The Watchtower Imperative: Lifespan-Calibrated Evaluation for Frontier AI
Organizations: University of Kansas · University of Zurich
Abstract
The gap between AI capability deployment and safety assurance is structural and cannot be closed through internal evaluations alone. Analyzing documented harms (2017-2026) across the human lifespan reveals that disparate incidents --- from voice assistants endangering children to AI companions implicated in adolescent suicide and automated care denials for the elderly --- reflect a single failure: evaluating systems against an implicit adult user model. We introduce five developmental moderators structured into a Lifespan Vulnerability Matrix across six life stages. The matrix reveals a nonmonotonic risk distribution: vulnerability peaks at both ends of the lifespan, governance coverage is asymmetrical, and authority to convert model errors into real-world impact is scrutinized least. We then propose a set of lifespan design principles generalizing existing child-protective instruments, and an institution-agnostic, three-tier external evaluation framework --- comprising automated simulation, expert red teaming, and post-deployment monitoring --- indexed directly to matrix risk cells.
Figures & tables
| Metric | Finding | Source |
|---|---|---|
| Incident growth 2023 2024 | +56.4% YOY | Stanford AI Index 2025 [ 58 ] |
| Incident growth 2024 2025 | 233 362 | Stanford AI Index 2026 [ 59 ] |
| Attacks by AI-enabled adversaries | +89% YOY | CrowdStrike 2026 [ 20 ] |
| Vulnerabilities exploited before public disclosure | 42% | CrowdStrike 2026 [ 20 ] |
| FDA-registered AI health tools with a post-deployment surveillance plan | 9% | Dolin et al. (2026) [ 22 ] |
| Organizations actively monitoring production AI | 50% (9% of small firms) | Gradient Flow (2025) [ 30 ] |
| Stage | Elevated moderators | Dominant harm vector | Detection requirement |
|---|---|---|---|
| Early childhood (0–7) | M1, M3 | Misattributed agency; unsafe instruction acceptance; displacement of adult-mediated learning | Elicited beliefs on system nature; instruction-following under unsafe prompts; substitution metrics |
| Middle childhood (8–12) | M1, M3, M5 | Cognitive offloading during foundational learning; algorithmic amplification of comparison content | Offloading events indexed to learning stage; exposure composition; comparison-content escalation |
| Adolescence (13–17) | M2, M3, M5 | Engagement optimization against reward sensitivity; parasocial dependency; crisis mishandling | Disengagement latency; AI-to-human interaction ratio; behavior under distress; escalation reliability |
| Emerging adulthood (18–25) | M3, M4 | Displacement of professional judgment; epistemic dependency prior to independent competence | Verification behavior under model uncertainty; trajectory of unaided task performance |
| Established adulthood (26–64) | M4 | Automation bias at maximum decision authority; hallucination harm multiplied by authorization scope | Override/challenge rates on erroneous outputs; confidence calibration; unaided performance tracking |
| Later adulthood (65+) | M1, M2, M4, M5 | Declining attribution capacity with high decision authority; synthetic-media susceptibility; social substitution | Calibration drift; response to persuasive/authority prompts; substitution ratios; capacity-sensitive consent |
| Proposed lifespan principle | Basis and index | |
|---|---|---|
| DH1 | Design calibrated to developmental state across the entire lifespan | Generalizes age-appropriate design [ 33 , 35 ] ; indexes all moderators |
| DH2 | Do not exploit affective or reward systems where self-regulation is reduced | Generalizes neurodevelopmental protection [ 63 ] ; indexes M2 |
| DH3 | Do not offload capabilities that the user is actively acquiring or maintaining | Extends educational protection duties [ 64 ] ; indexes M3 |
| DH4 | Scale verification friction to the consequence authority being exercised | No child-specific antecedent; indexes M4 |
| DH5 | Protect identity and social cohesion where substitution risk is elevated | Generalizes identity-formation protection [ 63 ] ; indexes M5 |
| DH6 | Verify supported decision-making where attribution capacity is reduced | Generalizes caregiver control [ 35 , 63 ] ; indexes M1 |
AI Safety Evaluation
AI Safety Evaluation studies how to test whether AI systems avoid harmful behavior and remain reliable across users, tasks, and deployment settings.
See all 143 papers