In the model, interventions that alter the feedback channel (verification visibility, social-proof damping) reduce overreliance and regret, while lowering verification cost alone does not lower regret¶

Claim · All claims
Evidence · 1 study (2 entries) · theoretical r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Strong verification visibility (sV = 1.0) triggers a counter-cascade to near-complete verification (overreliance 0.00, verification 1.00) and lowers regret to 0.07. → Ahana Biswas 2026 q2 i? Damping social proof (ρ = 0.6) gives a partial recovery; reducing verification friction (λV = 0.05) is the weakest lever and does not lower regret. → Ahana Biswas 2026 (2)

Evidence¶

Ahana Biswas 2026¶

Ahana Biswas. (2026). Modeling AI Overreliance as a Complex Adaptive System. arXiv. https://arxiv.org/abs/2608.19616 doi:10.48550/arxiv.2608.19616

q2 · i? · theoretical · r3

Intervention simulations run in the danger zone (Low AI, hard tasks, random+tagging) against a harmful baseline of visible unverified use at s = 0.3 (Fig. 5). Making verification visible "moves the population to near-complete verification (overreliance 0.00, verification 1.00) and drives regret down to 0.07".

"makingverificationvisible (s V =1.0) triggers a beneficial counter-cascade—a mirror of the harmful one, with its own tipping point—that, under this strong setting, moves the population to near-complete verification (overreliance 0.00, verification 1.00) and drives regret down to 0.07."

Ahana Biswas 2026 (2)¶

Ahana Biswas. (2026). Modeling AI Overreliance as a Complex Adaptive System. arXiv. https://arxiv.org/abs/2608.19616

q2 · i? · theoretical · r3

Same intervention simulations comparing the two weaker levers. Damping social proof gives "a partial recovery (overreliance 0.39, verification 0.17, regret 0.20)", while friction reduction "lifts verification only to 0.12 and doesnotlower regret (0.24)" because cheaper checking does not counter the social pull.

"Dampening social proof (ρ=0.6) gives a partial recovery (overreliance 0.39, verification 0.17, regret 0.20). Notably, merely reducing verification friction (λV =0.05) is the weakest lever: it lifts verification only to 0.12 and doesnotlower regret (0.24)"

Discussion¶

Related Claims¶

• In the agent-based model, task difficulty and AI quality set the baseline level of overreliance, with difficulty the dominant global factor — related
• In the model, social learning creates consensus in trust without shifting aggregate reliance; connectivity alone does not produce collective overreliance — related
• In the model, visible unverified peer use (social proof) suppresses verification and tips the population into collective overreliance, via a smooth crossover with no hysteresis — related
• In the model, high-quality AI on hard tasks produces the highest regret despite moderate overreliance, because agents over-defer and rarely self-rely — related
• Students given AI assistance in peer feedback tended to rely on the AI-generated feedback and struggled to perform the task when the assistance was removed — related
• Classroom verification norm is predicted to show threshold behavior and hysteresis rather than a smooth response to teacher-set requirements
• Feedback interventions show phase-specific effects: advice feedback reduces reminder bias, but ranking feedback alters beliefs without increasing offloading — related