Large Language Model (LLM) and Human Performance in Child Investigative Interviewing Question Formulation Tasks.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Forensic Psychology

Background:

• Investigative interviewing of children is crucial for legal proceedings.
• Assessing the capabilities of large language models (LLMs) in specialized communication tasks is an emerging area of research.
• Understanding question formulation strategies is key to effective interviewing.

Purpose of the Study:

• To compare the question formulation performance of LLMs and humans with varying expertise in child investigative interviewing.
• To evaluate LLM performance in both static and dynamic interview simulation tasks.
• To identify strengths and weaknesses of LLMs in interview contexts.

Main Methods:

• Two tasks were used: a static Interview Excerpt Task and a dynamic Avatar Interviewing Task.
• Participants included psychologists, naive individuals, students, and LLMs (GPT-4, Llama-2).
• Performance was evaluated based on question types, including recommended, non-recommended, invitations, and option-posing questions.

Main Results:

• In the dynamic task, LLMs used fewer recommended questions and more non-recommended questions than human professionals.
• In the static task, GPT-4 outperformed psychologists in using invitations and avoiding option-posing questions.
• LLMs demonstrated strong question formulation in controlled settings but faltered in adaptive dialogue.

Conclusions:

• LLMs possess potential for specific question formulation aspects in child interviews, particularly in static scenarios.
• LLMs require further development to effectively handle the nuances of adaptive, real-time dialogue in investigative interviews.
• Human expertise remains critical for navigating complex, dynamic interview situations with children.