Negotiative Alignment: An interactive approach to human-AI co-adaptation for clinical applications

Florence Xini Doo, Nikhil Shah, Pranav Kulkarni, Vishwa Sanjay Parekh, Heng Huang

Keywords: Negotiative alignment, Radiology, Clinical AI, Behavioral alignment, Dynamic and lifelong alignment, Human-in-the-loop learning, Continuous error detection, High-stakes alignment, Co-adaptation, Healthcare AI, Situated interaction, Co-evolving alignment

TL;DR: We introduce a framework for iterative, feedback-driven alignment in clinical radiology AI, showing how error detection and graded severity scoring can guide more aligned and trustworthy AI.

Abstract: We introduce a conceptual framework for ***negotiative alignment*** in high-stakes clinical AI, where human experts iteratively refine AI outputs rather than a binary accept/rejection. This approach uses graded feedback---including partial acceptance of useful insights---to systematically flag and score different types of clinical AI output errors. Although we do not present finalized experimental results, we outline a proof-of-concept using a chest radiograph image-report dataset and a multimodal model. These severity-scored errors might guide future targeted model updates. Negotiative alignment grounds each AI-generated report in a continuous, co-adaptive dialogue with clinicians, which has the potential to boost trust, transparency, and reliability in medical diagnostics and beyond.

Submission Type: Tiny Paper (2 Pages)

Archival Option: This is an archival submission

Presentation Venue Preference: ICLR 2025

Submission Number: 98