Can AI Extract Antecedent Factors of Human Trust in AI? An Application of Information Extraction for Scientific Literature in Behavioural and Computer Sciences
By
Melanie McGrath|Harrison Bailey|Necva Bölücü|Xiang Dai|Sarvnaz Karimi|Cecile Paris
Artificial intelligence is rapidly changing the world, but can we trust it? Researchers are exploring how to teach AI to understand the complex factors that build (or break) human trust in intelligent systems. This isn't just a philosophical question—it has real-world implications for the design and adoption of AI. A new study tackles this challenge by using information extraction techniques to analyze scientific literature on human-AI interaction. Think of it like training an AI detective to uncover the hidden clues in texts, revealing what makes people trust (or distrust) AI. This involves identifying 'trust factors'—everything from the AI's performance and explainability to a person's prior experiences and the specific context of interaction. The researchers created a specialized dataset called "Trust in AI" to train their AI models. Interestingly, they found that while large language models (LLMs) show promise, good old-fashioned supervised learning methods still outperform them on this complex task. This suggests that decoding human trust requires more than just pattern recognition—it needs a deeper understanding of human psychology and behavior. The research highlights the importance of creating structured datasets for complex domains like trust in AI. Future work will likely focus on refining these models, expanding the dataset to other languages, and addressing ethical considerations around data usage. Ultimately, this research aims to build a bridge between humans and AI, creating systems that are not only intelligent but also trustworthy.
🍰 Interesting in building your own agents?
PromptLayer provides the tools to manage and monitor prompts with your whole team. Get started for free.
Question & Answers
What information extraction techniques were used to analyze scientific literature on human-AI trust, and how did they compare to large language models?
The research employed supervised learning methods alongside large language models (LLMs) for information extraction from scientific texts. The supervised learning approaches surprisingly outperformed LLMs in analyzing trust factors in AI literature. The process involved: 1) Creating a specialized 'Trust in AI' dataset for training, 2) Implementing both traditional supervised learning and LLM-based approaches, and 3) Comparing their performance in identifying trust-related factors. For example, this could be applied in analyzing academic papers to automatically extract key factors that influence user trust in medical AI systems, helping developers build more trustworthy healthcare applications.
What are the key factors that influence human trust in AI systems?
Human trust in AI systems is influenced by multiple interconnected factors. The main elements include the AI's performance accuracy, transparency (how well it can explain its decisions), and reliability over time. User-specific factors also play a crucial role, such as prior experiences with AI systems and individual comfort levels with technology. These factors matter because they directly impact AI adoption rates across industries. For instance, in healthcare, patients are more likely to accept AI-driven diagnoses when the system can clearly explain its reasoning and has a proven track record of accuracy.
How can businesses build more trustworthy AI systems?
Building trustworthy AI systems requires a multi-faceted approach focusing on transparency, reliability, and user experience. Companies should prioritize explainable AI technologies that can clearly communicate their decision-making processes to users. Regular performance testing and validation ensure consistent reliability. Additionally, incorporating user feedback and maintaining clear communication about the AI's capabilities and limitations helps build trust. For example, a financial institution could build trust by showing customers exactly how their AI makes investment recommendations and maintaining transparent performance metrics.
PromptLayer Features
Testing & Evaluation
The paper's comparison of supervised learning vs LLM performance aligns with PromptLayer's testing capabilities for comparing different model approaches
Implementation Details
Set up A/B tests between traditional supervised models and LLMs using PromptLayer's testing framework, track performance metrics, and analyze results systematically
Key Benefits
• Quantitative comparison of different model approaches
• Reproducible evaluation pipeline
• Systematic performance tracking across iterations
.png){kind=icon}
.svg)
.svg)
.svg)
.svg)
