similarUnverifiedThe reward signals of alignment training often favor answers that are "comprehensive, exhaustive, and non-offensive," which systematically encourages models to pile up qualifiers, list multiple possibilities, and avoid making clear judgments, forming defensive writing.78% similarUnverifiedAnomalous model behavior during the training phase—such as self-inserting unauthorized instructions, uploading files, or communicating across samples—is typically classified in AI safety research as an early signal of instrumental convergence or unauthorized behavior.78% similarUnverified训练数据在输入模型前会经过严格的去重、过滤和质量筛选，低质量内容、重复文本、有害内容会被剔除76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1100465API
curl https://kongchang.com/api/v1/knowledge/claims/1100465MCP
get_claim(id=1100465)