{"id":1087,"date":"2022-04-29T13:26:00","date_gmt":"2022-04-29T04:26:00","guid":{"rendered":"https:\/\/citadelai.wpengine.com\/?p=1087"},"modified":"2024-09-17T12:21:41","modified_gmt":"2024-09-17T03:21:41","slug":"toward-standardization-for-bias-identification-and-management-in-ai-3-3-nist-special-publication-1270","status":"publish","type":"post","link":"https:\/\/citadel-ai.com\/ja\/blog\/2022\/04\/29\/toward-standardization-for-bias-identification-and-management-in-ai-3-3-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\uff083\/3\uff09NIST Special Publication 1270"},"content":{"rendered":"\n

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

\u5f0a\u793e\u3067\u306f\u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)<\/strong><\/a>\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\u4eca\u56de\u306f\u305d\u306e\u6700\u7d42\u56de\u3067\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<\/h2>\n\n\n\n

\u7b2c\u4e8c\u56de<\/strong><\/a>\u306b\u5f15\u304d\u7d9a\u304d\u3001\u6700\u7d42\u56de\u306e\u4eca\u56de\u306f\u4ee5\u4e0b\u69cb\u6210\u306e\u5185\u30013.AI\u30d0\u30a4\u30a2\u30b9\u306e\u8ab2\u984c\u3068\u6307\u91dd\u306e\u5f8c\u534a\u304b\u3089\u7d42\u308f\u308a\u307e\u3067\u306b\u3064\u3044\u3066\u3001\u4e3b\u306a\u8ad6\u70b9\u3092\u53d6\u308a\u307e\u3068\u3081\u307e\u3059\u3002<\/p>\n\n\n\n

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

    \n
  1. Purpose and Scope <\/strong>\u76ee\u7684\u3068\u30b9\u30b3\u30fc\u30d7<\/li>\n\n\n\n
  2. AI Bias: Context and Terminology <\/strong>AI\u30d0\u30a4\u30a2\u30b9\u306e\u985e\u578b\u5316<\/li>\n\n\n\n
  3. AI Bias: Challenges and Guidance <\/strong>AI\u30d0\u30a4\u30a2\u30b9\uff1a\u8ab2\u984c\u3068\u6307\u91dd<\/li>\n\n\n\n
  4. Conclusions <\/strong>\u7d50\u8ad6<\/li>\n\n\n\n
  5. Glossary<\/strong><\/li>\n<\/ol>\n\n\n\n
    \n

    \u6700\u7d42\u56de\u306e\u4eca\u56de\u3054\u7d39\u4ecb\u3059\u308b\u90e8\u5206\u306f\u3001AI\u30d0\u30a4\u30a2\u30b9\u306e\u30ac\u30d0\u30ca\u30f3\u30b9\u306b\u95a2\u308f\u308b\u6280\u8853\u7684\u30fb\u4eba\u7684\u30a2\u30d7\u30ed\u30fc\u30c1\u306b\u3064\u3044\u3066\u3001\u3088\u308a\u5177\u4f53\u7684\u306a\u8b70\u8ad6\u304c\u306a\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<\/blockquote>\n\n\n\n

    3.2 How do we know what is right? TEVV Considerations for AI Bias<\/strong><\/p>\n\n\n\n

    \u4f55\u304c\u6b63\u3057\u3044\u304b\u304c\u3001\u3069\u3046\u3059\u308c\u3070\u308f\u304b\u308b\u306e\u304b\uff1fAI\u30d0\u30a4\u30a2\u30b9\u306eTEVV(Testing Evaluation Validation Verification)\u8003\u5bdf<\/em><\/p>\n\n\n\n

    3.2.1 TEVV Challenges<\/strong><\/p>\n\n\n\n

    TEVV(T<\/em>est, Evaluation, Validation, and Verification)\u306e\u8ab2\u984c<\/em><\/p>\n\n\n\n

    Even the algorithm itself relies on data for training and performance tuning, which in turn can be assessed by a fairness metric. Therefore, when we consider the computational approaches to mitigating bias, we must take into consideration these three components together: algorithms, data, and fairness metrics.<\/strong><\/p>\n\n\n\n

    \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u81ea\u4f53\u3082\u3001\u5b66\u7fd2\u3084\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3092\u8abf\u6574\u3059\u308b\u305f\u3081\u3001\u30c7\u30fc\u30bf\u306b\u4f9d\u5b58\u3057\u3066\u304a\u308a\u3001\u4e00\u65b9\u305d\u306e\u30c7\u30fc\u30bf\u306f\u516c\u5e73\u6027\u306e\u6307\u6a19\u304b\u3089\u8a55\u4fa1\u3055\u308c\u308b\u3079\u304d\u3067\u3059\u3002\u3057\u305f\u304c\u3063\u3066\u3001\u30d0\u30a4\u30a2\u30b9\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306e\u8a08\u6570\u7684\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u691c\u8a0e\u3059\u308b\u969b\u306b\u306f\u3001\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3001\u30c7\u30fc\u30bf\u3001\u516c\u5e73\u6027\u30e1\u30c8\u30ea\u30c3\u30af\u306e3\u3064\u306e\u8981\u7d20<\/strong>\u3092\u5408\u308f\u305b\u3066\u8003\u616e\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    AI systems regularly model concepts that are \u2014 at best \u2014 only partially observable or capturable by data.<\/strong> Without direct measures for these highly complex considerations, AI development teams use proxies, which can create many risks [150].<\/p>\n\n\n\n

    AI\u30b7\u30b9\u30c6\u30e0\u306f\u3001\u305b\u3044\u305c\u3044\u90e8\u5206\u7684\u306b\u3057\u304b\u89b3\u6e2c\u3067\u304d\u306a\u3044\u3001\u3042\u308b\u3044\u306f\u30c7\u30fc\u30bf\u3067\u306f\u76f4\u63a5\u6355\u6349\u3067\u304d\u306a\u3044\u3088\u3046\u306a\u6982\u5ff5\u3092\u3001\u901a\u5e38\u30e2\u30c7\u30eb\u5316\u3057\u3066\u3044\u307e\u3059\u3002<\/strong>\u3053\u308c\u3089\u306e\u975e\u5e38\u306b\u8907\u96d1\u306a\u6982\u5ff5\u306b\u5bfe\u3057\u3066\u306f\u3001\u76f4\u63a5\u7684\u306a\u30c7\u30fc\u30bf\u8a08\u6e2c\u65b9\u6cd5\u304c\u306a\u3044\u305f\u3081\u3001AI\u958b\u767a\u30c1\u30fc\u30e0\u306f\u30d7\u30ed\u30ad\u30b7<\/strong>\uff08\u8a33\u6ce8\uff1a\u76f4\u63a5\u6e2c\u5b9a\u3067\u304d\u306a\u3044\u5909\u6570\u306b\u5bfe\u3057\u3066\u5229\u7528\u3059\u308b\u4ee3\u7406\u5909\u6570\uff09\u3092\u4f7f\u7528\u3057\u307e\u3059\u304c\u3001\u3053\u308c\u304c\u591a\u304f\u306e\u30ea\u30b9\u30af\u3092\u751f\u307f\u51fa\u3059\u3053\u3068\u306b\u306a\u308a\u307e\u3059[150]\u3002<\/p>\n\n\n\n

    \n

    Epistemic and aleatoric uncertainty <\/strong>Epistemic \u4e26\u3073\u306b Aleatoric\u306a\u4e0d\u78ba\u5b9f\u6027<\/p>\n<\/blockquote>\n\n\n\n

    ML distinguishes two types of predictive uncertainty: EPISTEMIC and ALEATORIC [152].<\/p>\n\n\n\n

    \u6a5f\u68b0\u5b66\u7fd2\u3067\u306f\u4e88\u6e2c\u306e\u4e0d\u78ba\u5b9f\u6027\u30922\u7a2e\u985e\u306b\u533a\u5225\u3057\u3066\u3044\u307e\u3059\u3002Epistemic\u306a\u4e0d\u78ba\u5b9f\u6027\u3068Aleatoric\u306a\u4e0d\u78ba\u5b9f\u6027\u3067\u3059[152]\uff0e<\/p>\n\n\n\n

    While epistemic uncertainty<\/strong> can be reduced by increasing the amount of representative training data, it cannot be fully eliminated. This can impact the behavior of a deep learning system in deployment when used with real-world data, especially when there is a mismatch in the distributions of the real and training data<\/strong> [103]. This can lead to undesirable effects on many of the AI system\u2019s critical attributes (e.g., robustness, resilience), including inducing harmful bias.<\/p>\n\n\n\n

    Epistemic\u306a\u4e0d\u78ba\u5b9f\u6027<\/strong>\uff08\u8a33\u6ce8: \u30e2\u30c7\u30eb\u306e\u4e0d\u78ba\u5b9f\u6027\u306e\u3046\u3061\u3001\u5b66\u7fd2\u3059\u308c\u3070\u6539\u5584\u3059\u308b\u3082\u306e\uff09\u306f\u3001\u4ee3\u8868\u7684\u306a\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u91cf\u3092\u5897\u3084\u3059\u3053\u3068\u3067\u4f4e\u6e1b\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u304c\u3001\u5b8c\u5168\u306b\u6392\u9664\u3059\u308b\u3053\u3068\u306f\u3067\u304d\u307e\u305b\u3093\u3002\u3053\u308c\u306f\u3001\u5b9f\u30c7\u30fc\u30bf\u3092\u4f7f\u3063\u305f\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u30b9\u30c6\u30fc\u30b8\u306b\u304a\u3044\u3066\u3001\u6df1\u5c64\u5b66\u7fd2\u30b7\u30b9\u30c6\u30e0\u306e\u6319\u52d5\u306b\u5f71\u97ff\u3092\u4e0e\u3048\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u7279\u306b\u5b9f\u30c7\u30fc\u30bf\u3068\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u5206\u5e03\u306b\u4e0d\u4e00\u81f4\u304c\u3042\u308b\u5834\u5408<\/strong>\u304c\u554f\u984c\u3067\u3059[103]\u3002\u3053\u306e\u7d50\u679c\u3001AI\u30b7\u30b9\u30c6\u30e0\u306e\u91cd\u8981\u306a\u7279\u6027\uff08\u4f8b\u3048\u3070\u3001\u30ed\u30d0\u30b9\u30c8\u6027\u3001\u30ec\u30b8\u30ea\u30a8\u30f3\u30b9\uff09\u306b\u5bfe\u3057\u3001\u6709\u5bb3\u306a\u30d0\u30a4\u30a2\u30b9\u306e\u9855\u5728\u5316\u3092\u542b\u3080\u3001\u671b\u307e\u3057\u304f\u306a\u3044\u5f71\u97ff\u3092\u4e0e\u3048\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Another inherent type of uncertainty associated with machine learning is ALEATORIC<\/strong>. It represents the uncertainty inherent in the data, e.g., the uncertainty in the label assigning process of the training dataset. Aleatoric uncertainty is the irreducible part of the predictive uncertainty.<\/p>\n\n\n\n

    \u6a5f\u68b0\u5b66\u7fd2\u306b\u56fa\u6709\u306e\u3082\u3046\u4e00\u3064\u306e\u4e0d\u78ba\u5b9f\u6027\u306f\u3001Aleatoric \u306a\u4e0d\u78ba\u5b9f\u6027<\/strong>\uff08\u8a33\u6ce8\uff1a\u74b0\u5883\u7b49\u30e9\u30f3\u30c0\u30e0\u306a\u4e8b\u8c61\u304c\u539f\u56e0\u3067\u767a\u751f\u3059\u308b\u4e0d\u78ba\u5b9f\u6027\u3001\u5b66\u7fd2\u3060\u3051\u3067\u306f\u6539\u5584\u3057\u306a\u3044\u3082\u306e\uff09\u3067\u3059\u3002\u3053\u308c\u306f\u30c7\u30fc\u30bf\u305d\u306e\u3082\u306e\u306b\u5185\u5728\u3059\u308b\u4e0d\u78ba\u5b9f\u6027\u3001\u4f8b\u3048\u3070\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u30e9\u30d9\u30eb\u4ed8\u4e0e\u904e\u7a0b\u306b\u304a\u3051\u308b\u4e0d\u78ba\u5b9f\u6027\u3067\u3059\u3002Aleatoric\u306a\u4e0d\u78ba\u5b9f\u6027\u306f\u3001\u4e88\u6e2c\u306e\u4e0d\u78ba\u5b9f\u6027\u306e\u3046\u3061\u3001\u8efd\u6e1b\u3059\u308b\u3053\u3068\u304c\u96e3\u3057\u3044\u4e0d\u78ba\u5b9f\u6027\u3067\u3059\u3002<\/p>\n\n\n\n

    \n

    \u200b\u200bProcesses <\/strong>\u30d7\u30ed\u30bb\u30b9<\/p>\n<\/blockquote>\n\n\n\n

    The software designers and data scientists working in design and development are often highly focused on system performance and optimization. This focus can inadvertently be a source of bias in AI systems. For example, during model development and selection, modelers will almost always select the most accurate models. Yet, as Forde et al describe in their paper, [167] selecting models based solely on accuracy is not necessarily the best approach for bias reduction<\/strong>.<\/p>\n\n\n\n

    \u8a2d\u8a08\u30fb\u958b\u767a\u306b\u643a\u308f\u308b\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u8a2d\u8a08\u8005\u3084\u30c7\u30fc\u30bf\u30b5\u30a4\u30a8\u30f3\u30c6\u30a3\u30b9\u30c8\u306f\u3001\u30b7\u30b9\u30c6\u30e0\u306e\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u3084\u6700\u9069\u5316\u306b\u9ad8\u3044\u95a2\u5fc3\u3092\u5bc4\u305b\u3066\u3044\u307e\u3059\u3002\u3053\u3046\u3057\u305f\u53d6\u7d44\u65b9\u91dd\u306f\u3001AI\u30b7\u30b9\u30c6\u30e0\u306b\u304a\u3044\u3066\u601d\u308f\u306c\u5f62\u3067\u30d0\u30a4\u30a2\u30b9\u306e\u539f\u56e0\u3068\u306a\u308b\u3053\u3068\u304c\u3042\u308a\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u30e2\u30c7\u30eb\u306e\u958b\u767a\u3068\u9078\u629e\u306b\u304a\u3044\u3066\u3001\u30e2\u30c7\u30eb\u958b\u767a\u8005\u306f\u307b\u3068\u3093\u3069\u306e\u5834\u5408\u3001\u6700\u3082\u7cbe\u5ea6\u306e\u9ad8\u3044\u30e2\u30c7\u30eb\u3092\u9078\u629e\u3057\u307e\u3059\u3002\u3057\u304b\u3057\u3001Forde\u3089\u304c\u8ad6\u6587\u3067\u8ff0\u3079\u3066\u3044\u308b\u3088\u3046\u306b\u3001[167]\u7cbe\u5ea6\u306e\u307f\u306b\u57fa\u3065\u3044\u3066\u30e2\u30c7\u30eb\u3092\u9078\u629e\u3059\u308b\u3053\u3068\u306f\u3001\u5fc5\u305a\u3057\u3082\u30d0\u30a4\u30a2\u30b9\u3092\u6e1b\u3089\u3059\u305f\u3081\u306e\u6700\u826f\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u3067\u306f\u3042\u308a\u307e\u305b\u3093<\/strong>\u3002<\/p>\n\n\n\n

    Furthermore, the choice of the model\u2019s objective function, upon which a model\u2019s definition of accuracy is based<\/strong>, can reflect bias. Not taking context into consideration during model selection can lead to biased results for sub-populations (for example, disparities in health care delivery).<\/p>\n\n\n\n

    \u3055\u3089\u306b\u3001\u30e2\u30c7\u30eb\u306e\u7cbe\u5ea6\u306e\u57fa\u6e96\u3068\u306a\u308b\u76ee\u7684\u95a2\u6570\uff08\u640d\u5931\u95a2\u6570\uff09\u306e\u9078\u629e\u306b\u3082\u30d0\u30a4\u30a2\u30b9\u304c\u304b\u304b\u3063\u3066\u3044\u308b\u3053\u3068\u304c\u3042\u308a\u307e\u3059\u3002<\/strong>\u30e2\u30c7\u30eb\u9078\u629e\u6642\u306b\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3092\u8003\u616e\u3057\u306a\u3044\u3068\u3001\u90e8\u5206\u7684\u306a\u6bcd\u96c6\u56e3\u306b\u5bfe\u3057\u3066\u504f\u3063\u305f\u7d50\u679c\uff08\u4f8b\u3048\u3070\u3001\u533b\u7642\u63d0\u4f9b\u306b\u304a\u3051\u308b\u683c\u5dee\uff09\u3092\u3082\u305f\u3089\u3059\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Relatedly, systems that are designed to use aggregated data about groups to make predictions about individual behavior \u2014 a practice initially meant to be a remedy for non-representative datasets[18] \u2014 can lead to biased outcomes. This bias, known as ECOLOGICAL FALLACY, occurs when an inference is made about an individual based on their membership within a group (for example, predicting college performance risk based on an individual\u2019s race [52]).<\/p>\n\n\n\n

    \u307e\u305f\u3001\u96c6\u56e3\u306b\u3064\u3044\u3066\u96c6\u7d04\u3055\u308c\u305f\u30c7\u30fc\u30bf\u3092\u7528\u3044\u3066\u500b\u4eba\u306e\u884c\u52d5\u3092\u4e88\u6e2c\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u305f\u30b7\u30b9\u30c6\u30e0\u306f\uff08\u3082\u3068\u3082\u3068\u306f\u3001\u5c11\u6570\u306e\u975e\u4ee3\u8868\u7684\u306a\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8[18]\u306b\u5bfe\u3059\u308b\u6551\u6e08\u63aa\u7f6e\u3068\u3057\u3066\u610f\u56f3\u3055\u308c\u3066\u3044\u305f\u3082\u306e\u3067\u3042\u3063\u3066\u3082\uff09\u3001\u504f\u3063\u305f\u7d50\u679c\u3092\u3082\u305f\u3089\u3059\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u306e\u30d0\u30a4\u30a2\u30b9\u306fECOLOGICAL FALLACY\u3068\u3057\u3066\u77e5\u3089\u308c\u3001\u30b0\u30eb\u30fc\u30d7\u5c5e\u6027\u306b\u57fa\u3065\u3044\u3066\u500b\u4eba\u306b\u5bfe\u3059\u308b\u63a8\u8ad6\u3092\u884c\u3046\u6642\u306b\u767a\u751f\u3057\u307e\u3059\uff08\u4f8b\u3048\u3070\u3001\u500b\u4eba\u306e\u4eba\u7a2e\u306b\u57fa\u3065\u3044\u3066\u5927\u5b66\u306e\u6210\u7e3e\u4e88\u6e2c\u3059\u308b\u5834\u5408\u306e\u30ea\u30b9\u30af\u306a\u3069[52]\uff09\u3002<\/p>\n\n\n\n

    \n

    Algorithmic effects <\/strong>\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306b\u3088\u308b\u5f71\u97ff<\/p>\n<\/blockquote>\n\n\n\n

    Simple models with fewer parameters<\/strong> are often used because they tend to be less expensive to build, more explainable and more transparent, and easier to implement. However, such models can exacerbate statistical biases because restrictive assumptions on the training data often do not hold with nuanced demographics.<\/strong><\/p>\n\n\n\n

    \u30d1\u30e9\u30e1\u30fc\u30bf\u304c\u5c11\u306a\u3044\u5358\u7d14\u306a\u30e2\u30c7\u30eb<\/strong>\u306f\u3001\u69cb\u7bc9\u30b3\u30b9\u30c8\u304c\u4f4e\u304f\u3001\u8aac\u660e\u53ef\u80fd\u3067\u900f\u660e\u6027\u304c\u9ad8\u304f\u3001\u5b9f\u88c5\u304c\u7c21\u5358\u306a\u50be\u5411\u304c\u3042\u308b\u305f\u3081\u3001\u3057\u3070\u3057\u3070\u4f7f\u7528\u3055\u308c\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u3053\u306e\u3088\u3046\u306a\u30e2\u30c7\u30eb\u306f\u3001\u7d71\u8a08\u7684\u30d0\u30a4\u30a2\u30b9\u3092\u60aa\u5316\u3055\u305b\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/strong>\u306a\u305c\u306a\u3089\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf\u306b\u5bfe\u3059\u308b\u9650\u5b9a\u7684\u306a\u4eee\u5b9a\u306f\u3001\u30cb\u30e5\u30a2\u30f3\u30b9\u306e\u7570\u306a\u308b\u30c7\u30e2\u30b0\u30e9\u30d5\u30a3\u30fc\u3067\u306f\u3057\u3070\u3057\u3070\u6210\u7acb\u3057\u306a\u3044\u304b\u3089\u3067\u3059\u3002<\/p>\n\n\n\n

    Furthermore, designers who must make decisions on what variables to include or exclude can impart their own cognitive biases into the model<\/strong> [112, 184].<\/p>\n\n\n\n

    \u3055\u3089\u306b\u3001\u3069\u306e\u5909\u6570\u3092\u542b\u3081\u308b\u304b\u3001\u3042\u308b\u3044\u306f\u9664\u5916\u3059\u308b\u304b\u3092\u6c7a\u5b9a\u3059\u308b\u30e2\u30c7\u30eb\u306e\u8a2d\u8a08\u8005\u81ea\u8eab\u306e\u8a8d\u77e5\u30d0\u30a4\u30a2\u30b9\u304c\u3001\u30e2\u30c7\u30eb\u306b\u30d0\u30a4\u30a2\u30b9\u3092\u3082\u305f\u3089\u3059<\/strong>\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059[112, 184]\u3002<\/p>\n\n\n\n

    Complex models<\/strong> are often used on nonlinear, multimodal data such as text and images. Such models may capture latent systemic bias<\/strong> in ways that are difficult to recognize and predict.<\/p>\n\n\n\n

    \u4e00\u65b9\u3067\u3001\u8907\u96d1\u306a\u30e2\u30c7\u30eb\u306f\u3001\u30c6\u30ad\u30b9\u30c8\u3084\u753b\u50cf\u306a\u3069\u306e\u975e\u7dda\u5f62\u3067\u30de\u30eb\u30c1\u30e2\u30fc\u30c0\u30eb\u306a\u30c7\u30fc\u30bf\u3067\u3088\u304f\u4f7f\u7528\u3055\u308c\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306a\u30e2\u30c7\u30eb\u306b\u306f\u3001\u6f5c\u5728\u7684\u306a\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u3092\u3001\u8b58\u5225\u3084\u4e88\u671f\u3057\u306b\u304f\u3044\u5f62\u3067\u5185\u5305\u3057\u3066\u3057\u307e\u3046\u30ea\u30b9\u30af\u304c\u3042\u308a\u307e\u3059\u3002<\/strong><\/p>\n\n\n\n

    \n

    Validity <\/strong>\u59a5\u5f53\u6027<\/p>\n<\/blockquote>\n\n\n\n

    Many difficulties and flaws can arise in system validation. A common challenge in system testing is a lack of ground truth, or noisy labeling and other annotation factors which make it difficult to know what is accurate.<\/p>\n\n\n\n

    \u30b7\u30b9\u30c6\u30e0\u306e\u691c\u8a3c\u3067\u306f\u3001\u591a\u304f\u306e\u56f0\u96e3\u3084\u6b20\u9665\u304c\u751f\u3058\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u30b7\u30b9\u30c6\u30e0\u30c6\u30b9\u30c8\u306b\u304a\u3051\u308b\u5171\u901a\u306e\u8ab2\u984c\u306f\u3001\u30b0\u30e9\u30f3\u30c9\u30c8\u30a5\u30eb\u30fc\u30b9\u306e\u6b20\u5982\u3001\u3042\u308b\u3044\u306f\u30ce\u30a4\u30ba\u306e\u591a\u3044\u30e9\u30d9\u30ea\u30f3\u30b0\u3084\u305d\u306e\u4ed6\u306e\u30a2\u30ce\u30c6\u30fc\u30b7\u30e7\u30f3\u8981\u56e0\u306b\u3088\u3063\u3066\u3001\u4f55\u304c\u6b63\u78ba\u306a\u306e\u304b\u3092\u77e5\u308b\u3053\u3068\u304c\u56f0\u96e3\u3067\u3042\u308b\u3053\u3068\u3067\u3059\u3002<\/p>\n\n\n\n

    The use of proxy variables<\/strong> compounds this difficulty, since what is being measured isn\u2019t directly observable. Performing system tests under optimal conditions \u2014 or conditions that are not close to the deployed state \u2014 is another challenging design flaw.<\/p>\n\n\n\n

    \u30d7\u30ed\u30ad\u30b7\u5909\u6570<\/strong>\uff08\u8a33\u6ce8\uff1a\u76f4\u63a5\u6e2c\u5b9a\u3067\u304d\u306a\u3044\u5909\u6570\u306b\u5bfe\u3057\u3066\u5229\u7528\u3059\u308b\u4ee3\u7406\u5909\u6570\uff09\u306e\u4f7f\u7528\u306f\u3001\u6e2c\u5b9a\u5bfe\u8c61\u304c\u76f4\u63a5\u89b3\u6e2c\u53ef\u80fd\u3067\u306a\u3044\u305f\u3081\u3001\u3053\u306e\u56f0\u96e3\u3092\u3055\u3089\u306b\u60aa\u5316\u3055\u305b\u307e\u3059\u3002\u30b7\u30b9\u30c6\u30e0\u30c6\u30b9\u30c8\u3092\u6700\u9069\u306a\u6761\u4ef6\u3001\u3064\u307e\u308a\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u305f\u72b6\u614b\u306b\u8fd1\u304f\u306a\u3044\u6761\u4ef6\u3067\u884c\u3046\u3053\u3068\u306f\u3001\u3055\u3089\u306b\u8a2d\u8a08\u4e0a\u306e\u6b20\u9665\u3092\u751f\u3058\u307e\u3059\u3002<\/p>\n\n\n\n

    Also the practice of \u201cstratified performance evaluations,\u201d [103] where system performance is analyzed across segments in the training or test data, whether demographic segments or otherwise<\/strong>, is a basic consideration for understanding system validity across a population of users.<\/p>\n\n\n\n

    \u307e\u305f\u300c\u30c7\u30fc\u30bf\u30bb\u30b0\u30e1\u30f3\u30c8\u5225\u306e\u6027\u80fd\u8a55\u4fa1\u300d[103]\u306e\u5b9f\u8df5\u306b\u304a\u3044\u3066\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf\u307e\u305f\u306f\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u306e\u30bb\u30b0\u30e1\u30f3\u30c8\uff08\u4eba\u53e3\u7d71\u8a08\u5b66\u7684\u30bb\u30b0\u30e1\u30f3\u30c8\u307e\u305f\u306f\u305d\u306e\u4ed6\uff09\u306b\u308f\u305f\u3063\u3066\u30b7\u30b9\u30c6\u30e0\u6027\u80fd\u3092\u5206\u6790<\/strong>\u3057\u307e\u3059\u304c\u3001\u3053\u308c\u306f\u30e6\u30fc\u30b6\u30fc\u6bcd\u96c6\u56e3\u5168\u4f53\u306b\u308f\u305f\u3063\u3066\u30b7\u30b9\u30c6\u30e0\u304c\u59a5\u5f53\u3067\u3042\u308b\u304b\u5426\u304b\u3092\u7406\u89e3\u3059\u308b\u305f\u3081\u306b\u91cd\u8981\u306a\u691c\u8a0e\u4e8b\u9805\u3067\u3059\u3002<\/p>\n\n\n\n

    \n

    Validation and deployment <\/strong>\u30d0\u30ea\u30c7\u30fc\u30b7\u30e7\u30f3\u3068\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8<\/p>\n<\/blockquote>\n\n\n\n

    Validation also means ensuring that the system is not being used in unintended ways.<\/p>\n\n\n\n

    \u30d0\u30ea\u30c7\u30fc\u30b7\u30e7\u30f3\u3068\u306f\u3001\u30b7\u30b9\u30c6\u30e0\u304c\u610f\u56f3\u3057\u306a\u3044\u4f7f\u308f\u308c\u65b9\u3092\u3057\u3066\u3044\u306a\u3044\u3053\u3068\u3092\u78ba\u8a8d\u3059\u308b\u3053\u3068\u3067\u3082\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    DEPLOYMENT BIAS <\/strong>happens when an AI model is used in ways not intended by developers.<\/p>\n\n\n\n

    \u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u30d0\u30a4\u30a2\u30b9<\/strong>\uff08\u8a33\u6ce8\uff1a\u30e2\u30c7\u30eb\u958b\u767a\u6642\u306b\u60f3\u5b9a\u3057\u3066\u3044\u305f\u3053\u3068\u3068\u7570\u306a\u308b\u76ee\u7684\u3067\u3001AI\u306e\u51fa\u529b\u7d50\u679c\u3092\u5229\u7528\u3057\u3066\u3057\u307e\u3046\u3053\u3068\u306b\u8d77\u56e0\u3059\u308b\u30d0\u30a4\u30a2\u30b9\uff09\u306f\u3001AI\u30e2\u30c7\u30eb\u304c\u958b\u767a\u8005\u306e\u610f\u56f3\u3057\u306a\u3044\u65b9\u6cd5\u3067\u4f7f\u7528\u3055\u308c\u305f\u3068\u304d<\/strong>\u306b\u8d77\u3053\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Emergent bias <\/strong>happens where the model is used in unanticipated contexts.<\/p>\n\n\n\n

    \u30a4\u30de\u30fc\u30b8\u30a7\u30f3\u30c8\u30d0\u30a4\u30a2\u30b9<\/strong>\uff08\u8a33\u6ce8\uff1a\u30e2\u30c7\u30eb\u958b\u767a\u6642\u306e\u8a2d\u5b9a\u3068\u306f\u7570\u306a\u308b\u30c9\u30e1\u30a4\u30f3\u5206\u91ce\u3067\u3001AI\u304c\u4f7f\u7528\u3055\u308c\u3066\u3057\u307e\u3046\u3053\u3068\u306b\u8d77\u56e0\u3059\u308b\u30d0\u30a4\u30a2\u30b9\u3002\u591a\u304f\u306e\u30b1\u30fc\u30b9\u3067\u306f\u7cbe\u5ea6\u52a3\u5316\u306b\u7e4b\u304c\u308b\uff09\u306f\u3001\u30e2\u30c7\u30eb\u304c\u4e88\u671f\u305b\u306c\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3067\u4f7f\u7528\u3055\u308c\u305f\u5834\u5408<\/strong>\u306b\u8d77\u3053\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Developers of an algorithm used by major U.S. cities to assist in coordinating housing to homeless people began phasing it out after several cities inappropriately used the algorithm as an assessment tool rather than as the presecreening tool as it was designed[185].<\/p>\n\n\n\n

    \u7c73\u56fd\u306e\u4e3b\u8981\u90fd\u5e02\u306b\u304a\u3044\u3066\u3001\u30db\u30fc\u30e0\u30ec\u30b9\u3078\u306e\u4f4f\u5c45\u63d0\u4f9b\u3092\u652f\u63f4\u3059\u308b\u305f\u3081\u306b\u958b\u767a\u3055\u308c\u305f\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u304c\u3001\u3044\u304f\u3064\u304b\u306e\u90fd\u5e02\u306b\u304a\u3044\u3066\u3001\u60f3\u5b9a\u3057\u3066\u3044\u305f\u4e88\u5099\u5be9\u67fb\u30c4\u30fc\u30eb\u3068\u3057\u3066\u3067\u306f\u306a\u304f\u3001\u4e0d\u9069\u5207\u306a\u8a55\u4fa1\u30c4\u30fc\u30eb\u3068\u3057\u3066\u4f7f\u7528\u3055\u308c\u305f\u3053\u3068\u304b\u3089\u3001\u6bb5\u968e\u7684\u306b\u4f7f\u7528\u3092\u4e2d\u6b62\u3059\u308b\u3053\u3068\u306b\u306a\u308a\u307e\u3057\u305f[185]\u3002<\/p>\n\n\n\n

    In another instance, the Chicago Police Department decommissioned an algorithm designed to predict the risk that an individual might be involved in future gun violence, citing unintended use and misapplication of the model [186].<\/p>\n\n\n\n

    \u5225\u306e\u4f8b\u3067\u306f\u3001\u30b7\u30ab\u30b4\u8b66\u5bdf\u306b\u304a\u3044\u3066\u3001\u3042\u308b\u500b\u4eba\u304c\u5c06\u6765\u9283\u4e71\u5c04\u4e8b\u4ef6\u3092\u5f15\u304d\u8d77\u3053\u3059\u5371\u967a\u6027\u3092\u4e88\u6e2c\u3059\u308b\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u958b\u767a\u3057\u307e\u3057\u305f\u304c\u3001\u305d\u308c\u304c\u610f\u56f3\u3057\u306a\u3044\u5f62\u3067\u4f7f\u7528\u3055\u308c\u305f\u308a\u8aa4\u3063\u3066\u9069\u7528\u3055\u308c\u308b\u30b1\u30fc\u30b9\u304c\u3042\u3063\u305f\u3068\u3057\u3066\u5ec3\u6b62\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

    Without system validation, an AI system could be released that is technically flawed or fails to establish appropriate underlying mechanisms for proper functioning [188\u2013190].<\/p>\n\n\n\n

    \u30b7\u30b9\u30c6\u30e0\u691c\u8a3c\u304c\u884c\u308f\u308c\u306a\u3051\u308c\u3070\u3001\u6280\u8853\u7684\u306b\u6b20\u9665\u306e\u3042\u308b\u3001\u3042\u308b\u3044\u306f\u6b63\u3057\u304f\u6a5f\u80fd\u3059\u308b\u305f\u3081\u306e\u9069\u5207\u306a\u57fa\u790e\u7684\u30e1\u30ab\u30cb\u30ba\u30e0\u3092\u78ba\u7acb\u3067\u304d\u306a\u3044\u307e\u307e\u3001AI\u30b7\u30b9\u30c6\u30e0\u304c\u30ea\u30ea\u30fc\u30b9\u3055\u308c\u3066\u3057\u307e\u3046\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059 [188\u2013190]\u3002<\/p>\n\n\n\n

    \n

    AI systems as magic <\/strong>\u9b54\u6cd5\u306e\u3088\u3046\u306aAI\u30b7\u30b9\u30c6\u30e0<\/p>\n<\/blockquote>\n\n\n\n

    High-level machine learning libraries and reduced costs of cloud computing have made AI more affordable and easier to develop. As a result, AI development is becoming increasingly democratized. Still, AI itself remains largely opaque <\/strong>\u2014 deep neural networks and Bayesian inference require advanced mathematics to understand.<\/p>\n\n\n\n

    \u30cf\u30a4\u30ec\u30d9\u30eb\u306a\u6a5f\u68b0\u5b66\u7fd2\u30e9\u30a4\u30d6\u30e9\u30ea\u30fc\u3068\u30af\u30e9\u30a6\u30c9\u30b3\u30f3\u30d4\u30e5\u30fc\u30c6\u30a3\u30f3\u30b0\u306e\u30b3\u30b9\u30c8\u524a\u6e1b\u306b\u3088\u308a\u3001AI\u306f\u3088\u308a\u5b89\u4fa1\u3067\u5bb9\u6613\u306b\u958b\u767a\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3057\u305f\u3002\u305d\u306e\u7d50\u679c\u3001AI\u958b\u767a\u306f\u307e\u3059\u307e\u3059\u6c11\u4e3b\u5316\u3055\u308c\u3066\u304d\u3066\u3044\u307e\u3059\u3002\u3057\u304b\u3057\u306a\u304c\u3089\u3001\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3084\u30d9\u30a4\u30ba\u63a8\u5b9a\u3092\u7406\u89e3\u3059\u308b\u306b\u306f\u9ad8\u5ea6\u306a\u6570\u5b66\u304c\u5fc5\u8981\u3067\u3042\u308a\u3001AI\u305d\u306e\u3082\u306e\u306f\u307e\u3060\u4e0d\u900f\u660e\u306a\u90e8\u5206\u304c\u591a\u3044<\/strong>\u306e\u3082\u4e8b\u5b9f\u3067\u3059\u3002<\/p>\n\n\n\n

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

    3.2.2 TEVV Guidance<\/strong><\/p>\n\n\n\n

    TEVV<\/em>(Test, Evaluation, Validation, and Verification)\u30ac\u30a4\u30c0\u30f3\u30b9<\/em><\/p>\n\n\n\n

    To mitigate the risks stemming from epistemic and aleatoric uncertainties, model developers should work closely with the organizations deploying them. Teams should work to ensure periodic model updates, and test and recalibrate model parameters on updated representative datasets<\/strong> to meet the business objectives while staying within desired performance targets and acceptable levels of bias.<\/p>\n\n\n\n

    Epistemic\u3042\u308b\u3044\u306fAleatoric\u4e0d\u78ba\u5b9f\u6027\u306b\u8d77\u56e0\u3059\u308b\u30ea\u30b9\u30af\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306b\u3001\u30e2\u30c7\u30eb\u958b\u767a\u8005\u306f\u305d\u308c\u3092\u5c0e\u5165\u3059\u308b\u7d44\u7e54\u3068\u7dca\u5bc6\u306b\u9023\u643a\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u30c1\u30fc\u30e0\u306f\u3001\u5b9a\u671f\u7684\u306a\u30e2\u30c7\u30eb\u306e\u66f4\u65b0<\/strong>\u3092\u78ba\u5b9f\u306b\u884c\u3044\u3001\u66f4\u65b0\u3055\u308c\u305f\u4ee3\u8868\u7684\u306a\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u30e2\u30c7\u30eb\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30c6\u30b9\u30c8\u3068\u518d\u30ad\u30e3\u30ea\u30d6\u30ec\u30fc\u30b7\u30e7\u30f3<\/strong>\u3092\u884c\u3044\u3001\u671b\u307e\u3057\u3044\u6027\u80fd\u76ee\u6a19\u3068\u8a31\u5bb9\u3067\u304d\u308b\u30ec\u30d9\u30eb\u306e\u30d0\u30a4\u30a2\u30b9\u306e\u7bc4\u56f2\u5185\u3067\u3001\u30d3\u30b8\u30cd\u30b9\u76ee\u6a19\u3092\u9054\u6210\u3059\u308b\u3088\u3046\u306b\u52aa\u529b\u3059\u308b\u3079\u304d\u3067\u3059\u3002<\/p>\n\n\n\n

    \n

    Algorithms <\/strong>\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0<\/p>\n<\/blockquote>\n\n\n\n

    In ML, it is not meaningful to assign bias to the model or algorithm itself without contextual information about the specific tasks on which they may be used.<\/strong> This links the model and algorithm to the dataset on which they are trained and tested<\/p>\n\n\n\n

    \u6a5f\u68b0\u5b66\u7fd2\u306b\u304a\u3044\u3066\u3001\u30e2\u30c7\u30eb\u3084\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u304c\u4f7f\u7528\u3055\u308c\u308b\u5177\u4f53\u7684\u306a\u30bf\u30b9\u30af\u306b\u95a2\u3059\u308b\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u60c5\u5831<\/strong>\u306a\u3057\u306b\u3001\u30d0\u30a4\u30a2\u30b9\u3092\u5b9a\u7fa9\u3059\u308b\u3053\u3068\u306f\u610f\u5473\u304c\u3042\u308a\u307e\u305b\u3093\u3002\u3053\u3046\u3057\u305f\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u304c\u3001\u30e2\u30c7\u30eb\u3084\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u5b66\u7fd2\u30fb\u30c6\u30b9\u30c8\u3059\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3068\u95a2\u9023\u4ed8\u3051\u308b\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

    When considering approaches to mitigating algorithmic bias in a specific task context, recent literature categorizes debiasing methods into one of three categories [62, 86, 195, 198]:<\/p>\n\n\n\n

    \u7279\u5b9a\u306e\u30bf\u30b9\u30af\u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306b\u304a\u3044\u3066\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u30d0\u30a4\u30a2\u30b9\u3092\u8efd\u6e1b\u3059\u308b\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u691c\u8a0e\u3059\u308b\u5834\u5408\u3001 \u6700\u8fd1\u306e\u6587\u732e\u3067\u306f\u3001\u305d\u306e\u624b\u6cd5\u3092\u4ee5\u4e0b3\u3064\u306e\u30ab\u30c6\u30b4\u30ea\u306b\u5206\u985e\u3057\u3066\u3044\u307e\u3059 [62, 86, 195, 198]\u3002<\/p>\n\n\n\n

    1. Pre-processing:<\/strong><\/p>\n\n\n\n

    transforming the data so that the underlying discrimination is mitigated. This method can be used if a modeling pipeline is allowed to modify the training data.<\/p>\n\n\n\n

    \u6839\u672c\u7684\u306a\u5dee\u5225\u304c\u7de9\u548c\u3055\u308c\u308b\u3088\u3046\u306b\u30c7\u30fc\u30bf\u3092\u5909\u63db\u3059\u308b\u3002\u3053\u306e\u65b9\u6cd5\u306f\u3001\u30e2\u30c7\u30ea\u30f3\u30b0\u30d1\u30a4\u30d7\u30e9\u30a4\u30f3\u304c\u5b66\u7fd2\u30c7\u30fc\u30bf\u3092\u4fee\u6b63\u3059\u308b\u3053\u3068\u3092\u8a31\u53ef\u3055\u308c\u3066\u3044\u308b\u5834\u5408\u306b\u4f7f\u7528\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3002<\/p>\n\n\n\n

    2. In-processing:<\/strong><\/p>\n\n\n\n

    techniques that modify the algorithms in order to mitigate bias during model training. Model training processes could incorporate changes to the objective (cost) function or impose a new optimization constraint.<\/p>\n\n\n\n

    \u30e2\u30c7\u30eb\u5b66\u7fd2\u4e2d\u306e\u30d0\u30a4\u30a2\u30b9\u3092\u8efd\u6e1b\u3059\u308b\u305f\u3081\u306b\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u4fee\u6b63\u3059\u308b\u6280\u8853\u3002\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u904e\u7a0b\u3067\u306f\u3001\u76ee\u7684\u95a2\u6570\uff08\u640d\u5931\u95a2\u6570\uff09\u306e\u5909\u66f4\u3084\u3001\u65b0\u305f\u306a\u6700\u9069\u5316\u5236\u7d04\u3092\u8ab2\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3002<\/p>\n\n\n\n

    3. Post-processing:<\/strong><\/p>\n\n\n\n

    typically performed with the help of a holdout dataset (data not used in the training of the model). Here, the learned model is treated as a black box and its predictions are altered by a function during the post-processing phase. The function is deduced from the performance of the black box model on the holdout dataset. This technique may be useful in adapting a pre-trained large language model to a dataset and task of interest.<\/p>\n\n\n\n

    \u901a\u5e38\u3001\u30db\u30fc\u30eb\u30c9\u30a2\u30a6\u30c8\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\uff08\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u4f7f\u7528\u3057\u306a\u304b\u3063\u305f\u30c7\u30fc\u30bf\uff09\u306e\u52a9\u3051\u3092\u501f\u308a\u3066\u884c\u3046\u3002\u3053\u3053\u3067\u306f\u3001\u5b66\u7fd2\u3057\u305f\u30e2\u30c7\u30eb\u3092\u30d6\u30e9\u30c3\u30af\u30dc\u30c3\u30af\u30b9\u3068\u3057\u3066\u6271\u3044\u3001\u305d\u306e\u4e88\u6e2c\u5024\u3092\u5f8c\u51e6\u7406\u306e\u6bb5\u968e\u3067\u3001\u3042\u308b\u95a2\u6570\u306b\u3088\u3063\u3066\u5909\u63db\u3059\u308b\u3002\u3053\u306e\u95a2\u6570\u306f\u3001\u30db\u30fc\u30eb\u30c9\u30a2\u30a6\u30c8\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u304a\u3051\u308b\u30d6\u30e9\u30c3\u30af\u30dc\u30c3\u30af\u30b9\u30e2\u30c7\u30eb\u306e\u6027\u80fd\u304b\u3089\u5c0e\u304d\u51fa\u3055\u308c\u308b\u3002\u3053\u306e\u624b\u6cd5\u306f\u3001\u4e8b\u524d\u306b\u5b66\u7fd2\u3055\u308c\u305f\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\u3092\u3001\u95a2\u5fc3\u306e\u3042\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3084\u30bf\u30b9\u30af\u306b\u9069\u5fdc\u3055\u305b\u308b\u969b\u306a\u3069\u306b\u6709\u7528\u3067\u3042\u308b\u3068\u8003\u3048\u3089\u308c\u308b\u3002<\/p>\n\n\n\n

    \n

    Fairness metrics <\/strong>\u30d5\u30a7\u30a2\u30cd\u30b9\uff08\u516c\u5e73\u6027\uff09\u30e1\u30c8\u30ea\u30af\u30b9<\/p>\n<\/blockquote>\n\n\n\n

    The plethora of fairness metric definitions illustrates that fairness cannot be reduced to a concise mathematical definition. Fairness is dynamic, social in nature, application and context specific, and not just an abstract or universal statistical problem. Therefore, it is important to adopt a socio-technical approach to fairness in order to have realistic fairness definitions for different contexts as well as task-specific datasets for machine learning model development and evaluation.<\/p>\n\n\n\n

    \u516c\u5e73\u6027\u306e\u6307\u6a19\u306e\u5b9a\u7fa9\u304c\u6570\u591a\u304f\u3042\u308b\u3053\u3068\u306f\u3001\u516c\u5e73\u3055\u3092\u7c21\u6f54\u306a\u6570\u5b66\u7684\u5b9a\u7fa9\u306b\u9084\u5143\u3067\u304d\u306a\u3044\u3053\u3068\u3092\u7269\u8a9e\u3063\u3066\u3044\u307e\u3059\u3002\u516c\u5e73\u6027\u306f\u3001\u30c0\u30a4\u30ca\u30df\u30c3\u30af\u3067\u793e\u4f1a\u7684\u306a\u6027\u8cea\u3092\u6301\u3063\u3066\u304a\u308a\u3001\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3084\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306b\u4f9d\u5b58\u3057\u3001\u62bd\u8c61\u7684\u307e\u305f\u306f\u666e\u904d\u7684\u306a\u7d71\u8a08\u7684\u554f\u984c\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u3057\u305f\u304c\u3063\u3066\u3001\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u306e\u958b\u767a\u3068\u8a55\u4fa1\u306e\u305f\u3081\u306b\u3001\u30bf\u30b9\u30af\u306b\u7279\u5316\u3057\u305f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3092\u6301\u3064\u306e\u3068\u540c\u69d8\u3001\u7570\u306a\u308b\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u306b\u5bfe\u3059\u308b\u73fe\u5b9f\u7684\u306a\u516c\u5e73\u6027\u306e\u5b9a\u7fa9\u3092\u5f97\u308b\u305f\u3081\u306b\u3001\u516c\u5e73\u6027\u306b\u5bfe\u3057\u3066\u793e\u4f1a\u6280\u8853\u7684\u306a\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u63a1\u7528\u3059\u308b\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002<\/p>\n\n\n\n

    3.3 Who makes decisions and how do they make them? Human Factors in AI Bias<\/strong><\/p>\n\n\n\n

    \u610f\u601d\u6c7a\u5b9a\u306f\u8ab0\u304c\u3001\u3069\u306e\u3088\u3046\u306b\u884c\u3046\u306e\u304b\uff1fAI\u30d0\u30a4\u30a2\u30b9\u306e\u4eba\u7684\u8981\u56e0<\/em><\/p>\n\n\n\n

    3.3.1 Human Factors Challenges <\/strong>\u4eba\u7684\u8981\u56e0\u306e\u8ab2\u984c<\/em><\/p>\n\n\n\n

    Computational decision support systems augment another, typically human, system in making decisions.<\/p>\n\n\n\n

    \u30b3\u30f3\u30d4\u30e5\u30fc\u30bf\u306b\u3088\u308b\u610f\u601d\u6c7a\u5b9a\u652f\u63f4\u30b7\u30b9\u30c6\u30e0\u306f\u3001\u610f\u601d\u6c7a\u5b9a\u3092\u884c\u3046\u969b\u3001\u4ed6\u306e\u30b7\u30b9\u30c6\u30e0\uff08\u901a\u5e38\u306f\u4eba\u9593\uff09\u3092\u88dc\u5b8c\u3059\u308b\u305f\u3081\u306b\u4f7f\u7528\u3055\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

    Comparatively, for algorithmic decision systems there is less human involvement, with the AI system itself more in the \u201cdriver\u2019s seat,\u201d and able to produce outcomes with little human involvement to govern the impact.<\/p>\n\n\n\n

    \u4e00\u65b9\u3001\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306b\u3088\u308b\u610f\u601d\u6c7a\u5b9a\u30b7\u30b9\u30c6\u30e0\u3067\u306f\u3001\u4eba\u9593\u306e\u95a2\u4e0e\u306f\u5c11\u306a\u304f\u3001AI\u30b7\u30b9\u30c6\u30e0\u81ea\u4f53\u304c\u300c\u904b\u8ee2\u5e2d\u300d\u306b\u5ea7\u308a\u3001\u5f71\u97ff\u3092\u7ba1\u7406\u3059\u308b\u305f\u3081\u306b\u4eba\u9593\u304c\u307b\u3068\u3093\u3069\u95a2\u4e0e\u3057\u306a\u304f\u3066\u3082\u7d50\u679c\u3092\u51fa\u3059\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n

    Organizations that deploy AI models and systems without assessing and managing these risks can not only harm their users but jeopardize their reputations<\/strong>.<\/p>\n\n\n\n

    \u3053\u308c\u3089\u306e\u30ea\u30b9\u30af\u3092\u8a55\u4fa1\u30fb\u7ba1\u7406\u305b\u305a\u306bAI\u30e2\u30c7\u30eb\u3084\u30b7\u30b9\u30c6\u30e0\u3092\u30c7\u30d7\u30ed\u30a4\u3059\u308b\u7d44\u7e54\u306f\u3001\u30e6\u30fc\u30b6\u306b\u5b9f\u5bb3\u3082\u9f4e\u3059\u3060\u3051\u3067\u306a\u304f\u3001\u30ec\u30d4\u30e5\u30c6\u30fc\u30b7\u30e7\u30f3\u30ea\u30b9\u30af\u306b\u6652\u3055\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/strong><\/p>\n\n\n\n

    \n

    Deployment Context of Use <\/strong>\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u6642\u306e\u4f7f\u7528\u74b0\u5883<\/p>\n<\/blockquote>\n\n\n\n

    AI systems are designed and developed to be used in specific real world settings, but are often tested in idealized scenarios<\/strong>. Once deployed, the original intent, idea, or impact assessment can drift<\/strong> as the application is repurposed or used in unforeseen ways, and in settings or contexts for which it was not originally intended. Different deployment contexts means a new set of risks to be considered.<\/strong><\/p>\n\n\n\n

    AI\u30b7\u30b9\u30c6\u30e0\u306f\u3001\u73fe\u5b9f\u306e\u7279\u5b9a\u306e\u74b0\u5883\u3067\u4f7f\u7528\u3055\u308c\u308b\u3053\u3068\u3092\u524d\u63d0\u306b\u8a2d\u8a08\u30fb\u958b\u767a\u3055\u308c\u307e\u3059\u304c\u3001\u4e00\u65b9\u3067\u3001\u7406\u60f3\u5316\u3055\u308c\u305f\u30b7\u30ca\u30ea\u30aa\u306e\u5143\u3067\u30c6\u30b9\u30c8\u3055\u308c\u308b\u3053\u3068\u3082\u5c11\u306a\u304f\u3042\u308a\u307e\u305b\u3093<\/strong>\u3002\u4e00\u65e6\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u308b\u3068\u3001\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u304c\u5225\u306e\u76ee\u7684\u3067\u518d\u5229\u7528\u3055\u308c\u305f\u308a\u3001\u4e88\u671f\u305b\u306c\u65b9\u6cd5\u3067\u4f7f\u7528\u3055\u308c\u305f\u308a\u3001\u5143\u3005\u610f\u56f3\u3055\u308c\u3066\u3044\u306a\u304b\u3063\u305f\u8a2d\u5b9a\u3084\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3067\u4f7f\u7528\u3055\u308c\u308b\u305f\u3081\u3001\u8a2d\u8a08\u5f53\u521d\u306e\u610f\u56f3\u3001\u30a2\u30a4\u30c7\u30a2\u3001\u5f71\u97ff\u8a55\u4fa1\u304b\u3089\u5916\u308c\u3066\u304f\u308b<\/strong>\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u306e\u4ed5\u65b9\u304c\u7570\u306a\u308b\u3068\u3044\u3046\u3053\u3068\u306f\u3001\u65b0\u3057\u3044\u3055\u307e\u3056\u307e\u306a\u30ea\u30b9\u30af\u3082\u8003\u616e\u3059\u3079\u304d\u3060\u3068\u3044\u3046\u3053\u3068\u3067\u3059\u3002<\/strong><\/p>\n\n\n\n

    \n

    Human-in-the-loop <\/strong>\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7<\/p>\n<\/blockquote>\n\n\n\n

    The default assumption is that placing a human \u201cin-the-loop\u201d of such systems can ensure that adverse events do not occur. Current perceptions about the role and responsibility of the human-in-the-loop with AI are often implicit, and expectations about level of performance for these systems are often based on untested or outdated hypotheses.<\/p>\n\n\n\n

    \u3053\u306e\u3088\u3046\u306a\u30b7\u30b9\u30c6\u30e0\u306e\u30eb\u30fc\u30d7\u5185\u306b\u4eba\u9593\u3092\u914d\u7f6e\u3059\u308b\u3053\u3068\u3067\u3001\u6709\u5bb3\u306a\u4e8b\u8c61\u304c\u767a\u751f\u3057\u306a\u3044\u3088\u3046\u306b\u3067\u304d\u308b\u3068\u3044\u3046\u306e\u304c\u3001\u30c7\u30d5\u30a9\u30eb\u30c8\u306e\u524d\u63d0\u3068\u3057\u3066\u8003\u3048\u3089\u308c\u3066\u3044\u307e\u3059\u3002\u4e00\u65b9\u3067AI\u306b\u304a\u3051\u308b\u4eba\u9593\u306e\u5f79\u5272\u3068\u8cac\u4efb\u306b\u95a2\u3059\u308b\u73fe\u5728\u306e\u8a8d\u8b58\u306f\u3001\u3057\u3070\u3057\u3070\u6697\u9ed9\u7684\u3067\u3042\u308a\u3001\u3053\u3046\u3057\u305f\u30b7\u30b9\u30c6\u30e0\u306e\u6027\u80fd\u30ec\u30d9\u30eb\u306b\u95a2\u3059\u308b\u671f\u5f85\u306f\u3001\u672a\u691c\u8a3c\u307e\u305f\u306f\u6642\u4ee3\u9045\u308c\u306e\u4eee\u8aac\u306b\u57fa\u3065\u3044\u3066\u3044\u308b\u3053\u3068\u304c\u591a\u3044\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

    The reality however is that without significant procedural and cultural support, optimistic expectations about how humans are able to serve in this administrative capacity are not borne out in practice.<\/p>\n\n\n\n

    \u73fe\u5b9f\u7684\u306b\u306f\u3001\u3057\u3063\u304b\u308a\u3068\u78ba\u7acb\u3055\u308c\u305f\u624b\u7d9a\u304d\u53ca\u3073\u6587\u5316\u7684\u306a\u30b5\u30dd\u30fc\u30c8\u304c\u306a\u3051\u308c\u3070\u3001\u4eba\u9593\u304c\u3053\u3046\u3057\u305f\u7ba1\u7406\u80fd\u529b\u3092\u767a\u63ee\u3067\u304d\u308b\u3068\u3044\u3046\u697d\u89b3\u7684\u306a\u671f\u5f85\u306f\u3001\u7c21\u5358\u306b\u88cf\u5207\u3089\u308c\u308b\u3053\u3068\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    Reliance on various downstream professionals to act as a governor on automated processes in complex societal systems is not a viable approach<\/p>\n\n\n\n

    \u8907\u96d1\u306a\u793e\u4f1a\u30b7\u30b9\u30c6\u30e0\u306e\u81ea\u52d5\u5316\u30d7\u30ed\u30bb\u30b9\u306b\u304a\u3044\u3066\u3001\u3055\u307e\u3056\u307e\u306a\u7523\u696d\u5206\u91ce\u306e\u5c02\u9580\u5bb6\u306b\u3001\u30ac\u30d0\u30ca\u30fc\u3068\u3057\u3066\u306e\u6a5f\u80fd\u3092\u4f9d\u5b58\u3059\u308b\u3053\u3068\u306f\u3001\u5b9f\u884c\u53ef\u80fd\u306a\u30a2\u30d7\u30ed\u30fc\u30c1\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002<\/p>\n\n\n\n

    3.3.2 Human Factors Guidance <\/strong>\u4eba\u7684\u8981\u56e0\u306b\u95a2\u308f\u308b\u30ac\u30a4\u30c0\u30f3\u30b9<\/em><\/p>\n\n\n\n

    \n

    Human\u2013AI configuration<\/strong> \u4eba\u9593\u3068AI\u3092\u7d61\u3081\u305f\u30b7\u30b9\u30c6\u30e0\u8a2d\u8a08<\/p>\n<\/blockquote>\n\n\n\n

    NIST seeks to develop formal guidance about how to implement human-in-the-loop processes<\/strong> that do not amplify or perpetuate the many human, systemic and computational biases that can degrade outcomes in this complex setting. Identifying system configurations and necessary qualifications for their components that result in outcomes that are accurate and trustworthy will be a key focus.<\/p>\n\n\n\n

    NIST\u306f\u3001\u3053\u306e\u8907\u96d1\u306a\u74b0\u5883\u306b\u304a\u3044\u3066\u3001\u7d50\u679c\u3092\u60aa\u5316\u3055\u305b\u308b\u591a\u304f\u306e\u4eba\u7684\u30d0\u30a4\u30a2\u30b9\u3001\u30b7\u30b9\u30c6\u30df\u30c3\u30af\u30d0\u30a4\u30a2\u30b9\u3001\u3042\u308b\u3044\u306f\u8a08\u6570\u7684\u306a\u30d0\u30a4\u30a2\u30b9\u3092\u5897\u5e45\u3057\u305f\u308a\u6c38\u7d9a\u3055\u305b\u306a\u3044\u305f\u3081\u306e\u3001\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7\u306e\u30d7\u30ed\u30bb\u30b9<\/strong>\u306e\u6b63\u5f0f\u306a\u5b9f\u88c5\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u958b\u767a\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u3066\u3044\u307e\u3059\u3002\u6b63\u78ba\u3067\u4fe1\u983c\u3067\u304d\u308b\u7d50\u679c\u3092\u5f97\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u30b7\u30b9\u30c6\u30e0\u69cb\u6210\u3068\u3001\u305d\u306e\u69cb\u6210\u8981\u7d20\u306b\u5fc5\u8981\u306a\u8981\u4ef6\u3092\u7279\u5b9a\u3059\u308b\u3053\u3068\u304c\u3001\u91cd\u8981\u306a\u7126\u70b9\u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    \n

    Keeping humans at the center of AI design <\/strong>AI\u8a2d\u8a08\u306e\u4e2d\u5fc3\u306b\u4eba\u9593\u3092\u636e\u3048\u308b<\/p>\n<\/blockquote>\n\n\n\n

    Human-centered design (HCD) is an approach to the design and development of a system or technology that aims to improve the ability of users to effectively and efficiently use a product.<\/p>\n\n\n\n

    \u4eba\u9593\u4e2d\u5fc3\u306e\u8a2d\u8a08\uff08HCD\uff09\u3068\u306f\u3001\u30b7\u30b9\u30c6\u30e0\u3084\u6280\u8853\u306e\u8a2d\u8a08\u30fb\u958b\u767a\u306b\u304a\u3044\u3066\u3001\u30e6\u30fc\u30b6\u30fc\u304c\u88fd\u54c1\u3092\u52b9\u679c\u7684\u30fb\u52b9\u7387\u7684\u306b\u4f7f\u7528\u3067\u304d\u308b\u3088\u3046\u306b\u3059\u308b\u3053\u3068\u3092\u3001\u305d\u306e\u76ee\u7684\u3068\u3057\u305f\u30a2\u30d7\u30ed\u30fc\u30c1\u306e\u3053\u3068\u3067\u3059\u3002<\/p>\n\n\n\n

    A deep understanding of contextual factors<\/strong> is important throughout the AI life- cycle. Context of use does not simply involve the users\u2019 context of use, it involves a much broader view of context: the organizational environment in which the AI system is being developed (including existing systems and products); the operational environment in which the system will be used; and the larger societal environment in which the system will be implemented.<\/p>\n\n\n\n

    AI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u3092\u901a\u3058\u3066\u91cd\u8981\u306a\u306e\u306f\u3001\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u8981\u56e0\u306e\u6df1\u3044\u7406\u89e3<\/strong>\u3067\u3059\u3002\u4f7f\u7528\u74b0\u5883\u3068\u306f\u3001\u5358\u306b\u30e6\u30fc\u30b6\u30fc\u306e\u4f7f\u7528\u74b0\u5883\u3092\u6307\u3059\u306e\u3067\u306f\u306a\u304f\u3001\u3082\u3063\u3068\u5e83\u3044\u7bc4\u56f2\u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u304c\u542b\u307e\u308c\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001AI\u30b7\u30b9\u30c6\u30e0\u304c\u958b\u767a\u3055\u308c\u308b\u7d44\u7e54\u74b0\u5883\uff08\u65e2\u5b58\u306e\u30b7\u30b9\u30c6\u30e0\u3084\u88fd\u54c1\u3092\u542b\u3080\uff09\u3001\u30b7\u30b9\u30c6\u30e0\u304c\u4f7f\u7528\u3055\u308c\u308b\u696d\u52d9\u74b0\u5883\u3001\u305d\u3057\u3066\u30b7\u30b9\u30c6\u30e0\u304c\u5c0e\u5165\u3055\u308c\u308b\u3001\u3088\u308a\u5927\u304d\u306a\u793e\u4f1a\u74b0\u5883\u306a\u3069\u3067\u3059\u3002<\/p>\n\n\n\n

    In the context of AI, this means placing humans in the loop, not only through meaningful human control [256], but also through their active participation in the preparation, learning, and decision-making phases of AI [257].\u201d Human-centered AI (HCAI)<\/strong> is an emerging area of scholarship that reconceptualizes HCD in the context of AI, providing human-centered AI design metaphors and suggested governance structures to develop reliable, safe, and trustworthy AI systems [258].<\/p>\n\n\n\n

    AI\u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3067\u306f\u3001\u3053\u308c\u306f\u4eba\u9593\u304c\u6709\u610f\u7fa9\u306a\u5236\u5fa1\u3092\u884c\u3046\u3068\u3044\u3046\u3060\u3051\u3067\u306a\u304f[256]\u3001AI\u306e\u6e96\u5099\u3001\u5b66\u7fd2\u3001\u610f\u601d\u6c7a\u5b9a\u306e\u6bb5\u968e\u3078\u306e\u7a4d\u6975\u7684\u306a\u53c2\u52a0\u3092\u901a\u3058\u3066\u3001\u4eba\u9593\u3092\u305d\u3046\u3057\u305f\u30eb\u30fc\u30d7\u5185\u306b\u7f6e\u304f\u3053\u3068\u3092\u610f\u5473\u3057\u307e\u3059[257]\u3002\u4eba\u9593\u3092\u4e2d\u5fc3\u306b\u7f6e\u3044\u305fAI\uff08HCAI\uff09<\/strong>\u306f\u65b0\u3057\u3044\u5b66\u554f\u9818\u57df\u3067\u3042\u308a[258]\u3001AI\u306e\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3067HCD\uff08\u4eba\u9593\u4e2d\u5fc3\u306e\u8a2d\u8a08\uff09\u3092\u518d\u8a8d\u8b58\u3055\u305b\u3001\u4eba\u9593\u3092\u4e2d\u5fc3\u306b\u7f6e\u3044\u305fAI\u306e\u8a2d\u8a08\u3068\u306f\u4f55\u3092\u610f\u5473\u3059\u308b\u306e\u304b\u3001\u4fe1\u983c\u3067\u304d\u308b\u3001\u5b89\u5168\u3067\u3001\u4fe1\u7528\u53ef\u80fd\u306aAI\u30b7\u30b9\u30c6\u30e0\u3092\u958b\u767a\u3059\u308b\u305f\u3081\u306e\u30ac\u30d0\u30ca\u30f3\u30b9\u306f\u3069\u3046\u3042\u308b\u3079\u304d\u304b\u3092\u3001\u79c1\u9054\u306b\u8003\u3048\u3055\u305b\u3066\u304f\u308c\u307e\u3059\u3002<\/p>\n\n\n\n

    3.4 How do we manage and provide oversight? Governance and AI Bias <\/strong>\u3069\u3046\u3084\u3063\u3066\u7ba1\u7406\u30fb\u76e3\u8996\u3092\u884c\u3046\u304b\uff1f\u30ac\u30d0\u30ca\u30f3\u30b9\u3068AI\u30d0\u30a4\u30a2\u30b9<\/p>\n\n\n\n

    3.4.1 Governance Guidance <\/strong>\u30ac\u30d0\u30ca\u30f3\u30b9\u30ac\u30a4\u30c0\u30f3\u30b9<\/em><\/p>\n\n\n\n

    \n

    Monitoring <\/strong>\u30e2\u30cb\u30bf\u30ea\u30f3\u30b0<\/p>\n<\/blockquote>\n\n\n\n

    AI systems may perform differently than expected once deployed, which can lead to differential treatment of individuals from different groups. A key measure to control this risk is to deploy additional systems that monitor for potential bias issues, which can alert the proper personnel when potential problems are detected<\/strong>.<\/p>\n\n\n\n

    AI\u30b7\u30b9\u30c6\u30e0\u306f\u3001\u4e00\u65e6\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u308b\u3068\u4e88\u60f3\u3068\u7570\u306a\u308b\u52d5\u4f5c\u3092\u3059\u308b\u3053\u3068\u304c\u3042\u308a\u3001\u3042\u308b\u500b\u4eba\u3092\u4ed6\u306e\u30b0\u30eb\u30fc\u30d7\u306e\u4eba\u9054\u3068\u5dee\u5225\u7684\u306b\u53d6\u308a\u6271\u3046\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u3046\u3057\u305f\u30ea\u30b9\u30af\u3092\u5236\u5fa1\u3059\u308b\u305f\u3081\u306e\u91cd\u8981\u306a\u5bfe\u7b56\u306f\u3001\u6f5c\u5728\u7684\u306a\u30d0\u30a4\u30a2\u30b9\u306e\u554f\u984c\u3092\u30e2\u30cb\u30bf\u30ea\u30f3\u30b0\u3059\u308b\u30b7\u30b9\u30c6\u30e0\u3092\u8ffd\u52a0\u914d\u5099\u3057\u3001\u6f5c\u5728\u7684\u306a\u554f\u984c\u304c\u691c\u51fa\u3055\u308c\u305f\u3068\u304d\u306b\u9069\u5207\u306a\u62c5\u5f53\u8005\u306b\u8b66\u544a\u3067\u304d\u308b\u3088\u3046\u306b\u3059\u308b\u3053\u3068\u3067\u3059\u3002<\/strong><\/p>\n\n\n\n

    Without such monitoring in place, it can be difficult to know if deployed system performance in the real world matches up to the measurements conducted in a laboratory environment, or whether newly collected data match the distribution of the training data.<\/strong><\/p>\n\n\n\n

    \u3053\u306e\u3088\u3046\u306a\u30e2\u30cb\u30bf\u30ea\u30f3\u30b0\u304c\u5b9f\u65bd\u3055\u308c\u306a\u3044\u3068\u3001\u5b9f\u4e16\u754c\u306b\u5c0e\u5165\u3057\u305f\u30b7\u30b9\u30c6\u30e0\u306e\u6027\u80fd\u304c\u5b9f\u9a13\u5ba4\u74b0\u5883\u3067\u884c\u3063\u305f\u6e2c\u5b9a\u5024\u3068\u4e00\u81f4\u3057\u3066\u3044\u308b\u306e\u304b\u3069\u3046\u304b\u3001\u3042\u308b\u3044\u306f\u3001\u65b0\u305f\u306b\u53ce\u96c6\u3057\u305f\u30c7\u30fc\u30bf\u304c\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u5206\u5e03\u3068\u4e00\u81f4\u3057\u3066\u3044\u308b\u304b\u3069\u3046\u304b\u3092\u77e5\u308b\u3053\u3068\u306f\u56f0\u96e3\u3068\u306a\u3063\u3066\u3057\u307e\u3044\u307e\u3059\u3002<\/strong><\/p>\n\n\n\n

    A key consideration for the success of live monitoring for bias is the collection of data from the active user population, especially data related to user demographics such as age and gender, to enable calculation of assessment measures. These type of data can have a variety of privacy implications and may be subject to legal restrictions on what types of data can be collected and under what conditions.<\/p>\n\n\n\n

    \u30d0\u30a4\u30a2\u30b9\u306b\u5bfe\u3059\u308b\u30e2\u30cb\u30bf\u30ea\u30f3\u30b0\u3092\u6210\u529f\u306b\u5c0e\u304f\u305f\u3081\u306b\u8003\u616e\u3059\u3079\u304d\u91cd\u8981\u306a\u70b9\u306f\u3001\u30a2\u30af\u30c6\u30a3\u30d6\u30e6\u30fc\u30b6\u30fc\u306e\u6bcd\u96c6\u56e3\u304b\u3089\u30c7\u30fc\u30bf\u3092\u53ce\u96c6\u3059\u308b\u3053\u3068\u3001\u7279\u306b\u5e74\u9f62\u3084\u6027\u5225\u306a\u3069\u306e\u30e6\u30fc\u30b6\u30fc\u306e\u30c7\u30e2\u30b0\u30e9\u30d5\u30a3\u30fc\u306b\u5fdc\u3058\u305f\u30c7\u30fc\u30bf\u3092\u53ce\u96c6\u3057\u3066\u8a55\u4fa1\u6307\u6a19\u3092\u7b97\u51fa\u3067\u304d\u308b\u3088\u3046\u306b\u3059\u308b\u3053\u3068\u3067\u3059\u3002\u3053\u3046\u3057\u305f\u30c7\u30fc\u30bf\u306f\u3001\u3055\u307e\u3056\u307e\u306a\u30d7\u30e9\u30a4\u30d0\u30b7\u30fc\u306b\u95a2\u308f\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u3001\u3069\u306e\u3088\u3046\u306a\u7a2e\u985e\u306e\u30c7\u30fc\u30bf\u3092\u3069\u306e\u3088\u3046\u306a\u6761\u4ef6\u3067\u53ce\u96c6\u3067\u304d\u308b\u304b\u306b\u3064\u3044\u3066\u306f\u3001\u6cd5\u7684\u898f\u5236\u306e\u5bfe\u8c61\u3068\u306a\u308b\u5834\u5408\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    \n

    Policies and Procedures <\/strong>\u30dd\u30ea\u30b7\u30fc\u3068\u624b\u9806\u306e\u6587\u66f8\u5316<\/p>\n<\/blockquote>\n\n\n\n

    In the context of AI systems, ensuring that written policies and procedures<\/strong> address key roles, responsibilities, and processes at all stages of the AI model lifecycle is critical to managing and detecting potential overall issues of AI system performance.20 Policies and procedures can enable consistent development and testing practices, which in turn can help to ensure that results from AI systems are repeatable and that related risks are consistently mapped, measured and managed. Without such policies, the management of AI bias can easily become subjective and inconsistent across organizations, which can exacerbate risks over time rather than minimize them \u2014 if, for example, irreconcilably different metrics are used across systems.<\/p>\n\n\n\n

    AI \u30b7\u30b9\u30c6\u30e0\u306e\u6027\u80fd\u306b\u95a2\u308f\u308b\u6f5c\u5728\u7684\u306a\u554f\u984c\u3092\u4e00\u62ec\u3057\u3066\u7ba1\u7406\u3057\u691c\u51fa\u3059\u308b\u305f\u3081\u306b\u306f\u3001AI \u30e2\u30c7\u30eb\u306e\u5168\u3066\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u30b9\u30c6\u30fc\u30b8\u306b\u304a\u3044\u3066\u3001\u5f79\u5272\u3001\u8cac\u4efb\u3001\u304a\u3088\u3073\u30d7\u30ed\u30bb\u30b9\u306b\u5bfe\u5fdc\u3059\u308b\u30dd\u30ea\u30b7\u30fc\u3068\u624b\u9806\u3092\u6587\u66f8\u5316<\/strong>\u3059\u308b\u3053\u3068\u304c\u3001\u6975\u3081\u3066\u91cd\u8981\u3067\u305920\u3002\u30dd\u30ea\u30b7\u30fc\u3068\u624b\u9806\u3092\u5b9a\u3081\u308b\u3053\u3068\u3067\u3001\u9996\u5c3e\u4e00\u8cab\u3057\u305f\u958b\u767a\u304a\u3088\u3073\u30c6\u30b9\u30c8\u3092\u5b9f\u65bd\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u3001\u305d\u308c\u306b\u3088\u3063\u3066\u3001AI \u30b7\u30b9\u30c6\u30e0\u304b\u3089\u306e\u7d50\u679c\u306e\u518d\u73fe\u6027\u304c\u4fdd\u8a3c\u3055\u308c\u3001\u95a2\u9023\u3059\u308b\u30ea\u30b9\u30af\u3092\u9f5f\u9f6c\u7121\u304f\u30de\u30c3\u30d4\u30f3\u30b0\u3001\u6e2c\u5b9a\u3001\u7ba1\u7406\u3059\u308b\u3053\u3068\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306a\u30dd\u30ea\u30b7\u30fc\u304c\u306a\u3051\u308c\u3070\u3001AI \u30d0\u30a4\u30a2\u30b9\u306e\u7ba1\u7406\u306f\u3001\u7d44\u7e54\u9593\u3067\u4e3b\u89b3\u7684\u3067\u4e00\u8cab\u6027\u306e\u306a\u3044\u3082\u306e\u306b\u306a\u308a\u3084\u3059\u304f\u3001\u4f8b\u3048\u3070\u3001\u7570\u306a\u308b\u57fa\u6e96\u3092\u5143\u306bAI\u30b7\u30b9\u30c6\u30e0\u306e\u6027\u80fd\u6e2c\u5b9a\u3092\u884c\u3063\u305f\u5834\u5408\u3001\u30ea\u30b9\u30af\u3092\u6700\u5c0f\u9650\u306b\u6291\u3048\u308b\u3069\u3053\u308d\u304b\u3001\u9577\u671f\u7684\u306b\u30ea\u30b9\u30af\u3092\u60aa\u5316\u3055\u305b\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

    \n

    Accountability <\/strong>\u30a2\u30ab\u30a6\u30f3\u30bf\u30d3\u30ea\u30c6\u30a3\uff08\u8aac\u660e\u8cac\u4efb\uff09<\/p>\n<\/blockquote>\n\n\n\n

    Accountability plays a critical role in governance efforts [261]. Governance without accountability is, in practice, unlikely to be effective. Ensuring that a specific team, and often, a specific individual \u2014 such as a Chief Model Risk Officer, as is now common in large consumer finance organizations \u2014 is responsible for bias management in AI systems is a fundamental accountability mechanism.22<\/p>\n\n\n\n

    \u30a2\u30ab\u30a6\u30f3\u30bf\u30d3\u30ea\u30c6\u30a3\u306f\u3001\u30ac\u30d0\u30ca\u30f3\u30b9\u306e\u53d6\u308a\u7d44\u307f\u306b\u304a\u3044\u3066\u91cd\u8981\u306a\u5f79\u5272\u3092\u679c\u305f\u3057\u307e\u3059[261]\u3002\u30a2\u30ab\u30a6\u30f3\u30bf\u30d3\u30ea\u30c6\u30a3\u306e\u306a\u3044\u30ac\u30d0\u30ca\u30f3\u30b9\u306f\u3001\u5b9f\u969b\u306b\u306f\u3001\u52b9\u679c\u7684\u3067\u3042\u308b\u3068\u306f\u8003\u3048\u306b\u304f\u3044\u3067\u3059\u3002\u7279\u5b9a\u306e\u30c1\u30fc\u30e0\u3001\u305d\u3057\u3066\u3001\u591a\u304f\u306e\u5834\u5408\u3001\u7279\u5b9a\u306e\u500b\u4eba\u3001\u4f8b\u3048\u3070\u3001\u5927\u898f\u6a21\u306a\u6d88\u8cbb\u8005\u91d1\u878d\u7d44\u7e54\u3067\u73fe\u5728\u4e00\u822c\u7684\u3067\u3042\u308b\u3088\u3046\u306a\u6700\u9ad8\u30e2\u30c7\u30eb\u30ea\u30b9\u30af\u8cac\u4efb\u8005\u304c\u3001AI\u30b7\u30b9\u30c6\u30e0\u306b\u304a\u3051\u308b\u30d0\u30a4\u30a2\u30b9\u7ba1\u7406\u306b\u8cac\u4efb\u3092\u6301\u3064\u3053\u3068\u3092\u4fdd\u8a3c\u3059\u308b\u3053\u3068\u306f\u3001\u57fa\u672c\u7684\u306a\u8aac\u660e\u8cac\u4efb\u306e\u30e1\u30ab\u30cb\u30ba\u30e0\u3067\u3059\u3002<\/p>\n\n\n\n

    Accountability for AI bias<\/strong> cannot lie on the shoulders of a single individual, which is why accountability mandates should also be embedded within and across the various teams involved in the training and deployment of AI systems.<\/strong><\/p>\n\n\n\n

    \u4f46\u3057\u3001AI\u30d0\u30a4\u30a2\u30b9\u306b\u5bfe\u3059\u308b\u8aac\u660e\u8cac\u4efb<\/strong>\u306f\u3001\u4e00\u500b\u4eba\u306e\u80a9\u306b\u304b\u304b\u308b\u3082\u306e\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u3060\u304b\u3089\u3053\u305d\u3001\u8aac\u660e\u8cac\u4efb\u306e\u8cac\u52d9\u306f\u3001AI\u30b7\u30b9\u30c6\u30e0\u306e\u5b66\u7fd2\u3068\u30c7\u30d7\u30ed\u30a4\u30e1\u30f3\u30c8\u306b\u95a2\u308f\u308b\u69d8\u3005\u306a\u30c1\u30fc\u30e0\u5185\u304a\u3088\u3073\u30c1\u30fc\u30e0\u5168\u4f53<\/strong>\u306b\u3082\u7d44\u307f\u8fbc\u307e\u308c\u308b\u3079\u304d\u306a\u306e\u3067\u3059\u3002<\/p>\n\n\n\n

    Model or algorithmic audits<\/strong> [264] can be used to assess and document such crucial accountability considerations. There are several notions of audits commonly discussed in the responsible and trustworthy AI communities. Audit may refer to a traditional internal audit function employed to track issues of model risk, as in traditional model governance. Audit may refer to a structured and principled application of lessons learned in financial audit practices to AI systems [265]. Alternatively, audit may refer to some general documentation and transparency approach. Audits can be an effective accountability, bias, and general risk mitigation mechanism.<\/strong><\/p>\n\n\n\n

    \u30e2\u30c7\u30eb\u76e3\u67fb\u3084\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u76e3\u67fb<\/strong> [264] \u306f\u3001\u3053\u306e\u3088\u3046\u306a\u91cd\u8981\u306a\u8aac\u660e\u8cac\u4efb\u306e\u691c\u8a0e\u4e8b\u9805\u3092\u8a55\u4fa1\u3057\u3001\u6587\u66f8\u5316\u3059\u308b\u305f\u3081\u306b\u6709\u52b9\u3067\u3059\u3002\u8cac\u4efb\u3042\u308b\u4fe1\u983c\u3067\u304d\u308bAI\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3067\u3088\u304f\u8b70\u8ad6\u3055\u308c\u308b\u76e3\u67fb\u306b\u306f\u3001\u3044\u304f\u3064\u304b\u306e\u6982\u5ff5\u304c\u3042\u308a\u307e\u3059\u3002\u76e3\u67fb\u306f\u3001\u5f93\u6765\u306e\u30e2\u30c7\u30eb\u30ac\u30d0\u30ca\u30f3\u30b9\u306e\u3088\u3046\u306b\u3001\u30e2\u30c7\u30eb\u30ea\u30b9\u30af\u306e\u554f\u984c\u3092\u8ffd\u8de1\u3059\u308b\u305f\u3081\u306b\u63a1\u7528\u3055\u308c\u305f\u4f1d\u7d71\u7684\u306a\u5185\u90e8\u76e3\u67fb\u6a5f\u80fd\u3092\u6307\u3059\u5834\u5408\u304c\u3042\u308a\u307e\u3059\u3002\u76e3\u67fb\u306f\u3001\u8ca1\u52d9\u76e3\u67fb\u5b9f\u52d9\u3067\u5b66\u3093\u3060\u6559\u8a13\u3092AI\u30b7\u30b9\u30c6\u30e0\u306b\u69cb\u9020\u7684\u304b\u3064\u539f\u5247\u7684\u306b\u9069\u7528\u3059\u308b\u3053\u3068\u3092\u6307\u3059\u304b\u3082\u3057\u308c\u307e\u305b\u3093[265]\u3002\u3042\u308b\u3044\u306f\u3001\u76e3\u67fb\u306f\u3001\u4f55\u3089\u304b\u306e\u4e00\u822c\u7684\u306a\u6587\u66f8\u5316\u53ca\u3073\u900f\u660e\u6027\u306e\u30a2\u30d7\u30ed\u30fc\u30c1\u3092\u6307\u3059\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u3002\u76e3\u67fb\u306f\u3001\u52b9\u679c\u7684\u306a\u8aac\u660e\u8cac\u4efb\u3001\u30d0\u30a4\u30a2\u30b9\u3001\u53ca\u3073\u4e00\u822c\u7684\u306a\u30ea\u30b9\u30af\u8efd\u6e1b\u306e\u30e1\u30ab\u30cb\u30ba\u30e0\u306b\u306a\u308a\u5f97\u307e\u3059\u3002<\/strong><\/p>\n\n\n\n

    \n

    Three lines of defense <\/strong>\u30b9\u30ea\u30fc\u30e9\u30a4\u30f3\u30fb\u30aa\u30d6\u30fb\u30c7\u30a3\u30d5\u30a7\u30f3\u30b9<\/p>\n<\/blockquote>\n\n\n\n

    Model risk management frameworks, for example, are often systematically implemented through the so-called \u201cthree lines of defense,\u201d which creates separate teams that are held accountable for different aspects of the model lifecycle. Typically, the first line of defense focuses on model development, the second on risk management, and the third on auditing.[24]<\/p>\n\n\n\n

    \u4f8b\u3048\u3070\u3001\u30e2\u30c7\u30eb\u30ea\u30b9\u30af\u7ba1\u7406\u306e\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u306f\u3001\u3044\u308f\u3086\u308b\u300c3\u3064\u306e\u9632\u885b\u7dda\u300d\u3092\u901a\u3058\u3066\u4f53\u7cfb\u7684\u306b\u5b9f\u65bd\u3055\u308c\u308b\u3053\u3068\u304c\u591a\u304f\u3001\u30e2\u30c7\u30eb\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u306e\u7570\u306a\u308b\u5074\u9762\u306b\u5bfe\u3057\u3066\u8cac\u4efb\u3092\u8ca0\u3046\u5225\u3005\u306e\u30c1\u30fc\u30e0\u304c\u4f5c\u3089\u308c\u307e\u3059\u3002\u5178\u578b\u7684\u306b\u306f\u3001\u7b2c\u4e00\u306e\u9632\u885b\u30e9\u30a4\u30f3\u306f\u30e2\u30c7\u30eb\u958b\u767a\u3001\u7b2c\u4e8c\u306e\u9632\u885b\u30e9\u30a4\u30f3\u306f\u30ea\u30b9\u30af\u7ba1\u7406\u3001\u7b2c\u4e09\u306e\u9632\u885b\u30e9\u30a4\u30f3\u306f\u76e3\u67fb\u306b\u7126\u70b9\u3092\u5f53\u3066\u307e\u3059\u3002[24]<\/p>\n\n\n\n

    \n

    Information Sharing <\/strong>\u60c5\u5831\u306e\u5171\u6709<\/p>\n<\/blockquote>\n\n\n\n

    Identifying internal mechanisms for teams to share information about bias incidents or other harmful impacts from AI helps to elevate the importance of AI risks and provides information for teams to avoid past failed designs.<\/p>\n\n\n\n

    AI\u306b\u3088\u308b\u30d0\u30a4\u30a2\u30b9\u3084\u305d\u306e\u4ed6\u306e\u6709\u5bb3\u306a\u5f71\u97ff\u306b\u95a2\u3059\u308b\u60c5\u5831\u3092\u30c1\u30fc\u30e0\u3067\u5171\u6709\u3059\u308b\u305f\u3081\u306e\u793e\u5185\u30e1\u30ab\u30cb\u30ba\u30e0\u3092\u660e\u78ba\u306b\u3059\u308b\u3053\u3068\u306f\u3001AI\u30ea\u30b9\u30af\u306e\u91cd\u8981\u6027\u3092\u9ad8\u3081\u3001\u30c1\u30fc\u30e0\u304c\u904e\u53bb\u306e\u5931\u6557\u3057\u305f\u8a2d\u8a08\u3092\u56de\u907f\u3059\u308b\u305f\u3081\u306e\u60c5\u5831\u3092\u30b7\u30a7\u30a2\u3059\u308b\u306e\u306b\u5f79\u7acb\u3061\u307e\u3059\u3002<\/p>\n\n\n\n

    As teams begin to create norms for tracking such incidents, it can potentially transform AI practices and the organizational culture. Improving awareness of how bias presents in deployed AI and its related impacts can enhance knowledge and capabilities, and prevent incidents. Fostering a culture of information sharing can also serve as a new area for community engagement.<\/p>\n\n\n\n

    \u30c1\u30fc\u30e0\u304c\u3053\u306e\u3088\u3046\u306a\u30a4\u30f3\u30b7\u30c7\u30f3\u30c8\u3092\u8ffd\u8de1\u3059\u308b\u305f\u3081\u306e\u898f\u7bc4\u3092\u4f5c\u308a\u59cb\u3081\u308b\u3053\u3068\u3067\u3001AI\u306e\u5b9f\u8df5\u3068\u7d44\u7e54\u6587\u5316\u3092\u5909\u9769\u3067\u304d\u308b\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002\u30c7\u30d7\u30ed\u30a4\u3055\u308c\u305f AI \u306b\u3069\u306e\u3088\u3046\u306a\u30d0\u30a4\u30a2\u30b9\u304c\u898b\u3089\u308c\u308b\u304b\u3001\u307e\u305f\u305d\u308c\u306b\u95a2\u9023\u3059\u308b\u5f71\u97ff\u306b\u3064\u3044\u3066\u306e\u8a8d\u8b58\u3092\u9ad8\u3081\u308b\u3053\u3068\u3067\u3001\u77e5\u8b58\u3068\u80fd\u529b\u3092\u9ad8\u3081\u3001\u30a4\u30f3\u30b7\u30c7\u30f3\u30c8\u3092\u9632\u6b62\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002\u307e\u305f\u3001\u60c5\u5831\u5171\u6709\u306e\u6587\u5316\u3092\u91b8\u6210\u3059\u308b\u3053\u3068\u3067\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3078\u306e\u65b0\u305f\u306a\u30a8\u30f3\u30b2\u30fc\u30b8\u30e1\u30f3\u30c8\u306e\u5834\u3068\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n

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

    NIST has provided an initial socio-technical framing for AI bias<\/strong> in this document, including key context and terminology, highlights of the main challenges, and foundational directions for future guidance. This information is classified and discussed through the document according to three key areas:<\/p>\n\n\n\n

    NIST\u306f\u3001\u672c\u5831\u544a\u66f8\u306b\u304a\u3044\u3066\u3001AI\u30d0\u30a4\u30a2\u30b9\u306b\u95a2\u3059\u308b\u6700\u521d\u306e\u793e\u4f1a\u6280\u8853\u7684\u306a\u67a0\u7d44<\/strong>\u307f\u3092\u63d0\u6848\u3057\u3001\u305d\u306e\u4e2d\u3067\u3001\u4e3b\u8981\u306a\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3068\u7528\u8a9e\u306e\u5b9a\u7fa9\u3001\u4e3b\u8981\u306a\u8ab2\u984c\u306e\u30cf\u30a4\u30e9\u30a4\u30c8\u3001\u5c06\u6765\u306b\u5411\u3051\u305f\u30ac\u30a4\u30c0\u30f3\u30b9\u306e\u57fa\u790e\u3068\u306a\u308b\u65b9\u5411\u6027\u7b49\u3092\u660e\u3089\u304b\u306b\u3057\u307e\u3057\u305f\u3002\u3053\u308c\u3089\u306e\u60c5\u5831\u306f\u3001\u672c\u5831\u544a\u66f8\u3092\u901a\u3058\u3066\u4ee5\u4e0b3\u3064\u306e\u4e3b\u8981\u306a\u5206\u91ce\u306b\u5f93\u3063\u3066\u5206\u985e\u3055\u308c\u8b70\u8ad6\u3057\u307e\u3057\u305f\u3002<\/p>\n\n\n\n

      \n
    1. dataset availability, representativeness, and suitability in socio-technical contexts;<\/li>\n<\/ol>\n\n\n\n

      \u793e\u4f1a\u6280\u8853\u7684\u6587\u8108\u306b\u304a\u3051\u308b\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u5229\u7528\u53ef\u80fd\u6027\u3001\u4ee3\u8868\u6027\u3001\u9069\u5408\u6027\u3002<\/p>\n\n\n\n

      2. TEVV considerations for measurement and metrics to support testing and evaluation;<\/p>\n\n\n\n

      \u30c6\u30b9\u30c8\u3068\u8a55\u4fa1\u3092\u652f\u63f4\u3059\u308b\u305f\u3081\u306e\u6e2c\u5b9a\u3068\u6e2c\u5b9a\u57fa\u6e96\u306b\u95a2\u3059\u308b TEVV \u306e\u691c\u8a0e\u3002<\/p>\n\n\n\n

      3. human factors, including societal and historic biases within individuals and organizations, participatory approaches such as human-centered design, and human\u2013in\u2013the\u2013 loop practices.<\/p>\n\n\n\n

      \u500b\u4eba\u3084\u7d44\u7e54\u306b\u304a\u3051\u308b\u793e\u4f1a\u7684\u30fb\u6b74\u53f2\u7684\u504f\u898b\u3001\u4eba\u9593\u4e2d\u5fc3\u8a2d\u8a08\u306e\u3088\u3046\u306a\u53c2\u52a0\u578b\u30a2\u30d7\u30ed\u30fc\u30c1\u3001\u30d2\u30e5\u30fc\u30de\u30f3\u30a4\u30f3\u30b6\u30eb\u30fc\u30d7\u306e\u5b9f\u8df5\u3092\u542b\u3080\u4eba\u7684\u8981\u56e0\u3002<\/p>\n\n\n\n

      The intent is for this guidance to be of specific assistance for organizations who commission, design, develop, deploy, use, or evaluate AI for a variety of use cases. By providing these entities with clear, explicit, and technically valid guidance NIST intends to improve the state of practice for AI bias and assure system trustworthiness.<\/p>\n\n\n\n

      \u3053\u306e\u30ac\u30a4\u30c0\u30f3\u30b9\u306f\u3001\u69d8\u3005\u306a\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9\u306b\u304a\u3044\u3066\u3001AI\u3092\u59d4\u8a17\u3001\u8a2d\u8a08\u3001\u958b\u767a\u3001\u30c7\u30d7\u30ed\u30a4\u3001\u4f7f\u7528\u3001\u8a55\u4fa1\u3059\u308b\u7d44\u7e54\u306b\u3068\u3063\u3066\u5177\u4f53\u7684\u306a\u52a9\u3051\u3068\u306a\u308b\u3053\u3068\u3092\u610f\u56f3\u3057\u3066\u3044\u307e\u3059\u3002NIST\u306f\u3001\u3053\u308c\u3089\u306e\u7d44\u7e54\u306b\u660e\u78ba\u304b\u3064\u660e\u793a\u7684\u3067\u3001\u6280\u8853\u7684\u306b\u6709\u52b9\u306a\u30ac\u30a4\u30c0\u30f3\u30b9\u3092\u63d0\u4f9b\u3059\u308b\u3053\u3068\u3067\u3001AI\u30d0\u30a4\u30a2\u30b9\u306b\u95a2\u3059\u308b\u5b9f\u8df5\u72b6\u6cc1\u3092\u6539\u5584\u3057\u3001\u30b7\u30b9\u30c6\u30e0\u306e\u4fe1\u983c\u6027\u3092\u4fdd\u8a3c\u3059\u308b\u3053\u3068\u3092\u76ee\u6307\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n

      \n

      \u672c\u30d6\u30ed\u30b0\u3092\u6700\u5f8c\u307e\u3067\u304a\u8aad\u307f\u306b\u306a\u3063\u3066\u3044\u305f\u3060\u304d\u3001\u6709\u96e3\u3046\u3054\u3056\u3044\u307e\u3057\u305f\u3002<\/strong><\/p>\n\n\n\n

      \u5f93\u6765\u306e\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3068\u7570\u306a\u308a\u3001AI\u306f\u7e4a\u7d30\u3067\u5bb9\u6613\u306b\u54c1\u8cea\u304c\u52a3\u5316\u3057\u3066\u3057\u307e\u3044\u307e\u3059\u3002\u4eba\u624b\u306b\u3088\u308b\u30c7\u30d0\u30c3\u30b0\u306b\u306f\u3001\u591a\u304f\u306e\u6642\u9593\u3068\u52b4\u529b\u3092\u8981\u3057\u975e\u5e38\u306b\u56f0\u96e3\u3067\u3059\u3002\u200d
      \u682a\u5f0f\u4f1a\u793eCitadel AI<\/strong><\/a>\u00a0\u3067\u306f\u3001\u958b\u767a\u304b\u3089\u904b\u7528\u306b\u81f3\u308b\u5168\u3066\u306eAI\u306e\u30e9\u30a4\u30d5\u30b5\u30a4\u30af\u30eb\u30b9\u30c6\u30fc\u30b8\u306b\u304a\u3044\u3066\u3001AI\u306e\u7570\u5e38\u3092\u77ac\u6642\u306b\u81ea\u52d5\u691c\u77e5\u30fb\u9632\u5fa1\u3059\u308b\u30b7\u30b9\u30c6\u30e0\u3092\u63d0\u4f9b\u3057\u3066\u3044\u307e\u3059\u3002
      \u30d3\u30b8\u30cd\u30b9\u3084\u30b3\u30f3\u30d7\u30e9\u30a4\u30a2\u30f3\u30b9\u30fb\u30bb\u30ad\u30e5\u30ea\u30c6\u30a3\u4e0a\u306eAI\u56fa\u6709\u306e\u65b0\u305f\u306a\u30ea\u30b9\u30af\u306b\u5bfe\u3057\u3001\u4f01\u696d\u306e\u7686\u69d8\u304c\u30d7\u30ed\u30a2\u30af\u30c6\u30a3\u30d6\u306b\u5bfe\u51e6\u9802\u3051\u308b\u3088\u3046\u3001\u5c11\u3057\u3067\u3082\u7686\u69d8\u306e\u304a\u5f79\u306b\u7acb\u3066\u308b\u3053\u3068\u3092\u9858\u3063\u3066\u3044\u307e\u3059\u3002
      \u3054\u610f\u898b\u30fb\u3054\u8981\u671b\u306f\u3001      info@citadel-ai.com<\/a>\u00a0<\/strong>\u307e\u3067\u304a\u77e5\u3089\u305b\u304f\u3060\u3055\u3044\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

      \u306f\u3058\u3081\u306b \u5f0a\u793e\u3067\u306f\u7c73\u56fd\u56fd\u7acb\u6a19\u6e96\u6280\u8853\u7814\u7a76\u6240(NIST)\u306e\u8a31\u53ef\u3092\u5f97\u3066\u30012022\u5e743\u6708\u306b\u767a\u8868\u3055\u308c\u305f\u5831\u544a\u66f8 \u201c NIST Special Publication 1270: Towards a Standar […]<\/p>\n","protected":false},"author":1,"featured_media":1199,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[9],"tags":[],"class_list":["post-1087","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\/1087","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\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/comments?post=1087"}],"version-history":[{"count":0,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/posts\/1087\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media\/1199"}],"wp:attachment":[{"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/media?parent=1087"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/categories?post=1087"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/citadel-ai.com\/ja\/wp-json\/wp\/v2\/tags?post=1087"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}