AI decision support nearly triples fraud-detection odds — but vulnerability is age- and scam-specific
In a 10-day, 2×2 online lab-in-the-field study of 72 users, real-time AI assistance cut risky behavioral intention by 0.56 points and reduced deceptive responses by ~34% lower odds — while vulnerability matched scam type. Younger users fell for social-media and consumer fraud; older users for government-impersonation and financial fraud.
Most fraud research describes WHO is vulnerable via static risk profiles or single-scam studies; this paper tests whether real-time AI support reduces fraud risk at the moment of exposure, and how its effects differ across age and scam type. Grounded in the Person–Task Fit framework, a 2×2 online lab-in-the-field experiment compared AI-assisted vs. unaided judgment across younger (18–30) and older (50+) groups over 10 days using ecologically realistic simulated fraud scenarios across six scam categories. AI-assisted support improved detection accuracy and reduced risky intention, but vulnerability was patterned rather than uniform — and AI produced the largest reductions exactly in each group's most-vulnerable scam categories.
Why it matters
Digital communication has broadened fraud's reach and complexity, making detection more urgent and cognitively demanding. Yet most studies examine static risk profiles or single scams, leaving real-time support in realistic settings underexplored. To address this gap, this study examined an AI-assisted decision support framework for digital fraud prevention. Using a 2 × 2 online lab-in-the-field experimental design, we compared an AI-assisted condition with a control condition across a younger group and an older group. A total of 72 participants completed a 10-day study involving ecologically realistic simulated fraud scenarios across six scam categories. The results showed that AI-assisted support improved fraud detection accuracy and reduced risky behavioral intention. Younger and older groups also showed different scam-specific vulnerability patterns, and AI support reduced participants' willingness to engage in risky actions in response to fraudulent communications in the scam categories to which each group was most vulnerable. Post-study questionnaire ratings and semi-structured interviews further indicated that participants generally perceived the system as useful and realistic, while trust depended on whether the AI output was interpretable and actionable.
Key findings
- AI-assisted support helps Relative to unaided judgment, AI assistance nearly tripled the odds of correctly identifying a scam (OR = 2.94, p < .001) and lowered risky behavioral intention by 0.56 points on a five-point scale (t = −6.22, p < .001).
- Older users more vulnerable overall The older group showed significantly lower detection accuracy (OR = 0.47), higher risky intention, and longer decision time (+7.9 s), though AI support benefited both age groups.
- Vulnerability is scam-specific In the control condition, the younger group had the highest deceptive-response rates for social-media scams (38.9%) and consumer fraud (36.1%); the older group peaked on government/institutional impersonation (38.9%) and financial/payment fraud (36.1%); the age × fraud-type interaction was significant (p = .003).
- AI supports where each group is weakest AI assistance was associated with ~34% lower odds of deceptive responses (OR = 0.66) and produced the largest reductions in exactly the scam categories posing the greatest baseline risk per group.
- Acceptance hinges on interpretability Post-study ratings of usefulness, realism, confidence, and future-use intention all exceeded the midpoint; interviews (four themes, Cohen's κ = 0.89) showed participants trusted the AI more when it explained WHY a message was suspicious and suggested a concrete action — not just a risk label.
2×2 online lab-in-the-field experiment with multilevel regression and qualitative analysis
- 2×2 between-participant design: AI-assisted vs. control × younger (18–30) vs. older (50+); 72 participants (18 per cell) in mainland China and Hong Kong.
- 10-day repeated-exposure web platform delivering realistic fraud scenarios at irregular intervals across six scam categories in email, SMS, app-notification and chat formats.
- Primary outcomes: fraud-detection accuracy (multilevel logistic regression) and risky behavioral intention (linear mixed-effects); secondary: confidence and decision time.
- The AI system output a structured aid per category: a probability-based risk assessment, an explanation of suspicious cues, and a recommended action.
What it means for practitioners
Design AI fraud-prevention tools as interpretable decision aids rather than opaque warning labels, and tailor them to scam type and user profile: for younger users, surface the hidden risk of low-cost, socially embedded promotional scams; for older users, guard against authority-based and institutionally framed deception. Don't fixate on large losses — small payments and low-threshold requests are powerful precisely because they seem too minor to scrutinize.
Get the paper & cite it
Official citation: Yicheng Sun, Jacky W. Keung, Hi Kuen Yu, Yuchen Cao (2026). At the point of risk: Can AI-assisted decision support reduce digital fraud vulnerability across age groups and scam types?. Computers in Human Behavior. DOI: 10.1016/j.chb.2026.109156.