Transparency isn’t trust. It can help establish initial confidence, but beyond a point, it becomes white noise, or worse, a control mechanism driven by latent distrust.

Consider the evolution of ride-hailing apps such as Uber. When Uber introduced live maps to track drivers, this transparency initially built confidence. Compared to old-school radio taxis, where we had no idea where the driver was, real-time visibility created an illusion of control, and thus, trust. But over time, this feature became the bare minimum, not a trust signal. Today, having a map doesn’t make users trust Uber, it’s their consistent reliability, frictionless service, and adherence to customer-aligned values that sustain trust (or not).

This reflects a broader failure - the transparency trap. It is the mistaken belief that more visibility always leads to more trust. If transparency is driven by a prior of distrust, it often creates a negative epistemic spiral. 

Imagine working for someone who constantly asks, “Show me exactly how you did that.” This isn’t about fostering trust, it’s about validating their pre-existing mistrust. Over time, this erodes confidence and creates an unstable, self-fulfilling failure loop. The more we are forced to prove ourselves, the more second-guessing infects the system, and the less trust is built.

This idea has direct consequences for AI alignment. There’s a growing push to make AI systems fully interpretable as a means to "build trust." But this frames the problem incorrectly. Transparency is not the goal, robustness, reliability, and safety are.

Paul Christiano, in his work on AI oversight and corrigibility, suggests that AI systems should be trained to be assistive and corrigible, not necessarily "trustworthy" in a human sense. The goal of AI transparency isn’t to manufacture trust, but to ensure continuous, structured oversight that allows for intervention when necessary (Christiano, 2019).

Similarly, mechanistic interpretability research (Olah et al., 2020, Anthropic, DeepMind) is not about making AI "explain itself" to build trust, it’s about making AI legible enough that we can detect failure modes before they manifest catastrophically. We audit not to make AI seem trustworthy, but to ensure fail-safes against emergent risks in complex, adaptive systems.

Trust should evolve in a Bayesian manner, through repeated interactions in an iterative game where risk exposure is controlled over time. The default assumption should not be that systems are untrustworthy until proven otherwise, but that trust updates as performance data accumulates. Over-engineering transparency as a trust mechanism risks creating fragility instead of robustness.

In the long run, the most trusted systems will not be those that explain themselves the most, but those that fail the least and are most consistently reliable in performance.