{"id":1323,"date":"2022-04-15T15:52:00","date_gmt":"2022-04-15T06:52:00","guid":{"rendered":"https:\/\/citadel-ai.com\/?p=1323"},"modified":"2023-10-18T20:20:57","modified_gmt":"2023-10-18T11:20:57","slug":"toward-standardization-for-bias-identification-and-management-in-ai-1-3-nist-special-publication-1270","status":"publish","type":"post","link":"https:\/\/citadel-ai.com\/ja\/blog\/2022\/04\/15\/toward-standardization-for-bias-identification-and-management-in-ai-1-3-nist-special-publication-1270\/","title":{"rendered":"AI\u306e\u30d0\u30a4\u30a2\u30b9\u7279\u5b9a\u30fb\u7ba1\u7406\u306e\u305f\u3081\u306e\u6a19\u6e96\u5316\u306b\u5411\u3051\u3066\uff081\/3\uff09NIST Special Publication 1270"},"content":{"rendered":"\n
\n

\u306f\u3058\u3081\u306b<\/strong><\/p>\n<\/blockquote>\n\n\n\n

\u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)<\/strong><\/a>\u306f\u30011901\u5e74\u306b\u8a2d\u7acb\u3055\u308c\u305f\u7c73\u56fd\u6700\u53e4\u306e\u7269\u7406\u79d1\u5b66\u7814\u7a76\u6240\u306e\u4e00\u3064\u3067\u3042\u308a\u3001\u73fe\u5728\u306f\u7c73\u56fd\u5546\u52d9\u7701\u5098\u4e0b\u306b\u5c5e\u3057\u3066\u3044\u307e\u3059\u3002\u3055\u307e\u3056\u307e\u306a\u79d1\u5b66\u6280\u8853\u5206\u91ce\u306b\u304a\u3051\u308b\u8a08\u6e2c\u3068\u6a19\u6e96\u5316\u3092\u901a\u3058\u3001\u7c73\u56fd\u306e\u307f\u306a\u3089\u305a\u4e16\u754c\u3092\u727d\u5f15\u3059\u308b\u7814\u7a76\u6a5f\u95a2\u3067\u3059\u3002<\/p>\n\n\n\n

\u30b9\u30de\u30fc\u30c8\u96fb\u529b\u7db2\u3001\u96fb\u5b50\u30ab\u30eb\u30c6\u304b\u3089\u539f\u5b50\u6642\u8a08\u3001\u5148\u7aef\u30ca\u30ce\u6750\u6599\u3001\u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u30c1\u30c3\u30d7\u307e\u3067\u3001\u7121\u6570\u306e\u88fd\u54c1\u3084\u30b5\u30fc\u30d3\u30b9\u304c\u3001\u4f55\u3089\u304b\u306e\u5f62\u3067NIST\u304c\u63d0\u4f9b\u3059\u308b\u6280\u8853\u3001\u6e2c\u5b9a\u3001\u6a19\u6e96\u306b\u4f9d\u62e0\u3057\u3066\u3044\u307e\u3059\u3002\u65e5\u672c\u3067\u306f\u60c5\u5831\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u5206\u91ce<\/strong>\u306e\u30ac\u30a4\u30c9\u30e9\u30a4\u30f3\u304c\u6709\u540d\u3067\u3059\u3002<\/p>\n\n\n\n

\u5f0a\u793e\u3067\u306f\u7c73\u56fdNIST\u306e\u8a31\u53ef\u3092\u5f97\u3066\u30012022\u5e743\u6708\u306b\u767a\u8868\u3055\u308c\u305f\u5831\u544a\u66f8 \u201c<\/strong>NIST Special Publication 1270: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence<\/strong><\/a>\u201d<\/strong>\u300cAI\u306e\u30d0\u30a4\u30a2\u30b9\u306e\u7279\u5b9a\u3068\u7ba1\u7406\u306e\u305f\u3081\u306e\u6a19\u6e96\u5316\u306b\u5411\u3051\u3066\u300d<\/strong>\u306b\u3064\u3044\u3066\u3001\u305d\u306e\u30ad\u30fc\u30dd\u30a4\u30f3\u30c8\u3068\u601d\u308f\u308c\u308b\u3068\u3053\u308d\u3092\u629c\u7c8b\u3057\u90a6\u8a33\u306e\u4e0a\u30013\u56de\u306b\u5206\u3051\u3066\u30d6\u30ed\u30b0\u3067\u516c\u958b\u3059\u308b\u3053\u3068\u3068\u81f4\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

\u540c\u5831\u544a\u66f8\u3067\u306f\u3001\u30d0\u30a4\u30a2\u30b9\u306e\u6709\u5bb3\u306a\u5f71\u97ff\u3092\u7279\u5b9a\u3057\u7ba1\u7406\u3057\u3001\u4fe1\u983c\u3067\u304d\u308bAI\u30b7\u30b9\u30c6\u30e0\u3092\u958b\u767a\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u305d\u306e\u6a5f\u68b0\u5b66\u7fd2\u30d7\u30ed\u30bb\u30b9\u3084\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u6280\u8853\u7684\u306a\u691c\u8a3c\u306b\u52a0\u3048\u3001AI\u306e\u958b\u767a\u3084\u4f7f\u7528\u306b\u95a2\u308f\u308b\u793e\u4f1a\u7684\u30fb\u7d44\u7e54\u7684\u30fb\u4eba\u7684\u8981\u56e0\u3084\u80cc\u666f\u306b\u3082\u6ce8\u8996\u3057\u305f\u300c\u793e\u4f1a\u6280\u8853\u7684\u30a2\u30d7\u30ed\u30fc\u30c1(Socio-Technical Perspective)<\/strong>\u300d\u304c\u91cd\u8981\u3067\u3042\u308b\u3068\u6307\u6458\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u90a6\u8a33\u4e26\u3073\u306b\u30d6\u30ed\u30b0\u3067\u306e\u8ee2\u8f09\u306e\u8a31\u53ef\u3092\u9802\u304d\u307e\u3057\u305fNIST\u306e\u3054\u539a\u610f\u306b\u6df1\u8b1d\u7533\u3057\u4e0a\u3052\u307e\u3059\u3068\u5171\u306b\u3001\u672c\u8a18\u4e8b\u304c\u3001\u4eca\u5f8c\u306e\u65e5\u672c\u306e\u793e\u4f1a\u3084\u4f01\u696d\u306b\u304a\u3051\u308bAI\u306e\u5b9f\u88c5\u306b\u304a\u3044\u3066\u3001\u5c11\u3057\u3067\u3082\u7686\u69d8\u306e\u304a\u5f79\u306b\u7acb\u3066\u308b\u3053\u3068\u3092\u9858\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\u5c1a\u3001\u90a6\u8a33\u306b\u95a2\u308f\u308b\u629c\u7c8b\u7b87\u6240\u306e\u9078\u5b9a\u306f\u3001\u3042\u304f\u307e\u3067\u5f0a\u793e\u306e\u79c1\u898b\u30fb\u8cac\u4efb\u306b\u5247\u3063\u305f\u3082\u306e\u3067\u3042\u308a\u3001NIST\u306e\u610f\u56f3\u3092\u73fe\u3057\u305f\u3082\u306e\u3067\u306f\u3054\u3056\u3044\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

Translated with permission courtesy of the National Institute of Standards and Technology (NIST), not an official US Government translation. All rights reserved, US Secretary of Commerce.<\/em><\/p>\n\n\n\n

\u3010NIST Special Publication 1270\u306e\u69cb\u6210\u3011<\/h4>\n\n\n\n

\u4eca\u56de\u306f\u3001Executive Summary\u304b\u30892. AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316\u306e\u5192\u982d\u306e\u90e8\u5206\u307e\u3067\u3092\u8a18\u8f09\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

Executive Summary<\/strong><\/p>\n\n\n\n

1 Purpose and Scope <\/strong>\u76ee\u7684\u3068\u30b9\u30b3\u30fc\u30d7<\/p>\n\n\n\n

2 AI Bias: Context and Terminology <\/strong>AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316<\/p>\n\n\n\n

3 AI Bias: Challenges and Guidance <\/strong>AI\u30d0\u30a4\u30a2\u30b9\uff1a\u8ab2\u984c\u3068\u6307\u91dd<\/p>\n\n\n\n

4. Conclusions <\/strong>\u7d50\u8ad6<\/p>\n\n\n\n

5. Glossary<\/strong><\/p>\n\n\n\n

\n

Executive Summary<\/strong><\/p>\n<\/blockquote>\n\n\n\n

Current attempts for addressing the harmful effects of AI bias remain focused on computational factors such as representativeness of datasets and fairness 
of machine learning algorithms. These remedies are vital for mitigating bias, and more work remains. Yet, as illustrated in Fig. 1, human and systemic institutional and societal factors are significant sources of AI bias as well, and are currently overlooked.<\/p>\n\n\n\n

AI\u306e\u30d0\u30a4\u30a2\u30b9\u306e\u6709\u5bb3\u306a\u5f71\u97ff\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u306e\u591a\u304f\u306f\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u4ee3\u8868\u6027\u3084\u6a5f\u68b0\u5b66\u7fd2\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u516c\u5e73\u6027\u3068\u3044\u3063\u305f\u8a08\u6570\u7684\u306a\u8981\u56e0\u306b\u7126\u70b9\u304c\u5f53\u3066\u3089\u308c\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u6539\u5584\u7b56\u306f\u3001\u30d0\u30a4\u30a2\u30b9\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306b\u4e0d\u53ef\u6b20\u3067\u3042\u308a\u3001\u3084\u308b\u3079\u304d\u3053\u3068\u306f\u307e\u3060\u6ca2\u5c71\u6b8b\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u56f31\u306b\u793a\u3059\u3088\u3046\u306b\u3001\u4eba\u7684\u8981\u56e0\u3042\u308b\u3044\u306f\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u3067\u7d44\u7e54\u793e\u4f1a\u7684\u306a\u8981\u56e0\u3082AI\u30d0\u30a4\u30a2\u30b9\u306e\u91cd\u8981\u306a\u8981\u56e0\u3067\u3042\u308a\u3001\u591a\u304f\u306e\u30b1\u30fc\u30b9\u3067\u898b\u904e\u3054\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The importance of transparency, datasets, and test, evaluation, validation, and verification (TEVV) cannot be overstated. Human factors such as participatory design techniques and multi-stakeholder approaches, and a human-in-the-loop are also important for mitigating risks related to AI bias. However none of these practices individually or in concert are a panacea against bias and each brings its own set of pitfalls.<\/p>\n\n\n\n

\u30b7\u30b9\u30c6\u30e0\u306e\u900f\u660e\u6027\u3084\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3001\u30c6\u30b9\u30c8\u30fb\u8a55\u4fa1\u30fb\u6709\u52b9\u6027\u306e\u78ba\u8a8d\u30fb\u691c\u8a3c\uff08TEVV\uff09\u306e\u91cd\u8981\u6027\u306f\u3001\u3044\u304f\u3089\u5f37\u8abf\u3057\u3066\u3082\u3057\u904e\u304e\u308b\u3053\u3068\u306f\u306a\u3044\u3067\u3057\u3087\u3046\u3002\u307e\u305f\u3001\u53c2\u52a0\u578b\u8a2d\u8a08\u3084\u30de\u30eb\u30c1\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u30a2\u30d7\u30ed\u30fc\u30c1\u3001\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7\u306a\u3069\u306e\u4eba\u7684\u624b\u6cd5\u3082\u3001AI\u306e\u30d0\u30a4\u30a2\u30b9\u306b\u95a2\u3059\u308b\u30ea\u30b9\u30af\u3092\u8efd\u6e1b\u3059\u308b\u4e0a\u3067\u91cd\u8981\u3067\u3059\u3002\u3057\u304b\u3057\u3001\u3053\u308c\u3089\u306e\u624b\u6cd5\u306f\u3001\u5358\u72ec\u3067\u3082\u9023\u643a\u3057\u3066\u3082\u3001\u30d0\u30a4\u30a2\u30b9\u306b\u5bfe\u3059\u308b\u4e07\u80fd\u85ac\u3067\u306f\u306a\u304f\u3001\u305d\u308c\u305e\u308c\u306b\u843d\u3068\u3057\u7a74\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

What is missing from current remedies is guidance from a broader SOCIO-TECHNICAL perspective that connects these practices to societal values. Experts in the area of Trustworthy and Responsible AI counsel that to successfully manage the risks of AI bias we must operationalize these values and create new norms around how AI is built and deployed. This document, and work by the National Institute of Standards and Technology (NIST) in the area of AI bias, is based on a socio-technical perspective<\/strong>.<\/p>\n\n\n\n

\u73fe\u5728\u306e\u30d0\u30a4\u30a2\u30b9\u89e3\u6c7a\u7b56\u306b\u6b20\u3051\u3066\u3044\u308b\u306e\u306f\u3001\u3053\u308c\u3089\u306e\u624b\u6cd5\u3068\u793e\u4f1a\u7684\u4fa1\u5024\u3068\u3092\u7d50\u3073\u3064\u3051\u308b\u3001\u3088\u308a\u5e83\u7bc4\u306a\u793e\u4f1a\u6280\u8853\u7684\u89b3\u70b9\u304b\u3089\u306e\u30ac\u30a4\u30c0\u30f3\u30b9\u3067\u3059\u3002\u4fe1\u983c\u3067\u304d\u308b\u8cac\u4efb\u3042\u308bAI \u306e\u5206\u91ce\u306e\u5c02\u9580\u5bb6\u306f\u3001AI\u306e\u30d0\u30a4\u30a2\u30b9\u30ea\u30b9\u30af\u3092\u3046\u307e\u304f\u7ba1\u7406\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u3053\u308c\u3089\u306e\u4fa1\u5024\u304c\u7dcf\u5408\u7684\u306b\u6a5f\u80fd\u3059\u308b\u3088\u3046\u3001AI\u306e\u69cb\u7bc9\u30fb\u904b\u7528\u624b\u6cd5\u306b\u95a2\u3059\u308b\u65b0\u3057\u3044\u898f\u7bc4\u3092\u4f5c\u6210\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u3068\u52a9\u8a00\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u5831\u544a\u66f8\u3001\u4e26\u3073\u306bAI \u30d0\u30a4\u30a2\u30b9\u306e\u5206\u91ce\u306b\u304a\u3051\u308bNIST\u306e\u7814\u7a76\u306f\u3001\u3053\u3046\u3057\u305f\u793e\u4f1a\u6280\u8853\u7684\u306a\u30a2\u30d7\u30ed\u30fc\u30c1<\/strong>\u306b\u57fa\u3065\u3044\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

The intent of this document is to surface the salient issues in the challenging area of AI bias, and to provide a first step on the roadmap for developing detailed socio-technical guidance for identifying and managing AI bias. Specifically, this special publication:<\/p>\n\n\n\n

\u2022 describes the stakes and challenge of bias in artificial intelligence and provides examples of how and why it can chip away at public trust;<\/p>\n\n\n\n

\u2022 identifies three categories of bias in AI\u200a<\/strong>\u2014\u200asystemic, statistical, and human\u200a<\/strong>\u2014\u200aand describes how and where they contribute to harms;<\/p>\n\n\n\n

\u2022 describes three broad challenges for mitigating bias\u200a<\/strong>\u2014\u200adatasets, testing and evaluation, and human factors\u200a<\/strong>\u2014\u200aand introduces preliminary guidance for addressing them.<\/p>\n\n\n\n

\u672c\u5831\u544a\u66f8\u306e\u76ee\u7684\u306f\u3001AI\u30d0\u30a4\u30a2\u30b9\u306e\u89e3\u6c7a\u56f0\u96e3\u306a\u9818\u57df\u306b\u304a\u3051\u308b\u9855\u8457\u306a\u8ab2\u984c\u3092\u7099\u308a\u51fa\u3057\u3001AI\u30d0\u30a4\u30a2\u30b9\u3092\u7279\u5b9a\u3057\u7ba1\u7406\u3059\u308b\u305f\u3081\u3001\u3088\u308a\u8a73\u7d30\u306a\u793e\u4f1a\u6280\u8853\u7684\u306a\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u69cb\u7bc9\u3059\u308b\u30ed\u30fc\u30c9\u30de\u30c3\u30d7\u306e\u7b2c\u4e00\u6b69\u3092\u63d0\u4f9b\u3059\u308b\u3053\u3068\u3067\u3059\u3002\u5177\u4f53\u7684\u306b\u306f\u3001\u4eca\u56de\u306e\u3053\u306e\u7279\u5225\u51fa\u7248\uff08\u8a33\u6ce8\uff1aNIST SP 1270\uff09\u306f<\/p>\n\n\n\n

    \n
  • AI\u306b\u304a\u3051\u308b\u30d0\u30a4\u30a2\u30b9\u306e\u554f\u984c\u70b9\u3068\u8ab2\u984c\u3092\u8aac\u660e\u3057\u3001\u305d\u308c\u304c\u3069\u306e\u3088\u3046\u306b\u3001\u305d\u3057\u3066\u306a\u305c\u793e\u4f1a\u7684\u4fe1\u983c\u3092\u640d\u306a\u3044\u3046\u308b\u304b\u306e\u4f8b\u3092\u63d0\u793a\u3057\u307e\u3059\u3002<\/li>\n\n\n\n
  • AI\u306b\u304a\u3051\u308b\u30d0\u30a4\u30a2\u30b9\u306e3\u3064\u306e\u30ab\u30c6\u30b4\u30ea\u30fc\uff08\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u3001\u7d71\u8a08\u7684\u30fb\u8a08\u6570\u7684\u30d0\u30a4\u30a2\u30b9\u3001\u4eba\u7684\u30d0\u30a4\u30a2\u30b9\uff09<\/strong>\u3092\u7279\u5b9a\u3057\u3001\u305d\u308c\u3089\u304c\u3069\u306e\u3088\u3046\u306a\u5834\u9762\u3067\u5982\u4f55\u306b\u3057\u3066\u5bb3\u3092\u53ca\u307c\u3059\u304b\u3092\u8aac\u660e\u3057\u307e\u3059\u3002<\/li>\n\n\n\n
  • \u30d0\u30a4\u30a2\u30b9\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306e3\u3064\u306e\u5927\u304d\u306a\u8ab2\u984c\uff08\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3001\u30c6\u30b9\u30c8\u3068\u8a55\u4fa1\u3001\u4eba\u7684\u8981\u56e0\uff09<\/strong>\u3092\u8aac\u660e\u3057\u3001\u305d\u308c\u3089\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u306e\u4e88\u5099\u7684\u306a\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/li>\n<\/ul>\n\n\n\n

    Bias is neither new nor unique to AI and it is not possible to achieve zero risk of bias in an AI system. NIST intends to develop methods for increasing assurance, GOVERNANCE and practice improvements for identifying, understanding, measuring, managing, and reducing bias. To reach this goal, techniques are needed that are flexible, can be applied across contexts regardless of industry, and are easily communicated to different stakeholder groups.<\/p>\n\n\n\n

    \u30d0\u30a4\u30a2\u30b9\u306fAI\u306b\u3068\u3063\u3066\u65b0\u3057\u3044\u3082\u306e\u3067\u3082\u7279\u6b8a\u306a\u3082\u306e\u3067\u3082\u306a\u304f\u3001AI\u30b7\u30b9\u30c6\u30e0\u306b\u304a\u3051\u308b\u30d0\u30a4\u30a2\u30b9\u306e\u30ea\u30b9\u30af\u3092\u30bc\u30ed\u306b\u3059\u308b\u3053\u3068\u306f\u4e0d\u53ef\u80fd\u3067\u3059\u3002NIST\u306f\u3001\u30d0\u30a4\u30a2\u30b9\u3092\u7279\u5b9a\u3057\u3001\u7406\u89e3\u3057\u3001\u6e2c\u5b9a\u3057\u3001\u7ba1\u7406\u30fb\u4f4e\u6e1b\u3059\u308b\u305f\u3081\u306e\u3001\u78ba\u5ea6\u3092\u9ad8\u3081\u3001\u30ac\u30d0\u30ca\u30f3\u30b9\u3057\u3001\u5b9f\u8df5\u7684\u306b\u6539\u5584\u3059\u308b\u624b\u6cd5\u306e\u958b\u767a\u3092\u76ee\u6307\u3057\u307e\u3059\u3002\u3053\u306e\u76ee\u6a19\u3092\u9054\u6210\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u67d4\u8edf\u3067\u3001\u3055\u307e\u3056\u307e\u306a\u696d\u754c\u74b0\u5883\u3092\u8d85\u3048\u3066\u9069\u7528\u3067\u304d\u3001\u7570\u306a\u308b\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u306b\u7c21\u4fbf\u306b\u8a34\u6c42\u3067\u304d\u308b\u624b\u6cd5\u304c\u5fc5\u8981\u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

      \n
    1. Purpose and Scope <\/strong>\u76ee\u7684\u3068\u30b9\u30b3\u30fc\u30d7<\/em><\/li>\n<\/ol>\n\n\n\n

      Working with the AI community, NIST has identified the following technical and socio-technical characteristics needed to cultivate trust in AI systems: accuracy, explainability and interpretability, privacy, reliability, robustness, safety, and security resilience\u200a\u2014\u200aand that harmful biases are mitigated or controlled.<\/p>\n\n\n\n

      NIST\u306f\u3001AI\u306b\u95a2\u308f\u308b\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u5354\u529b\u3057\u3001AI\u30b7\u30b9\u30c6\u30e0\u306e\u4fe1\u983c\u3092\u91b8\u6210\u3057\u3001\u6709\u5bb3\u306a\u30d0\u30a4\u30a2\u30b9\u3092\u7de9\u548c\u307e\u305f\u306f\u5236\u5fa1\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u7cbe\u5ea6\u3001\u8aac\u660e\u53ef\u80fd\u6027\u3068\u89e3\u91c8\u53ef\u80fd\u6027\u3001\u30d7\u30e9\u30a4\u30d0\u30b7\u30fc\u3001\u4fe1\u983c\u6027\u3001\u30ed\u30d0\u30b9\u30c8\u6027\u3001\u5b89\u5168\u6027\u3001\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u8010\u6027\u3068\u3044\u3063\u305f\u6280\u8853\u9762\u4e26\u3073\u306b\u793e\u4f1a\u6280\u8853\u9762\u3067\u306e\u8981\u7d20\u304c\u91cd\u8981\u3067\u3042\u308b\u3053\u3068\u3092\u660e\u3089\u304b\u306b\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

      While AI has significant potential as a transformative technology, it also poses inherent risks. Since trust and risk are closely related, NIST\u2019s work in the area of trustworthy and responsible AI centers around development of a voluntary Risk Management Framework (RMF).<\/p>\n\n\n\n

      AI\u306f\u9769\u65b0\u7684\u306a\u6280\u8853\u3068\u3057\u3066\u5927\u304d\u306a\u53ef\u80fd\u6027\u3092\u79d8\u3081\u3066\u3044\u307e\u3059\u304c\u3001\u540c\u6642\u306bAI\u56fa\u6709\u306e\u30ea\u30b9\u30af\u3082\u5185\u5305\u3057\u3066\u3044\u307e\u3059\u3002\u4fe1\u983c\u3068\u30ea\u30b9\u30af\u306f\u5bc6\u63a5\u306b\u95a2\u9023\u3057\u3066\u3044\u308b\u305f\u3081\u3001NIST\u306f\u3001\u4fe1\u983c\u3067\u304d\u308b\u8cac\u4efb\u3042\u308bAI\u306e\u5206\u91ce\u3067\u3001\u81ea\u4e3b\u7684\u306a\u30ea\u30b9\u30af\u7ba1\u7406\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\uff08RMF\uff09\u306e\u958b\u767a\u306b\u91cd\u70b9\u3092\u7f6e\u3044\u3066\u53d6\u308a\u7d44\u3093\u3067\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      The unique challenges of AI require a deeper understanding of how AI risks differ from other domains. The NIST AI RMF is intended to address risks in the design, development, use, and evaluation of AI products, services, and systems for such tasks as recommendation, diagnosis, pattern recognition, and automated planning and decision making.<\/p>\n\n\n\n

      AI\u7279\u6709\u306e\u8ab2\u984c\u3068\u3057\u3066\u3001AI\u306e\u30ea\u30b9\u30af\u304c\u4ed6\u306e\u9818\u57df\u3068\u3069\u306e\u3088\u3046\u306b\u7570\u306a\u308b\u304b\u3092\u3088\u308a\u6df1\u304f\u7406\u89e3\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u30ea\u30b3\u30e1\u30f3\u30c7\u30fc\u30b7\u30e7\u30f3\u3001\u8a3a\u65ad\u3001\u30d1\u30bf\u30fc\u30f3\u8a8d\u8b58\u3001\u8a08\u753b\u7acb\u6848\u3084\u610f\u601d\u6c7a\u5b9a\u306e\u81ea\u52d5\u5316\u306a\u3069\u306b\u95a2\u308f\u308bAI\u306e\u88fd\u54c1\u3001\u30b5\u30fc\u30d3\u30b9\u3001\u30b7\u30b9\u30c6\u30e0\u306b\u5bfe\u3057\u3066\u3001NIST\u306eAI RMF\u306f\u3001\u305d\u306e\u8a2d\u8a08\u3001\u958b\u767a\u3001\u4f7f\u7528\u3001\u8a55\u4fa1\u306b\u304a\u3051\u308b\u30ea\u30b9\u30af\u306b\u5bfe\u51e6\u3059\u308b\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      The framework is intended to enable the development and use of AI in ways that will increase trustworthiness, advance usefulness, and address potential harms. NIST is leveraging a multi-stakeholder approach to creating and maintaining actionable practice guides via the RMF that is broadly adoptable.<\/p>\n\n\n\n

      \u3053\u306eRMF\u306e\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306f\u3001AI\u306e\u958b\u767a\u304a\u3088\u3073\u4f7f\u7528\u306b\u969b\u3057\u3001\u4fe1\u983c\u6027\u3092\u9ad8\u3081\u3001\u6709\u7528\u6027\u3092\u5411\u4e0a\u3055\u305b\u3001\u6f5c\u5728\u7684\u306a\u554f\u984c\u306b\u5bfe\u51e6\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u3066\u3044\u307e\u3059\u3002NIST\u3067\u306f\u3001RMF\u3092\u901a\u3058\u305f\u3001\u5e83\u7bc4\u56f2\u306b\u9069\u7528\u3067\u304d\u5b9f\u884c\u53ef\u80fd\u306a\u5b9f\u8df5\u30ac\u30a4\u30c9\u3092\u4f5c\u6210\u30fb\u7dad\u6301\u3059\u308b\u305f\u3081\u306b\u3001\u30de\u30eb\u30c1\u30b9\u30c6\u30fc\u30af\u30db\u30eb\u30c0\u30fc\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u63a1\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      2. AI Bias: Context and Terminology <\/strong>\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3068\u7528\u8a9e\u5b9a\u7fa9<\/em><\/p>\n\n\n\n

      AI is neither built nor deployed in a vacuum, sealed off from societal realities of discrimination or unfair practices. Understanding AI as a socio-technical system acknowledges that the processes used to develop technology are more than their mathematical and computational constructs. A socio-technical approach to AI takes into account the values and behavior modeled from the datasets, the humans who interact with them, and the complex organizational factors that go into their commission, design, development, and ultimate deployment.<\/p>\n\n\n\n

      AI\u306f\u3001\u5dee\u5225\u3084\u4e0d\u516c\u6b63\u306a\u884c\u70ba\u3068\u3044\u3063\u305f\u793e\u4f1a\u7684\u73fe\u5b9f\u304b\u3089\u906e\u65ad\u3055\u308c\u305f\u3001\u307e\u3063\u3055\u3089\u306a\u7a7a\u9593\u3067\u69cb\u7bc9\u30fb\u904b\u7528\u3055\u308c\u308b\u3082\u306e\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002AI\u3092\u793e\u4f1a\u6280\u8853\u30b7\u30b9\u30c6\u30e0\u3068\u3057\u3066\u7406\u89e3\u3059\u308b\u3068\u3044\u3046\u3053\u3068\u306f\u3001\u305d\u306e\u6280\u8853\u958b\u767a\u30d7\u30ed\u30bb\u30b9\u304c\u3001\u5358\u306b\u6570\u5b66\u7684\u30fb\u8a08\u6570\u7684\u306a\u8981\u7d20\u3060\u3051\u304b\u3089\u3067\u304d\u3066\u3044\u308b\u306e\u3067\u306f\u306a\u304f\u3001\u305d\u308c\u4ee5\u4e0a\u306e\u3082\u306e\u3067\u3042\u308b\u3053\u3068\u3092\u8a8d\u8b58\u3059\u308b\u3068\u3044\u3046\u3053\u3068\u3067\u3059\u3002AI\u3078\u306e\u793e\u4f1a\u6280\u8853\u7684\u30a2\u30d7\u30ed\u30fc\u30c1\u3068\u306f\u3001\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u304b\u3089\u30e2\u30c7\u30eb\u5316\u3055\u308c\u308b\u3079\u304d\u4fa1\u5024\u89b3\u3084\u884c\u52d5\u3001\u305d\u308c\u3089\u3068\u76f8\u4e92\u4f5c\u7528\u3059\u308b\u4eba\u9593\u3001\u305d\u3057\u3066\u305d\u306e\u8a66\u884c\u3001\u8a2d\u8a08\u3001\u958b\u767a\u3001\u6700\u7d42\u7684\u306a\u5c55\u958b\u306b\u95a2\u308f\u308b\u8907\u96d1\u306a\u7d44\u7e54\u7684\u8981\u56e0\u3092\u3082\u8003\u616e\u306b\u5165\u308c\u308b\u3068\u3044\u3046\u3053\u3068\u3067\u3059\u3002<\/p>\n\n\n\n

      2.1 Characterizing AI bias <\/strong>AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316<\/em><\/p>\n\n\n\n

      2.1.1 Contexts for addressing AI bias<\/strong><\/p>\n\n\n\n

      AI\u30d0\u30a4\u30a2\u30b9\u306b\u5bfe\u51e6\u3059\u308b\u305f\u3081\u306e\u30b3\u30f3\u30c6\u30af\u30b9\u30c8<\/em><\/p>\n\n\n\n

      \n

      Statistical context<\/strong> \u7d71\u8a08\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8<\/p>\n<\/blockquote>\n\n\n\n

      In technical systems, bias is most commonly understood and treated as a statistical phenomenon.<\/p>\n\n\n\n

      \u6280\u8853\u30b7\u30b9\u30c6\u30e0\u306b\u304a\u3044\u3066\u3001\u30d0\u30a4\u30a2\u30b9\u306f\u7d71\u8a08\u7684\u306a\u73fe\u8c61\u3068\u3057\u3066\u7406\u89e3\u3055\u308c\u3001\u6271\u308f\u308c\u308b\u3053\u3068\u304c\u6700\u3082\u4e00\u822c\u7684\u3067\u3059\u3002<\/p>\n\n\n\n

      \n

      Legal context<\/strong> \u6cd5\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8<\/p>\n<\/blockquote>\n\n\n\n

      This section is presented not as legal guidance, rather as a reminder for developers, deployers, and users of AI that they must be cognizant of legal considerations in their work, particularly with regard to bias testing.<\/p>\n\n\n\n

      \u3053\u306e\u30bb\u30af\u30b7\u30e7\u30f3\u3067\u306f\u3001\u6cd5\u7684\u306a\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u63d0\u793a\u3057\u3066\u3044\u308b\u306e\u3067\u306f\u306a\u304f\u3001AI\u306e\u958b\u767a\u8005\u3001\u5c0e\u5165\u8005\u3001\u304a\u3088\u3073\u5229\u7528\u8005\u304c\u3001\u7279\u306b\u30d0\u30a4\u30a2\u30b9\u306b\u3064\u3044\u3066\u306e\u30c6\u30b9\u30c8\u3092\u884c\u3046\u969b\u3001\u305d\u306e\u696d\u52d9\u4e0a\u3001\u6cd5\u7684\u914d\u616e\u304c\u5fc5\u8981\u3067\u3042\u308b\u3053\u3068\u3092\u6ce8\u610f\u559a\u8d77\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      \n

      Cognitive and societal context<\/strong> \u8a8d\u77e5\u30fb\u793e\u4f1a\u7684\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8<\/p>\n<\/blockquote>\n\n\n\n

      The teams involved in AI system design and development bring their cognitive biases, both individual and group, into the process [73]. Bias is prevalent in the assumptions about which data should be used, what AI models should be developed, where the AI system should be placed\u200a\u2014\u200aor if AI is required at all. There are systemic biases at the institutional level that affect how organizations and teams are structured and who controls the decision making processes, and individual and group heuristics and cognitive\/perceptual biases throughout the AI lifecycle (as described in Section 2.4).<\/p>\n\n\n\n

      AI\u30b7\u30b9\u30c6\u30e0\u306e\u8a2d\u8a08\u3068\u958b\u767a\u306b\u643a\u308f\u308b\u30c1\u30fc\u30e0\u306f\u3001\u305d\u308c\u305e\u308c\u306e\u500b\u4eba\u306e\u3001\u305d\u3057\u3066\u30b0\u30eb\u30fc\u30d7\u3068\u3057\u3066\u306e\u3001\u4e21\u65b9\u306e\u8a8d\u77e5\u30d0\u30a4\u30a2\u30b9\u3092\u30d7\u30ed\u30bb\u30b9\u306b\u9f4e\u3057\u307e\u3059[73]\u3002\u3069\u306e\u30c7\u30fc\u30bf\u3092\u4f7f\u3046\u3079\u304d\u304b\u3001\u3069\u306e\u3088\u3046\u306aAI\u30e2\u30c7\u30eb\u3092\u958b\u767a\u3059\u3079\u304d\u304b\u3001AI\u30b7\u30b9\u30c6\u30e0\u3092\u3069\u3053\u306b\u914d\u7f6e\u3059\u3079\u304d\u304b\u3001\u3042\u308b\u3044\u306fAI\u304c\u5168\u304f\u5fc5\u8981\u306a\u3044\u306e\u304b\u3001\u3068\u3044\u3063\u305f\u4eee\u5b9a\u306e\u4e2d\u306b\u30d0\u30a4\u30a2\u30b9\u304c\u904d\u5728\u3057\u3066\u3044\u307e\u3059\u3002\u7d44\u7e54\u3084\u30c1\u30fc\u30e0\u304c\u3069\u306e\u3088\u3046\u306b\u69cb\u6210\u3055\u308c\u3001\u8ab0\u304c\u610f\u601d\u6c7a\u5b9a\u30d7\u30ed\u30bb\u30b9\u3092\u30b3\u30f3\u30c8\u30ed\u30fc\u30eb\u3059\u308b\u304b\u3068\u3044\u3063\u305f\u3053\u3068\u306b\u5f71\u97ff\u3092\u4e0e\u3048\u308b\u7d44\u7e54\u30ec\u30d9\u30eb\u3067\u306e\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u3084\u3001\uff08\u30bb\u30af\u30b7\u30e7\u30f3 2.4 \u3067\u8aac\u660e\u3059\u308b\u3088\u3046\u306b\uff09AI \u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u3092\u901a\u3058\u305f\u3001\u500b\u4eba\u3084\u30b0\u30eb\u30fc\u30d7\u306e\u7d4c\u9a13\u5247\u3001\u3042\u308b\u3044\u306f\u8a8d\u77e5\u30fb\u77e5\u899a\u306b\u4f34\u3046\u30d0\u30a4\u30a2\u30b9\u304c\u5b58\u5728\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

      2.1.2 Categories of AI bias <\/strong>AI\u306e\u30d0\u30a4\u30a2\u30b9\u306e\u30ab\u30c6\u30b4\u30ea\u30fc<\/em><\/p>\n\n\n\n

      Based on previous academic work to classify AI bias [81\u201391] and discussions with thought leaders in the field, it is possible to identify three dominant categories of AI bias.<\/p>\n\n\n\n

      AI\u30d0\u30a4\u30a2\u30b9\u3092\u5206\u985e\u3057\u305f\u3053\u308c\u307e\u3067\u306e\u5b66\u8853\u7684\u7814\u7a76[81\u201391]\u3068\u3001\u3053\u306e\u5206\u91ce\u306e\u30aa\u30d4\u30cb\u30aa\u30f3\u30ea\u30fc\u30c0\u30fc\u3068\u306e\u8b70\u8ad6\u306b\u57fa\u3065\u304d\u3001AI\u30d0\u30a4\u30a2\u30b9\u306e\u4e3b\u8981\u56e0\u306f\u3001\u4ee5\u4e0b3\u3064\u306e\u30ab\u30c6\u30b4\u30ea\u30fc\uff08\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u3001\u7d71\u8a08\u7684\u30fb\u8a08\u6570\u7684\u30d0\u30a4\u30a2\u30b9\u3001\u4eba\u7684\u30d0\u30a4\u30a2\u30b9\uff09<\/strong>\u306b\u985e\u578b\u5316\u3067\u304d\u308b\u3068\u8003\u3048\u3089\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

      \n

      Systemic<\/strong> \u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\uff08\u7d44\u7e54\u5168\u4f53\u306b\u53ca\u3076\u3088\u3046\u306a\u30d0\u30a4\u30a2\u30b9\uff09<\/p>\n<\/blockquote>\n\n\n\n

      Systemic biases<\/strong> result from procedures and practices of particular institutions that operate in ways which result in certain social groups being advantaged or favored and others be ing disadvantaged or devalued. This need not be the result of any conscious prejudice or discrimination but rather of the majority following existing rules or norms. Institutional racism and sexism are the most common examples [92].<\/p>\n\n\n\n

      \u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9<\/strong>\u306f\u3001\u7279\u5b9a\u306e\u6a5f\u95a2\u306e\u624b\u7d9a\u304d\u3084\u6163\u884c\u306b\u8d77\u56e0\u3059\u308b\u3082\u306e\u3067\u3001\u305d\u308c\u304c\u7279\u5b9a\u306e\u793e\u4f1a\u96c6\u56e3\u306b\u306f\u6709\u5229\u306b\u306a\u308a\u3001\u305d\u306e\u4ed6\u306e\u96c6\u56e3\u306f\u4e0d\u5229\u306b\u306a\u308b\u3088\u3046\u306a\u65b9\u6cd5\u3067\u904b\u55b6\u3055\u308c\u3066\u3044\u308b\u305f\u3081\u306b\u751f\u3058\u307e\u3059\u3002\u3053\u308c\u306f\u3001\u5fc5\u305a\u3057\u3082\u610f\u8b58\u7684\u306a\u504f\u898b\u3084\u5dee\u5225\u306e\u7d50\u679c\u3068\u306f\u9650\u3089\u305a\u3001\u3080\u3057\u308d\u591a\u304f\u306f\u3001\u65e2\u5b58\u306e\u30eb\u30fc\u30eb\u3084\u898f\u7bc4\u304c\u539f\u56e0\u3067\u3059\u3002\u5236\u5ea6\u7684\u306a\uff08\u4f01\u696d\u3084\u793e\u4f1a\u306e\uff09\u4eba\u7a2e\u4e3b\u7fa9\u3084\u6027\u5dee\u5225\u306f\u3001\u3053\u306e\u6700\u3082\u4e00\u822c\u7684\u306a\u4e8b\u4f8b\u3067\u3059[92]\u3002<\/p>\n\n\n\n

      \n

      Statistical and Computational<\/strong> \u7d71\u8a08\u7684\u30fb\u8a08\u6570\u7684\u30d0\u30a4\u30a2\u30b9<\/p>\n<\/blockquote>\n\n\n\n

      Statistical and computational biases<\/strong> stem from errors that result when the sample is not representative of the population. These biases arise from systematic as opposed to random error and can occur in the absence of prejudice, partiality, or discriminatory intent [98].<\/p>\n\n\n\n

      \u7d71\u8a08\u7684\u30fb\u8a08\u6570\u7684\u30d0\u30a4\u30a2\u30b9<\/strong>\u306f\u3001\u30b5\u30f3\u30d7\u30eb\u304c\u6bcd\u96c6\u56e3\u3092\u4ee3\u8868\u3057\u3066\u3044\u306a\u3044\u5834\u5408\u306b\u751f\u3058\u308b\u30a8\u30e9\u30fc\u306b\u8d77\u56e0\u3057\u307e\u3059\u3002\u3053\u308c\u3089\u306e\u30d0\u30a4\u30a2\u30b9\u306f\u30e9\u30f3\u30c0\u30e0\u8aa4\u5dee\u3068\u306f\u5bfe\u7167\u7684\u306b\u3001\u898f\u5247\u7684\u306a\u8aa4\u5dee\u304b\u3089\u751f\u3058\u307e\u3059\u3002\u307e\u305f\u3001\u504f\u898b\u3001\u4e0d\u516c\u5e73\u3001\u5dee\u5225\u7684\u306a\u610f\u56f3\u304c\u306a\u304f\u3066\u3082\u767a\u751f\u3057\u307e\u3059[98]<\/p>\n\n\n\n

      In AI systems, these biases are present in the datasets and algorithmic processes used in the development of AI applications, and often arise when algorithms are trained on one type of data and cannot extrapolate beyond those data.<\/p>\n\n\n\n

      AI\u30b7\u30b9\u30c6\u30e0\u306b\u304a\u3044\u3066\u306f\u3001\u3053\u308c\u3089\u306e\u30d0\u30a4\u30a2\u30b9\u306fAI\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306e\u958b\u767a\u306b\u4f7f\u7528\u3055\u308c\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3068\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u30d7\u30ed\u30bb\u30b9\u306b\u5185\u5728\u3057\u307e\u3059\u3002\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u304c\u3042\u308b\u7279\u5b9a\u306e\u30c7\u30fc\u30bf\u9818\u57df\u3067\u8a13\u7df4\u3055\u308c\u3001\u305d\u306e\u30c7\u30fc\u30bf\u9818\u57df\u3092\u8d85\u3048\u305f\u5916\u633f\u304c\u9069\u7528\u3067\u304d\u306a\u3044\u5834\u5408\u306b\u3057\u3070\u3057\u3070\u767a\u751f\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

      The error may be due to heterogeneous data, representation of complex data in simpler mathematical representations, wrong data, and algorithmic biases such as over- and under-fitting, the treatment of outliers, and data cleaning and imputation factors.<\/p>\n\n\n\n

      \u3053\u3046\u3057\u305f\u30a8\u30e9\u30fc\u306f\u3001\u30d8\u30c6\u30ed\u30b8\u30cb\u30a2\u30b9\u30c7\u30fc\u30bf\uff08\u8a33\u6ce8\uff1a\u7a2e\u985e\u3084\u5f62\u5f0f\u306e\u3070\u3089\u3064\u304d\u304c\u5927\u304d\u3044\u30c7\u30fc\u30bf\uff09\u3084\u3001\u5358\u7d14\u306a\u6570\u5b66\u7684\u8868\u73fe\u3067\u7e8f\u3081\u3089\u308c\u305f\u8907\u96d1\u306a\u30c7\u30fc\u30bf\u3001\u8aa4\u3063\u305f\u30c7\u30fc\u30bf\u7b49\u304c\u5143\u3068\u306a\u3063\u3066\u3044\u305f\u308a\u3001\u3042\u308b\u3044\u306f\u30aa\u30fc\u30d0\u30fc\u30d5\u30a3\u30c3\u30c8\u3084\u30a2\u30f3\u30c0\u30fc\u30d5\u30a3\u30c3\u30c8\u3001\u5916\u308c\u5024\u306e\u53d6\u308a\u6271\u3044\u3001\u30c7\u30fc\u30bf\u30af\u30ec\u30f3\u30b8\u30f3\u30b0\u3084\u30a4\u30f3\u30d4\u30e5\u30c6\u30fc\u30b7\u30e7\u30f3\uff08\u6b20\u640d\u5024\u88dc\u5b8c\uff09\u8981\u56e0\u306a\u3069\u3001\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u4e0a\u306e\u30d0\u30a4\u30a2\u30b9\u304c\u539f\u56e0\u3067\u3042\u3063\u305f\u308a\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

      \n

      Human<\/strong> \u4eba\u7684\u30d0\u30a4\u30a2\u30b9<\/p>\n<\/blockquote>\n\n\n\n

      Human biases<\/strong> reflect systematic errors in human thought based on a limited number of heuristic principles and predicting values to simpler judgmental operations [99]. These biases are often implicit and tend to relate to how an individual or group perceives information (such as automated AI output) to make a decision or fill in missing or unknown information.<\/p>\n\n\n\n

      \u4eba\u7684\u30d0\u30a4\u30a2\u30b9<\/strong>\u306f\u3001\u3053\u308c\u307e\u3067\u306e\u9650\u3089\u308c\u305f\u7d4c\u9a13\u5247\u306b\u57fa\u3065\u3044\u3066\u3001\u7269\u4e8b\u3092\u5358\u7d14\u5316\u3057\u3066\u5224\u65ad\u3092\u3057\u3066\u3057\u307e\u3046\u4eba\u9593\u306e\u601d\u8003\u30a8\u30e9\u30fc\u306b\u8d77\u56e0\u3057\u3066\u3044\u307e\u3059[99]\u3002\u3053\u308c\u3089\u306e\u30d0\u30a4\u30a2\u30b9\u306f\u3001\u3057\u3070\u3057\u3070\u6697\u9ed9\u7684\u3067\u3042\u308a\u3001\u500b\u4eba\u307e\u305f\u306f\u30b0\u30eb\u30fc\u30d7\u304c\u5982\u4f55\u306b\uff08\u81ea\u52d5\u5316\u3055\u308c\u305fAI\u51fa\u529b\u306a\u3069\u306e\uff09\u60c5\u5831\u3092\u8a8d\u8b58\u3057\u3001\u6c7a\u5b9a\u3092\u4e0b\u3057\u3001\u3042\u308b\u3044\u306f\u6b20\u843d\u60c5\u5831\u3084\u672a\u77e5\u60c5\u5831\u3092\u57cb\u3081\u3066\u3057\u307e\u3046\u304b\u3068\u3044\u3063\u305f\u3053\u3068\u306b\u95a2\u9023\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      There is a wide variety of human biases. Cognitive and perceptual biases show themselves in all domains and are not unique to human interactions with AI. Rather, they are a fundamental part of the human mind.<\/p>\n\n\n\n

      \u4eba\u7684\u30d0\u30a4\u30a2\u30b9\u306f\u591a\u7a2e\u591a\u69d8\u3067\u3059\u3002\u8a8d\u77e5\u30fb\u77e5\u899a\u30d0\u30a4\u30a2\u30b9\u306f\u3042\u3089\u3086\u308b\u9818\u57df\u3067\u898b\u3089\u308c\u308b\u3082\u306e\u3067\u3042\u308a\u3001\u4eba\u9593\u3068AI\u3068\u306e\u76f8\u4e92\u4f5c\u7528\u3060\u3051\u306b\u7279\u6709\u306a\u3082\u306e\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u3080\u3057\u308d\u3001\u305d\u308c\u3089\u306f\u4eba\u9593\u306e\u5fc3\u306e\u57fa\u672c\u7684\u306a\u90e8\u5206\u3067\u3059\u3002<\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n

      2.2 How AI bias contributes to harms<\/strong> AI\u30d0\u30a4\u30a2\u30b9\u304c\u9f4e\u3059\u554f\u984c<\/em><\/p>\n\n\n\n

      Technology based on AI has tighter connections to and broader impacts on society than traditional software. Applications that utilize AI are often deployed across sectors and contexts for decision-support and decision-making.<\/p>\n\n\n\n

      AI\u306b\u57fa\u3065\u304f\u30c6\u30af\u30ce\u30ed\u30b8\u30fc\u306f\u3001\u5f93\u6765\u306e\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3088\u308a\u3082\u793e\u4f1a\u3068\u306e\u7d50\u3073\u3064\u304d\u304c\u5f37\u304f\u3001\u793e\u4f1a\u306b\u5bfe\u3057\u3066\u3088\u308a\u5e83\u3044\u5f71\u97ff\u3092\u4e0e\u3048\u307e\u3059\u3002AI\u3092\u6d3b\u7528\u3057\u305f\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u306f\u3001\u610f\u601d\u6c7a\u5b9a\u3084\u305d\u306e\u30b5\u30dd\u30fc\u30c8\u306a\u3069\u3001\u5206\u91ce\u3084\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\uff08\u8a33\u6ce8\uff1a\u4f7f\u7528\u74b0\u5883\u3084\u80cc\u666f\uff09\u3092\u8d85\u3048\u3066\u6d3b\u7528\u3055\u308c\u308b\u3053\u3068\u3082\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

      Yet, ML models tend to exhibit \u201cunexpectedly poor behavior when deployed in real world domains<\/strong>\u201d without domain-specific constraints supplied by human operators [103].<\/p>\n\n\n\n

      \u3057\u304b\u3057\u306a\u304c\u3089\u3001\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u304c\u4eba\u9593\u306e\u30aa\u30da\u30ec\u30fc\u30bf\u306b\u3088\u308b\u30c9\u30e1\u30a4\u30f3\u56fa\u6709\u306e\u5236\u7d04\u6761\u4ef6\u306a\u3057\u306b\u904b\u7528\u3055\u308c\u305f\u5834\u5408\u3001\u300c\uff08\u8a33\u6ce8\uff1a\u958b\u767a\u74b0\u5883\u3068\u306f\u7570\u306a\u308b\uff09\u5b9f\u4e16\u754c\u306e\u30c9\u30e1\u30a4\u30f3\u306b\u304a\u3044\u3066\u3001\u4e88\u671f\u305b\u306c\u5f62\u3067\u60aa\u3044\u632f\u308b\u821e\u3044\u3092\u793a\u3059\u4e8b\u614b<\/strong>\u300d\u304c\u3057\u3070\u3057\u3070\u767a\u751f\u3057\u307e\u3059[103]\u3002<\/p>\n\n\n\n

      These biases can negatively impact individuals and society by amplifying and reinforcing discrimination at a speed and scale far beyond the traditional discriminatory practices that can result from implicit human or institutional biases such as racism, sexism, ageism or ableism.<\/p>\n\n\n\n

      \u4eba\u7a2e\u5dee\u5225\u3001\u6027\u5dee\u5225\u3001\u5e74\u9f62\u5dee\u5225\u3001\u80fd\u529b\u5dee\u5225\u306a\u3069\u3001\u4eba\u9593\u3084\u7d44\u7e54\u306b\u6f5c\u3080\u6697\u9ed9\u7684\u306a\u504f\u898b\u306b\u57fa\u3065\u304f\u5f93\u6765\u306e\u5dee\u5225\u306b\u6bd4\u3079\u3001AI\u30d0\u30a4\u30a2\u30b9\u306f\u3001\u306f\u308b\u304b\u306b\u901f\u3044\u30b9\u30d4\u30fc\u30c9\u3068\u30b9\u30b1\u30fc\u30eb\u3067\u5dee\u5225\u3092\u8a98\u767a\u30fb\u5897\u5f37\u3057\u3001\u500b\u4eba\u3068\u793e\u4f1a\u306b\u30cd\u30ac\u30c6\u30a3\u30d6\u306a\u5f71\u97ff\u3092\u53ca\u307c\u3059\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

      \u306f\u3058\u3081\u306b \u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)\u306f\u30011901\u5e74\u306b\u8a2d\u7acb\u3055\u308c\u305f\u7c73\u56fd\u6700\u53e4\u306e\u7269\u7406\u79d1\u5b66\u7814\u7a76\u6240\u306e\u4e00\u3064\u3067\u3042\u308a\u3001\u73fe\u5728\u306f\u7c73\u56fd\u5546\u52d9\u7701\u5098\u4e0b\u306b\u5c5e\u3057\u3066\u3044\u307e\u3059\u3002\u3055\u307e\u3056\u307e\u306a\u79d1\u5b66\u6280\u8853\u5206\u91ce\u306b\u304a\u3051\u308b\u8a08\u6e2c\u3068\u6a19\u6e96\u5316\u3092\u901a\u3058\u3001\u7c73\u56fd\u306e\u307f\u306a\u3089\u305a\u4e16\u754c\u3092\u727d\u5f15\u3059 […]<\/p>\n","protected":false},"author":11,"featured_media":1329,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9],"tags":[],"class_list":["post-1323","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog"],"_links":{"self":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts\/1323","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/comments?post=1323"}],"version-history":[{"count":0,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts\/1323\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media\/1329"}],"wp:attachment":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media?parent=1323"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/categories?post=1323"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/tags?post=1323"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}