Scholarlink

<p>Should artificial intelligence (AI) companies be taxed, and if so, how? This paper presents two arguments which together support an ability-indexed tax on AI systems. First: AI companies—including frontier labs and traditional firms using AI in production—do not deserve all of their income. Because AI companies do not contribute important elements of their models, including (in part) their algorithms and the public data on which they were trained, on a desert-based normative framework they have limited moral entitlement to their income. I formally decompose AI models using the game-theoretic concept of <i>Shapley value</i>, showing that the general point about limited AI company deserts is robust. Indeed, the deserved income is plausibly small. Second, I turn to the public finance question of how the excess income can be efficiently redistributed. Drawing on <i>optimal tax theory</i>, I present a model in which an ability-indexed AI tax avoids the intensive margin distortion familiar from human taxation. This is because AI relaxes a key constraint of optimal tax theory: Unlike with humans, it is possible to observe AI system ability. Another ramification of the model is that relatively high-ability/high-profit AI systems will be taxed down to a common level of profit. If individuals, firms, and others ought to get what they deserve, there is a strong moral case for taxing AI companies, and good theoretical grounds to believe the revenue can be generated and redistributed efficiently.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

What do AI companies deserve? Shapley attribution and optimal tax theory

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Thomas Mulligan

摘要

Should artificial intelligence (AI) companies be taxed, and if so, how? This paper presents two arguments which together support an ability-indexed tax on AI systems. First: AI companies—including frontier labs and traditional firms using AI in production—do not deserve all of their income. Because AI companies do not contribute important elements of their models, including (in part) their algorithms and the public data on which they were trained, on a desert-based normative framework they have limited moral entitlement to their income. I formally decompose AI models using the game-theoretic concept of Shapley value, showing that the general point about limited AI company deserts is robust. Indeed, the deserved income is plausibly small. Second, I turn to the public finance question of how the excess income can be efficiently redistributed. Drawing on optimal tax theory, I present a model in which an ability-indexed AI tax avoids the intensive margin distortion familiar from human taxation. This is because AI relaxes a key constraint of optimal tax theory: Unlike with humans, it is possible to observe AI system ability. Another ramification of the model is that relatively high-ability/high-profit AI systems will be taxed down to a common level of profit. If individuals, firms, and others ought to get what they deserve, there is a strong moral case for taxing AI companies, and good theoretical grounds to believe the revenue can be generated and redistributed efficiently.