TL;DR
This paper introduces a two-stage "activation-integration" model to explain how humans use AI advice. The core finding is that while our prior beliefs about AI vs. human performance determine if we listen to a suggestion, once we decide to use it, we incorporate AI and human advice in exactly the same way.
What Problem Does This Paper Address?
As AI becomes a standard fixture in decision-support roles—like a "human-in-the-loop" medical diagnostic tool—understanding the psychology of trust is paramount. Do we blindly follow algorithms (automation bias) or irrationally reject them (algorithm aversion)?
Previous research has been inconsistent, often hampered by small sample sizes or limited task types. This paper seeks to provide a comprehensive, large-scale analysis of how humans incorporate advice from AI versus human peers. The goal is to move beyond just measuring if people trust AI and start modeling how they process that information.
Methods
The researchers employed the Judge-Advisor Paradigm (JAS), a standard psychological framework. Participants make an initial judgment, receive advice, and then provide a final judgment.
To isolate the effect of the "AI" label, the researchers used a clever experimental design: the advice shown in both the "AI" and "Human" conditions was identical. It was generated from the average responses of a previous group of human labelers. This ensured that any difference in behavior was due to the participant's perception of the source, not the quality of the advice itself.
The authors proposed and validated the Activation-Integration Model:
- Activation Stage: The user decides whether to modify their initial response at all.
- Integration Stage: If "activated," the user decides how much weight to give the advice relative to their own knowledge.
Datasets
The study involved over 1,100 crowdworkers and a group of expert dermatologists across five diverse tasks:
- Art (Image): Identifying art periods (e.g., Renaissance vs. Baroque). Art Task
- Cities (Image): Identifying US cities from photos. Cities Task
- Sarcasm (Text): Detecting sarcasm in Reddit posts. Sarcasm Task
- Census (Tabular): Predicting income levels from demographic data. Census Task
- Biopsy (Expert): Dermatologists deciding whether to biopsy skin lesions based on images.
Key Insights and Ideas
- Prior Beliefs are the Gatekeepers: The "Activation" stage is heavily influenced by a person's prior belief about who is better at a specific task. People generally believed AI is better at visual tasks (Art, Cities) and humans are better at social/linguistic tasks (Sarcasm).
- Integration is Source-Agnostic: This is the "aha moment." Once someone decides to change their answer, they shift it by the same amount regardless of whether they think the advice came from a computer or a person. The source affects the "gate," but not the "weight."
- Confidence Dynamics: Users are more likely to use advice when they are unconfident in their own initial guess or when the advice is presented with high confidence.
- Experts vs. Laypeople: Interestingly, expert dermatologists showed almost no difference in how they used human vs. AI advice, suggesting that professional expertise might mitigate some of the biases seen in the general public.
Key Results
- Accuracy Gains: Advice improved performance across the board. In the Art task, AI advice led to an 11.8% accuracy boost compared to 7.1% for human advice, simply because people were more likely to "activate" for the AI.
- Activation Differences: In the Sarcasm task, the trend reversed: the activation rate was 39.8% for human advice vs. 34.0% for AI advice.
- Global Consistency: These patterns were remarkably consistent across participants from the US, UK, and Asia.
- Integration Parity: Statistical tests (Kolmogorov–Smirnov) showed no significant difference in the distribution of the "Weight of Advice" between human and AI conditions across all four layperson datasets.
Limitations and Future Work
- Anonymous Peers: In this study, "human advice" came from an unknown peer. In real-world professional settings, advice often comes from trusted, known colleagues, which might change the integration dynamics.
- Simple Advice Formats: The study focused on simple labels and confidence scores. The authors suggest that future work should examine more complex advice, such as AI-generated explanations (XAI).
- Static Interactions: The experiments were one-off tasks. Real-world trust is often built (or destroyed) over repeated interactions, a dynamic not captured here.
Summary updated