Trust Should Follow Evidence

A specialist AI system asks people to rely on its judgment. That makes trust a practical design question: what has the system demonstrated, under which conditions, and what should happen when the evidence is weak? A useful goal is calibrated reliance, where the authority granted to an agent matches its tested capabilities and the consequences of failure.

The distinction starts with the foundations. Mayer, Davis and Schoorman’s organizational trust model separates trust from the characteristics that help produce it. Ability concerns relevant skills; benevolence concerns the other party’s interests; integrity concerns adherence to acceptable principles. The model also considers the trusting party’s propensity to trust and perceived risk. These dimensions provide a vocabulary for examining relationships. They do not, by themselves, certify an AI system or supply a universal deployment score. Mayer, Davis & Schoorman (1995) ↗

Lerman and Dover bring this vocabulary to language models’ evaluations of humans. Their experiments found partial alignment with human judgments, alongside greater rigidity and demographic effects that varied by model and scenario. These are findings about emulated interpersonal trust. They do not validate a specialist-agent product, demonstrate that fine-tuning inevitably amplifies bias, or establish a tested method for network-wide trust. Prompt sensitivity and limits to generalization remain important. Lerman & Dover (2026) ↗

For a proposed PersonaSI trust architecture, the next step should therefore be concrete evaluation. A specialist agent could be tested on realistic tasks with independently defined success criteria. The test set should include uncertainty, conflicting information and situations requiring escalation. Results should distinguish task accuracy from policy adherence and demographic sensitivity. A single favorable score should not erase a serious failure in another category.

This approach is consistent with NIST’s AI Risk Management Framework. NIST treats trustworthiness as multidimensional and context-dependent. Its measurement guidance calls for documented test sets and metrics, evaluation under conditions resembling deployment, production monitoring, and explicit attention to fairness and bias. It also calls for documenting limits to generalization and involving independent assessors or experts outside the front-line development team. These are evaluation practices, not a certification granted simply by referring to the framework. NIST AI RMF 1.0 ↗

For connected agents, an additional design question is what evidence travels with each handoff. A proposed system could record which agent produced an output, the task it was evaluated for, the relevant model version, and any unresolved uncertainty. The receiving agent could then apply a task-specific acceptance rule. An evaluator’s approval should have a defined scope and expiry condition, particularly when models, prompts or data change. This is a design proposal requiring its own validation.

Any public description should make the boundary between evidence and ambition visible. Research findings belong beside their sources. Proposed discovery, orchestration and verification roles should be labeled as proposed. Audit profiles and marketplace admission rules should be presented as implemented capabilities only when their operation can be demonstrated. The useful promise is a process people can inspect: defined tests, visible limitations, accountable decisions and a clear route to challenge an outcome.