<p>Artificial intelligence (AI) is reshaping society, raising questions about trust, risks, and the asymmetries between public and academic perspectives. We examine how the German public (N = 1,110), comprising individuals who interact with or are affected by AI, and academic AI experts (N = 119, mainly from Germany), who contribute to research, educate practitioners, and inform policymaking, construct mental models of AI’s capabilities and impacts across 71 scenarios. These scenarios span diverse domains (including sustainability, healthcare, employment, inequality, art, and warfare) and were evaluated across four dimensions using the psychometric model: likelihood, perceived risk, perceived benefit, and overall value. Across scenarios, academic experts generally anticipated higher probabilities of occurrence, perceived lower risks, and reported greater benefits than the public, while also expressing more positive overall evaluations of AI. Beyond differences in absolute assessments, the two groups exhibited systematically different evaluative patterns: experts’ value judgments were driven primarily by perceived benefits (<InlineEquation ID="IEq100"> <EquationSource Format="TEX">$\beta = 0.623$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>β</mi> <mo>=</mo> <mn>0.623</mn> </mrow> </math></EquationSource> </InlineEquation> vs. <InlineEquation ID="IEq101"> <EquationSource Format="TEX">$\beta = 0.623$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>β</mi> <mo>=</mo> <mn>0.623</mn> </mrow> </math></EquationSource> </InlineEquation>), whereas public evaluations placed more weight on perceived risks (<InlineEquation ID="IEq103"> <EquationSource Format="TEX">$\beta = 0.703$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>β</mi> <mo>=</mo> <mn>0.703</mn> </mrow> </math></EquationSource> </InlineEquation> vs. <InlineEquation ID="IEq104"> <EquationSource Format="TEX">$\beta = 0.361$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>β</mi> <mo>=</mo> <mn>0.361</mn> </mrow> </math></EquationSource> </InlineEquation>), reflecting distinct risk–benefit trade-offs. Visual mappings indicate convergent domains (e.g., medical diagnoses and criminal use) and tension points (e.g., justice and political decision-making) that may warrant targeted communication or policy attention. While this study does not assess AI systems or design practices directly, the observed divergence in mental models suggests that the research, implementation, and use of AI may inadvertently neglect the risk-related priorities of the public. Such biases in research and implementation may yield “procrustean AI”—systems insufficiently aligned with the needs of the affected public (akin to the Bed of Procrustes). We address the socio-technical challenge of expert-centric governance and advocate for participatory practices.</p>

Charting the AI perception gap: divergent views on risk, benefit, and value between experts and the public challenge the societal acceptance of AI

期刊论文
出版模式：

开放获取
发表日期：

2026年4月29日

查看全文

Philipp Brauner,
Felix Glawe,
Gian Luca Liehner,
Luisa Vervier,
Martina Ziefle

摘要

Artificial intelligence (AI) is reshaping society, raising questions about trust, risks, and the asymmetries between public and academic perspectives. We examine how the German public (N = 1,110), comprising individuals who interact with or are affected by AI, and academic AI experts (N = 119, mainly from Germany), who contribute to research, educate practitioners, and inform policymaking, construct mental models of AI’s capabilities and impacts across 71 scenarios. These scenarios span diverse domains (including sustainability, healthcare, employment, inequality, art, and warfare) and were evaluated across four dimensions using the psychometric model: likelihood, perceived risk, perceived benefit, and overall value. Across scenarios, academic experts generally anticipated higher probabilities of occurrence, perceived lower risks, and reported greater benefits than the public, while also expressing more positive overall evaluations of AI. Beyond differences in absolute assessments, the two groups exhibited systematically different evaluative patterns: experts’ value judgments were driven primarily by perceived benefits ( $\beta = 0.623$ $β = 0.623$ vs. $\beta = 0.623$ $β = 0.623$ ), whereas public evaluations placed more weight on perceived risks ( $\beta = 0.703$ $β = 0.703$ vs. $\beta = 0.361$ $β = 0.361$ ), reflecting distinct risk–benefit trade-offs. Visual mappings indicate convergent domains (e.g., medical diagnoses and criminal use) and tension points (e.g., justice and political decision-making) that may warrant targeted communication or policy attention. While this study does not assess AI systems or design practices directly, the observed divergence in mental models suggests that the research, implementation, and use of AI may inadvertently neglect the risk-related priorities of the public. Such biases in research and implementation may yield “procrustean AI”—systems insufficiently aligned with the needs of the affected public (akin to the Bed of Procrustes). We address the socio-technical challenge of expert-centric governance and advocate for participatory practices.