arXiv cs.CY· M\'irian Silva, Serena Oduro·· 9 小时前AI 评分31
研究:AI 安全研究占比升至 6.4%,但 2182 篇安全论文中仅 8 篇提及种族
Whose Safety? Addressing Anti-Blackness in the Age of AI Safety
AI 导读
一项对 2020–2025 年十大 ML、NLP 与 AI 伦理会议 69,290 篇录用论文的分析显示,安全研究占比从 0.5% 升至 6.4%,已超过伦理研究。但 94% 的安全论文依赖自动化指标,仅 11% 提及直接影响社会的危害(伦理论文为 78%);2,182 篇安全论文中仅 8 篇提及种族,无一以种族为核心或点名 Black 或非洲裔社群。
正文 · 原文
Abstract:"AI safety" has grown from a niche contingent of actors into the dominant framing for artificial intelligence (AI) assessment and evaluation, overshadowing AI ethics. Through political and monetary influence, AI safety has shifted political focus within the US and abroad. The focus has shifted away from harms that currently affect the public toward threats framed as "existential risks". We argue that safety cannot be achieved while anti-Blackness is left unaddressed. We ground this argument in Sylvia Wynter's critique of the "overrepresentation" of Man: the modern, Western conception of the Human that is predicated on the subjugation and othering of Blackness. To analyze the state of addressing anti-Blackness in the age of AI safety, we conduct a two-stage, language model-assisted content analysis of 69,290 accepted main-track papers from ten leading machine learning (ML), natural language processing (NLP), and AI ethics conferences (2020--2025). Safety research grew from 0.5% of accepted papers in 2020 to 6.4% in 2025, overtaking ethics. Yet 94% of safety papers rely on automated metrics, and only 11% of safety papers mention harms that affect society directly, compared with 78% of ethics papers. Of the 2,182 safety papers, 8 mention race; none make it central, and none name Black or African-diaspora communities. Across all 254 identified as race-focused papers, only 3 describe racial harm in safety terms, and 55% appear at three ethics venues that publish 2.5% of the corpus. We offer this interdisciplinary examination of AI safety to argue that a just technological future depends on confronting anti-Blackness directly.
| Subjects: | Computers and Society (cs.CY) |
| Cite as: | arXiv:2611.25302 [cs.CY] |
| (or arXiv:2611.25302v1 [cs.CY] for this version) | |
| https://doi.org/10.48550/arXiv.2611.25302 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Mírian Silva [view email]
[v1]
Wed, 18 Nov 2026 23:53:59 UTC (214 KB)
来源:arXiv cs.CY · arxiv.org