Scholarlink

<p>Debates about political bias in artificial intelligence often conflate ideological leanings with epistemic error. This paper introduces the Viewpoint Preferences–Epistemic Integrity (VPEI) framework to separate these phenomena. The framework distinguishes between a system’s directional political affinities (viewpoint preferences) and the quality of its reasoning processes (epistemic integrity) when generating politically laden outputs. Within the VPEI framework, political bias arises only when viewpoint preferences coincide with epistemic distortion, such as inconsistent application of evidentiary standards, failures of perspective-taking, or systematic errors in factual recall and forecasting. Existing policy debates about “biased AI”, including Trump Executive Order on “Woke AI” (14319, 2025), often assume that ideological asymmetry itself indicates epistemic failure. VPEI challenges this assumption. By disentangling ideological orientation from epistemic performance, the framework shows how AI systems may be politically asymmetric yet epistemically disciplined or appear neutral while failing basic standards of truth-seeking. It thereby provides a more coherent basis for evaluating and regulating AI systems in politically contested domains.</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

Assessing political bias in AI systems: a framework for disentangling viewpoint preferences from epistemic integrity

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • David Rozado,
  • Philip E. Tetlock

摘要

Debates about political bias in artificial intelligence often conflate ideological leanings with epistemic error. This paper introduces the Viewpoint Preferences–Epistemic Integrity (VPEI) framework to separate these phenomena. The framework distinguishes between a system’s directional political affinities (viewpoint preferences) and the quality of its reasoning processes (epistemic integrity) when generating politically laden outputs. Within the VPEI framework, political bias arises only when viewpoint preferences coincide with epistemic distortion, such as inconsistent application of evidentiary standards, failures of perspective-taking, or systematic errors in factual recall and forecasting. Existing policy debates about “biased AI”, including Trump Executive Order on “Woke AI” (14319, 2025), often assume that ideological asymmetry itself indicates epistemic failure. VPEI challenges this assumption. By disentangling ideological orientation from epistemic performance, the framework shows how AI systems may be politically asymmetric yet epistemically disciplined or appear neutral while failing basic standards of truth-seeking. It thereby provides a more coherent basis for evaluating and regulating AI systems in politically contested domains.