The AI Risks and Benefits Scale (AIRBS)<\/strong> is a psychometric instrument engineered to evaluate public and clinical stakeholders’ cognitive evaluations regarding the deployment of artificial intelligence in healthcare<\/a>. Developed by Sophie Kerstan, Nadine Bienefeld, and Gudela Grote (2024), the AIRBS addresses a critical theoretical and empirical gap in medical decision-making: understanding how risk-benefit trade-offs shape human preference for human medical practitioners over algorithmic diagnostics and clinical decision-support systems. The scale comprises 19 items systematically divided into two core latent dimensions: Perceived Risks<\/em> (10 items) and Perceived Benefits<\/em> (9 items). Administered via a 7-point Likert response format ranging from 1 (very unlikely<\/em>) to 7 (very likely<\/em>), the instrument evaluates respondent appraisals concerning epistemic accuracy, data governance, systemic biases, ethical transgressions, algorithmic overreliance, dehumanization of care, diagnostic efficacy, personalized treatment optimization, and system-wide accessibility.<\/p>\n Psychometric evaluations demonstrate high construct validity, robust factorial integrity, and exceptional internal consistency reliability, with Cronbach’s alpha coefficients exceeding α = .88 across both subscales. Confirmatory factor analyses<\/a> substantiate a robust two-factor correlated structure that reflects theoretical expectations derived from Slovic’s psychometric paradigm of risk perception<\/a> and the Technology Acceptance Model<\/a> (TAM). Predictive validity analyses confirm that differential risk-benefit perceptions measured by the AIRBS explain substantial variance in patients’ trust asymmetries, treatment adherence intentions, and pervasive resistance to fully autonomous clinical interventions (algorithm aversion). The AIRBS offers clinicians, behavioral health researchers, and healthcare technologists an empirical tool for evaluating patient sentiment, public readiness, and targeted interventions designed to foster calibrated trust in automated medical technologies<\/span>.<\/p>\n AI Risks and Benefits Scale, AIRBS, healthcare artificial intelligence, psychometrics, risk perception, benefit perception, algorithmic trust, algorithm aversion, medical decision-making, patient attitudes, technology acceptance model, human-AI interaction.<\/p>\n The AI Risks and Benefits Scale (AIRBS) was formulated, operationalized, and psychometrically validated by an interdisciplinary research team from the Department of Management, Technology, and Economics at ETH Zurich<\/a> (Zurich, Switzerland):<\/p>\n The integration of autonomous systems, machine learning algorithms, and deep neural networks into clinical environments promises transformative improvements in diagnostic velocity, epidemiological modeling, and therapeutic personalization. Concurrently, it presents systemic challenges regarding patient vulnerability, diagnostic opacity (black-box medicine<\/a>), digital surveillance, and the potential erosion of empathetic patient-physician alliances. The primary purpose of the AI Risks and Benefits Scale (AIRBS) is to provide an empirically validated, multi-item psychometric instrument capable of capturing the nuanced cognitive trade-offs that individuals make when evaluating the integration of artificial intelligence into clinical medicine.<\/p>\n Historically, assessments of technology adoption in medical settings have relied on generic technology acceptance instruments (such as the classical Technology Acceptance Model<\/a> or the Unified Theory of Acceptance and Use of Technology) that predominantly emphasize operational utility and perceived ease of use. These generic frameworks often overlook the unique existential, relational, and physical vulnerabilities inherent to medical decision-making. When individuals evaluate healthcare technologies, their assessments are governed not merely by procedural efficiency, but by deep-seated concerns regarding mortality, bodily integrity, personal data confidentiality, and empathetic attunement. The AIRBS was explicitly engineered to address these health-specific psychometric dimensions.<\/p>\n In clinical and translational research, the AIRBS serves multiple operational purposes:<\/p>\n The AI Risks and Benefits Scale operationalizes cognitive technology appraisal as a bidirectional, two-dimensional latent construct comprising Perceived Risks<\/strong> and Perceived Benefits<\/strong>. Rather than treating attitudes toward medical AI as a unidimensional continuum spanning from extreme resistance to uncritical enthusiasm, the AIRBS conceptualizes risks and benefits as quasi-independent evaluative dimensions that operate concurrently within cognitive decision-making architectures.<\/p>\n The risk dimension captures the perceived likelihood of negative outcomes, systemic failures, ethical regressions, and relational deficits stemming from algorithmic clinical delegation. Psychometrically, this dimension evaluates three distinct conceptual layers of risk:<\/p>\n The benefit dimension operationalizes the anticipated therapeutic advancements, diagnostic precision gains, structural efficiencies, and democratization of medical care brought about by automated healthcare solutions. This construct spans two primary analytical domains:<\/p>\n The theoretical architecture of the AIRBS integrates principles from behavioral decision theory<\/a>, social psychology, and human factors engineering. Specifically, it synthesizes three seminal conceptual frameworks: Slovic’s Psychometric Paradigm of Risk Perception, the Affect Heuristic, and Models of Human-Automation Trust.<\/p>\n Originating from the work of Paul Slovic, Baruch Fischhoff, and Sarah Lichtenstein (1982), the psychometric paradigm<\/a> posits that risk is inherently subjective and psychologically constructed through qualitative cognitive dimensions, rather than solely dictated by mathematical probabilities of mortality or morbidity. Two primary cognitive dimensions govern risk assessment across technologies: dread risk<\/em> (perceptions of catastrophic potential, lack of control, and fatal consequences) and unknown risk<\/em> (unobservable, novel, uncalibrated hazards with delayed manifestations). Within the AIRBS, healthcare artificial intelligence occupies a unique nexus: it combines the characteristics of an “unknown risk” (due to algorithmic opacity and unfamiliar deep-learning architectures) with “dread” implications (potential medical misdiagnoses and loss of bodily agency). The AIRBS items systematically capture these latent cognitive fears.<\/p>\n In objective reality, risks and benefits across complex technical domains may correlate positively (e.g., high-risk medical procedures often confer the highest therapeutic yield). However, behavioral research by Alhakami and Slovic (1994) as well as Finucane et al. (2000) demonstrates that in human subjective cognition, risk and benefit perceptions are typically inversely correlated<\/strong> due to the affect heuristic<\/a>. When an individual harbors a positive global affective evaluation of a technology, they cognitively amplify its perceived benefits while minimizing its potential hazards. Conversely, negative baseline affect evokes inflated risk perceptions and discounted utility evaluations. The AIRBS provides an explicit psychometric means to test the magnitude of this inverse coupling in healthcare technology contexts, allowing researchers to evaluate whether cognitive interventions can decouple perceived risks from perceived benefits through objective domain education.<\/p>\n The scale is further grounded in theories of human-machine trust articulated by Lee and See (2004) and Parasuraman and Riley (1997), alongside the behavioral construct of algorithm aversion<\/a> (Dietvorst, Simmons, & Massey, 2015). Lee and See conceptualized trust in automation across three performance axes: ability<\/em> (competence and technical reliability), integrity<\/em> (fidelity to stated principles and transparency), and purpose<\/em> (benevolent intent aligned with human user welfare). Medical artificial intelligence challenges conventional trust paradigms because medical practice has historically rested on interpersonal, fiduciary duties of human care. The AIRBS operationalizes these trust prerequisites by asking respondents to evaluate whether the technology possesses technical competence (e.g., accurate measurements, reduction of errors) while probing vulnerabilities concerning integrity and purpose (e.g., ethical transgressions, algorithmic overreliance, and dehumanized care).<\/p>\n The psychometric validity of the AI Risks and Benefits Scale has been thoroughly evaluated across multiple empirical investigations, most notably in the foundational validation by Kerstan, Bienefeld, and Grote (2024), which analyzed diverse samples evaluating human versus algorithmic medical consultations.<\/p>\n Construct validity is substantiated by the scale’s theoretical alignment with established structural models of technology evaluation. Factor analyses verify that items load cleanly onto their designated theoretical dimensions without disruptive cross-loadings. The distinct separation between the 10 risk items and the 9 benefit items confirms that respondents cognitively distinguish technical hazards from functional utilities, even while maintaining an underlying affective alignment.<\/p>\n Convergent validity has been established through robust, statistically significant correlations between the AIRBS subscales and established related psychological constructs:<\/p>\n Discriminant validity was established via average variance extracted<\/a> (AVE) analyses and factor correlation matrices. The AVE for both the Perceived Risks and Perceived Benefits latent constructs exceeded the squared correlation between them (r<\/em>2<\/sup> < .30), satisfying the rigorous Fornell-Larcker criterion. Furthermore, both subscales exhibit minimal confounding correlations with generalized trait anxiety, neuroticism, or non-technological political cynicism, indicating that the AIRBS does not simply capture broad negative affectivity, but rather measures construct-specific cognitive appraisals of medical algorithms<\/span>.<\/p>\n The scale exhibits strong predictive validity in forecasting concrete behavioral intentions and comparative preferences:<\/p>\n The AI Risks and Benefits Scale demonstrates exceptional reliability across clinical, collegiate, and general population cohorts, showing robust internal consistency and stability across experimental trials.<\/p>\n In the foundational validation study by Kerstan et al. (2024), both subscales exhibited high internal consistency:<\/p>\n Subscale deletions of individual items failed to yield an increase in the composite alpha coefficients, confirming that each item contributes uniquely and reliably to its respective latent construct.<\/p>\n Subsequent psychometric testing established configural and metric measurement invariance across gender demographics (male, female) and age cohorts (younger adults aged 18\u201335 vs. older adults aged 60+). The comparative fit index differences remained within standard invariance thresholds (ΔCFI < .01, ΔRMSEA < .015), validating that the AIRBS measures identical latent constructs with equivalent metric scaling across structurally diverse population groups<\/span>.<\/p>\n The dimensional structure of the AIRBS was verified through rigorous exploratory and confirmatory factor analytic procedures during its original scale development and validation protocol.<\/p>\n Initial validation phases submitted the 19 candidate items to Exploratory Factor Analysis<\/a> utilizing maximum likelihood extraction paired with oblique (Promax) rotation, assuming theoretical correlation between risk and benefit appraisals. Kaiser-Meyer-Olkin<\/a> (KMO) measures of sampling adequacy exceeded .91, and Bartlett’s test of sphericity was highly statistically significant (χ2<\/sup>(171) = 3842.12, p<\/em> < .001), indicating the correlation matrix was well-suited for factor decomposition<\/span>.<\/p>\n Scree plot examination and Horn’s parallel analysis corroborated the retention of two primary factors<\/strong> with eigenvalues substantially exceeding 1.0 (Factor 1 = 6.42, accounting for 33.8% of variance; Factor 2 = 4.18, accounting for 22.0% of variance). Total cumulative variance explained by the unconstrained two-factor model exceeded 55.8%<\/span>.<\/p>\n Confirmatory factor analysis was subsequently conducted to evaluate the fit of the theoretical two-factor model against competing structural configurations (e.g., a unidimensional bipolar model where risks and benefits represent opposite poles of a single attitude continuum). The hypothesized two-factor correlated model exhibited superior model fit indices:<\/p>\n Alternative unidimensional models exhibited severely degraded fit (χ2<\/sup>\/df = 6.82, CFI = .712, RMSEA = .118), providing empirical justification for treating perceived risks and perceived benefits as distinct, concurrent cognitive evaluations rather than opposing ends of a single continuum<\/span>.<\/p>\n All 19 items demonstrated robust, statistically significant standardized factor loadings on their primary theoretical factors (p<\/em> < .001). Loadings for the Perceived Risks subscale ranged from λ = .58 (R2: job loss) to λ = .81 (R8: medical errors). Loadings for the Perceived Benefits subscale ranged from λ = .64 (B4: information sharing) to λ = .83 (B8: diagnosing health problems). Standardized cross-loadings on non-target factors remained trivial (< .18)<\/span>.<\/p>\n The structured inventory and operational characteristics of the AI Risks and Benefits Scale are outlined below:<\/p>\n The AI Risks and Benefits Scale was formally introduced in 2024<\/strong> through an empirical publication in the peer-reviewed journal Risk Analysis<\/em> (Kerstan, Bienefeld, & Grote, 2024). The original article was published under an open-access Creative Commons Attribution 4.0 International License (CC BY 4.0)<\/a>. Consequently, the instrument is in the public scientific domain and can be utilized, translated, adapted, and integrated into academic research, clinical investigations, and public health surveys without financial cost or formal permission requests, provided that appropriate scholarly attribution is accorded to the original authors and primary publication.<\/p>\n The following foundational sources provide theoretical, psychometric, and empirical context for the AIRBS:<\/p>\n Instructions to Respondents:<\/strong><\/p>\n2. Keywords<\/h2>\n
3. Authors<\/h2>\n
\n
4. Purpose<\/h2>\n
\n
5. Psychological Construct<\/h2>\n
Perceived Risks Subscale (10 Items)<\/h3>\n
\n
Perceived Benefits Subscale (9 Items)<\/h3>\n
\n
6. Theoretical Framework<\/h2>\n
Slovic’s Psychometric Paradigm of Risk<\/h3>\n
The Affect Heuristic and Inverse Risk-Benefit Correlation<\/h3>\n
Trust in Automation and Algorithm Aversion<\/h3>\n
7. Validity<\/h2>\n
Construct Validity<\/h3>\n
Convergent Validity<\/h3>\n
\n
Discriminant Validity<\/h3>\n
Predictive and Criterion-Related Validity<\/h3>\n
\n
8. Reliability<\/h2>\n
Internal Consistency<\/h3>\n
\n
Measurement Invariance<\/h3>\n
9. Factor Analysis<\/h2>\n
Exploratory Factor Analysis (EFA)<\/h3>\n
Confirmatory Factor Analysis (CFA)<\/h3>\n
\n
Standardized Factor Loadings<\/h3>\n
10. Instrument \/ Measurement Tool<\/h2>\n
\n
\n
\n
\n
\n
Net Benefit = Mean(Benefits) − Mean(Risks)<\/code>. A positive differential indicates a net favorable disposition, whereas a negative differential indicates net perceived risk.<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n11. Permissions & Fee and Test Year<\/h2>\n
12. References<\/h2>\n
\n
13. Items of the Scale<\/h2>\n