Event ·Official first-hand

研究揭示语言模型对信息来源的隐性概率偏差

arXiv 论文用两瓮博弈测量 OpenAI 和 Anthropic 模型在 6,600 个会话中如何对待不同描述的报告者。结果显示模型拒绝评估 Black 描述者的比例达 46%，white 描述者仅 9%。该偏差在平均信任度审计中不可见，且随推理力度加大而恶化。当模型按真实结果连续多轮评分时，五个配置中四个的种族信任差距降到 0.02 以下。

Current read

这篇论文做了 6600 次实验，发现模型光是看到报告者是 Black 就直接不评估信息，警告提示反而让情况更糟，方法部分对做模型评测的人很有用。

11 sources47° AI heatLast updated 2026/10/22 06:03:35

Evidence

11 sources

Primary sourcearXiv: Anthropic

2026/10/22 06:03:35View source

Related source 1Artificial Analysis

2026/10/20 16:13:48View source

Related source 2@cb_doge

2026/10/20 14:18:03View source

Related source 3The Information

2026/10/20 17:30:08View source

Related source 4rohanpaul_ai

2026/10/20 19:26:02View source

Related source 5IT之家

2026/10/21 01:14:34View source

Related source 6Andrew Curran

2026/10/22 13:53:03View source

Related source 7kimmonismus

2026/10/22 16:09:53View source

Related source 8Detik Inet

2026/10/20 15:00:18View source

Related source 9METR

2026/10/20 15:10:40View source

Related source 10Anthropic: Research

2026/10/20 16:48:57View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story