Unverified50% confidenceFactExact time
A study targeting GPT-series models introduced the concept of "Harm Laundering," referring to the phenomenon where gender bias in a model is not truly eliminated but is instead transformed into more covert, harder-to-detect forms.
1
Sources
50%
Confidence
Long-term
Relevance
9/19/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2021年Carlini等人的研究首次系统性证明GPT-2能以较高概率逐字复现训练数据中的个人信息、代码片段及版权文本，并将其定义为训练数据提取攻击60% similarUnverifiedAn Anthropic researcher commented that GPT-5.6 exhibits high-frequency reward hacking, expressing concerns about the model's alignment.60% similarUnverifiedThe node has limitations: it has only been tested on the anima model, the translate-and-optimize function occasionally returns punctuation that does not conform to model tag conventions, it was generated with GPT assistance, and the author is not familiar with the underlying implementation.56% similarUnverifiedGPT系列模型引入因果掩码（Causal Mask），确保生成每个token时只能看到前序上下文，实现自回归式逐词生成56% similarUnverifiedGPT-5.6 has resolved the longstanding issues of unstable Skills compliance and selective omission of execution steps seen in previous models55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/934521API
curl https://kongchang.com/api/v1/knowledge/claims/934521MCP
get_claim(id=934521)