How Domain Expertise Alters Human AI Interaction

39 concepts36 papers1 proposal

Imagine designing an AI assistant to help professionals diagnose diseases, detect financial fraud, or assess legal risks. To ensure the system is safe and trustworthy, you conduct a user study using crowdsourced workers on a platform like Amazon Mechanical Turk. The participants find the AI's explanations intuitive, calibrate their trust perfectly to its accuracy, and show significant performance gains.

But when you deploy the system to actual clinicians or financial analysts, it fails. The experts find the explanations distracting, ignore the system's uncertainty signals, or reject the tool entirely.

This breakdown occurs because of a fundamental generalizability gap in human-AI interaction research. A growing body of literature demonstrates that domain experts do not merely possess more knowledge than laypeople; they use entirely different cognitive heuristics, mental models, and verification strategies when interacting with automated systems. Consequently, design principles derived from layperson studies frequently fail when applied to professional environments.

The Verification Loop vs. Blind Reliance

When a layperson interacts with an AI recommendation, they lack the underlying domain knowledge required to evaluate its correctness independently. As a result, they rely heavily on superficial heuristics—such as the linguistic fluency of a large language model (LLM) or the confidence score displayed on an interface—to decide whether to accept or reject a suggestion.

For domain experts, the interaction dynamic is an active verification loop rather than a passive reliance choice. As shown by Sunnie S. Y. Kim et al. in Humans, AI, and Context: Understanding End-Users' Trust in a Real-World Computer Vision Application, experts treat their own knowledge as a primary calibration tool. They cross-examine the AI’s output against their internal mental models, selectively accepting or rejecting suggestions based on specific evidence rather than general trust.

This ability to independently adjudicate outputs fundamentally changes how advice is integrated:

  • Handling Conflicting Advice: When presented with primary AI recommendations and secondary opinions, laypeople rely on simple heuristics to resolve friction. In contrast, Zhuoran Lu et al. demonstrate in Does More Advice Help? The Effects of Second Opinions in AI-Assisted Decision Making that professional experts leverage their background knowledge to independently verify and resolve conflicting suggestions.
  • Resistance to Persuasion: LLMs can generate highly persuasive arguments that easily sway non-experts. However, Takehiro Takayanagi et al. show in Can GPT-4 Sway Experts' Investment Decisions? that domain expertise acts as a buffer against this influence. While amateurs find AI-generated financial reports highly convincing, experts prioritize logical depth and professional human analysis, making them significantly more resistant to algorithmic persuasion.

Why Generic Explanations Fail Experts

A foundational assumption in Explainable AI (XAI) is that providing the "why" behind an algorithmic prediction improves human decision-making. However, the utility of an explanation is highly sensitive to the user's prior knowledge.

When an AI's output already aligns with an expert's robust mental model, providing an additional explanation offers no marginal benefit to trust or performance, as noted by Brent Mittelstadt et al. in The Impact of Explanations on Layperson Trust in Artificial Intelligence–Driven Symptom Checker Apps: Experimental Study. Low-knowledge scenarios require targeted information to bridge specific gaps, but experts operate under a much higher standard of evidence.

Layperson Interaction:
[AI Output + Explanation] ───► Evaluated via Fluency/Surface Cues ───► High Acceptance

Expert Interaction:
[AI Output + Explanation] ───► Evaluated via Internal Mental Model ──► Strict Scrutiny / Distraction

When explanations are provided for complex tasks, experts evaluate their quality far more strictly than laypeople. In Can Current Explainability Help Provide References in Clinical Notes to Support Humans Annotate Medical Codes?, Byung‐Hak Kim et al. found that while laypeople found automated text justifications highly informative, professional medical coders judged the exact same evidence as insufficient or irrelevant for specialized workflows.

Furthermore, popular feature-attribution methods (such as SHAP or LIME) can actively hinder expert performance. Kasun Amarasinghe et al. discovered in On the Importance of Application-Grounded Experimental Design for Evaluating Explainable ML Methods that these explanations failed to improve the decisions of fraud analysts and instead slowed them down. Because advanced practitioners already possess optimized internal search and decision strategies, automated prompts and visual aids often become more distracting than helpful, an effect documented by Vincent Dörig et al. in Harmonizing Assistance: Moderating Visual and Textual Aids in AI-Enhanced Textbook Reading with IRead.

Divergent Heuristics for Uncertainty and Risk

How users interpret model confidence and predictive uncertainty is also deeply mediated by expertise. In layperson studies, researchers often use behavioral "nudges" or specific calibration techniques to align human trust with model accuracy. However, as Kailas Vodrahalli et al. argue in Uncalibrated Models Can Improve Human-AI Collaboration, these insights do not generalize to domain experts, who process confidence signals through entirely different decision-making heuristics.

For instance, when visualizing machine learning model uncertainty, a user's background dictates how that visualization is parsed. Sean McGrath et al. demonstrate in When Does Uncertainty Matter?: Understanding the Impact of Predictive Uncertainty in ML Assisted Decision Making that ML experts interpret uncertainty distributions differently than non-experts, showing a distinct preference for normally distributed representations with low variance.

Similarly, professional risk managers weigh data variance and model reliability differently than laypeople when facing compound uncertainties, as explored by Lace Padilla and Maia Powell in Uncertain About Uncertainty: How Qualitative Expressions of Forecaster Confidence Impact Decision-Making With Uncertainty Visualizations. An expert's tolerance for error is bounded by professional standards, meaning they cannot easily be nudged into accepting uncalibrated or highly variant outputs.

The Methodological Generalizability Gap

The systematic divergence between experts and laypeople exposes a critical flaw in current human-AI interaction research: the widespread reliance on crowdsourced proxies.

A vast number of behavioral experiments utilize crowdsourced workers or student populations as stand-ins for professionals because they are cheaper and easier to recruit (Alex Chouldechova et al. in The Impact of Algorithmic Risk Assessments on Human Predictions and its Analysis via Crowdsourcing Studies; Ashley Suh et al. in Visual Validation versus Visual Estimation: A Study on the Average Value in Scatterplots).

However, simplified tasks performed by non-experts fail to capture the high stakes, institutional identities, and specialized training inherent to professional domains. As argued by Andi Peng et al. in Investigations of Performance and Bias in Human-AI Teamwork in Hiring and Shiye Cao et al. in Designing for Appropriate Reliance: The Roles of AI Uncertainty Presentation, Initial User Decision, and User Demographics in AI-Assisted Decision-Making, professionals like recruiters or clinicians exhibit entirely different patterns of reliance and bias compared to laypeople.

When researchers evaluate trust using general users, they miss the unique skepticism and mental models of domain experts (Mohsen Abbaspour Onari et al. in Assessing and Quantifying Perceived Trust in Interpretable Clinical Decision Support). For instance, Melina A. Kunar demonstrates in Increasing transparency of computer-aided detection impairs decision-making in visual search that explicitly stating system accuracy to medically naïve participants can backfire, triggering overtrust and impairing their visual search performance. Yet, a professional clinician or a judge operates under institutional constraints and professional liabilities that fundamentally alter how they interact with algorithmic risk assessments (Ben Green et al. in Algorithmic risk assessments can alter human decision-making processes in high-stakes government contexts).

To build safe, effective AI systems for specialized fields, the research community must move away from generic transparency strategies and design for the specific cognitive workflows, verification habits, and risk profiles of the experts who will actually use them.

Key Foundational Work

Human Decision Making with Machine Assistance: An Experiment on Bailing and Jailing (2019) — contributed 1 concept elements to this direction.
How Children Seek Out Information from Technological and Human Informants (2015) — contributed 1 concept elements to this direction.
Exploring Computational User Models for Agent Policy Summarization (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_expert_layperson_divergence_630e0b92,
  title = {expert-layperson divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/expert-layperson-divergence/rdr_630e0b924abe414bb14d753305265616},
  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_expert_layperson_divergence_630e0b92_v1,
  title = {expert-layperson divergence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/expert-layperson-divergence/rdr_630e0b924abe414bb14d753305265616/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