ACE, Action and Control via Explanations: A Proposal for LLMs to Provide Human-Centered Explainability for Multimodal AI Assistants

Explainability enables user control: The paper proposes that Large Language Models can generate human-readable "semantic frames" that help end users understand AI decisions and provide feedback to improve the system, rather than treating AI systems as black boxes that users must simply accept.

Participatory design is essential for manufacturing AI: The authors identify that involving actual workers in the design and training of multimodal AI assistants (which combine vision, speech, and document inputs) is critical for building systems that genuinely support human performance rather than creating new problems.

Collaboration improves accuracy: By using explanations to facilitate back-and-forth communication between humans and AI systems, the ACE paradigm aims to build more accurate models of human activities and behaviors, suggesting that human oversight and input during deployment—not just during initial training—leads to better outcomes.

Multimodal AI requires human alignment: The paper emphasizes that AI systems combining multiple input types (computer vision, speech recognition, documents) need continuous human guidance to align their different components, highlighting that technical integration alone is insufficient without human-centered design.

Transparency supports accountability: By making AI reasoning explicit through semantic frames, the approach enables workers to verify whether the system is making decisions for the right reasons, supporting accountability in high-stakes domains like manufacturing where errors can affect worker safety and productivity.