Automated rationale generation: a technique for explainable AI and its effects on human perceptions
summary
Engineers build a Frogger environment, collect a think‑aloud corpus with automated speech‑to‑text and real‑time edits, train a sequence-to-sequence neural rationale generator, and test two configurations Complete View (full screen + noise) and Focused View (local window). A 128‑participant MTurk study shows generated rationales beat random baseline and align with human judgments on confidence, humanlikeness, justification and understandability; a follow‑up within‑subject study finds Complete View wins for confidence, failure explanation and unexpected behavior because added detail improves mental models. Proof‑of‑concept works but is data and labor intensive.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:35Rationale generation approach
- 3:22Data collection pipeline
- 5:31Model configurations and training
- 6:56Perception study results
- 11:11Preference comparison study
- 15:03Summary and future directions
- 17:17Q&A session
related talks
42mReasonable AI
01h 05mAdaptive Generalization: The Role of Dynamic Evaluation in Low-Resource Settings
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
15mIntroduction to Explainable AI
43mAI Engineer Ethics: Shaping Responsible Generative AI
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
24mThe Rise of Reasoning Models
46mCan you trust your AI?