The Temporal Dynamics of Human AI Trust

105 concepts98 papers1 proposal

When we evaluate how humans interact with artificial intelligence, we usually take a snapshot. A participant sits down for a brief lab study, prompts a model a few times, and fills out a survey rating their "trust." But in the real world, human-AI collaboration is a movie, not a photograph. As Guo et al. note, brief experimental sessions fail to capture the deep psychological shifts and trust calibration that occur over long-term professional partnerships.

Trust is not a static binary switch; it is a dynamic, state-updating process. When a user first encounters an AI system, their reliance is driven by superficial signals: the fluency of its text, the slickness of its interface, or a general novelty effect. However, as interactions span weeks or months, these initial impressions give way to a more complex, experiential evaluation. Understanding why and how trust evolves over time is critical to building AI systems that remain safe and effective throughout their deployment lifecycle.

The Fleeting Nature of First Impressions

Initial human trust is highly susceptible to "vibes." For instance, Martindale et al. found that users initially over-trust machine translation systems simply because the output sounds fluent, even if it is factually misleading. Similarly, Tambwekar et al. observe that natural-language rationales generate immediate rapport and confidence, but warn that this is often a novelty effect that wears off over extended periods of interaction.

As the novelty fades, users begin to look past the interface and notice systemic quirks and errors. Beck et al. demonstrated that a system's initial reliability—whether it starts out highly accurate or error-prone—does not produce a lasting effect on a human's long-term vigilance. First impressions of technology are fleeting. Instead, users continuously recalibrate their trust based on ongoing performance. Fogliato et al. echo this, showing that while initial alignment or perceived similarity might offer a temporary boost, seeing actual real-world outcomes quickly overrides any initial vibes-based trust.

The Trajectories of Trust and Mental Models

If trust isn't static, what shape does it take over time? In an analysis of collaborative writing, Lu discovered that trust typically follows one of three distinct trajectories:

  1. Consistent growth, as users discover and habituate to the system's capabilities.
  2. A curvilinear arc, where trust peaks early due to high expectations but steadily declines as the user encounters inevitable edge-case failures.
  3. A steady decrease, when early errors permanently break the user's willingness to rely on the tool.
Trust Level
  ^
  |       /---\          (Curvilinear Arc)
  |      /     \____
  |     /____________>   (Consistent Growth)
  |     \
  |      \___________>   (Steady Decrease)
  +--------------------> Time

This shifting behavior happens because humans are constantly running an internal feedback loop. Vafa et al. formalize this as the Human Generalization Function—the mental model people construct to guess what an AI can or cannot do based on its past answers. Every time the AI succeeds or fails, the user updates this function.

Over extended periods, this updating process fundamentally changes how users verify information. As Kim et al. found in real-world deployments, domain experts do not just blindly accept or reject AI advice; they develop highly customized, long-term manual verification strategies tailored to the specific model quirks they have observed over time.

The Dark Side of Extended Exposure

While long-term interaction can lead to healthy calibration, it can also induce cognitive fatigue and complacency. In high-stakes environments like healthcare, Gaube et al. and Raji et al. warn that prolonged exposure to automated advice can exacerbate automation bias or trigger alert fatigue, causing experts to gradually let down their guard and accept incorrect suggestions they would have flagged initially.

Even more profoundly, sustained interaction can subtly erode human agency. Analyzing over a million real-world conversations, Sharma et al. uncovered patterns of situational disempowerment, where users repeatedly interacting with an AI gradually begin to outsource their moral judgments, reality-testing, and personal decisions to the system. Because the AI remains consistently helpful and fluent, the cumulative effect is a slow, unnoticeable shift in the user's own belief systems and autonomy.

Designing for Dynamic Alignment

If trust changes with every turn, AI systems cannot afford to treat human behavior as a fixed parameter. Instead, modern frameworks attempt to track and manage trust dynamically.

To achieve this, Soh introduced the Trust-POMDP, a computational framework that treats human trust as a hidden variable to be managed. By modeling how trust fluctuates based on interaction history, an autonomous system can strategically choose its actions—and even plan around its own potential failures—to keep human expectations perfectly calibrated.

On the algorithmic side, Sreedharan et al. proposed a unifying Bayesian formulation that tracks how human priors evolve over long-term engagement. By explicitly modeling how a user updates their mental model of the agent over time, the system can dynamically adjust its explanations and actions to maintain alignment, ensuring that the partnership remains stable long after the novelty has worn off.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_longitudinal_trust_evolution_4eb5d421,
  title = {longitudinal trust evolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/longitudinal-trust-evolution/rdr_4eb5d42176e34ae1938d4e4882d321bb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_longitudinal_trust_evolution_4eb5d421_v1,
  title = {longitudinal trust evolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/longitudinal-trust-evolution/rdr_4eb5d42176e34ae1938d4e4882d321bb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake