Human--AI Interactions as First-Class Data: A Formal Interaction Model and Runtime for Explainable AI Collaboration

Michela Taufer; Kin Ng; Ian Lumsden; Aashish Pandey; Rajasi Gore; Angelina Ju; Sukanya Moudgalya; Ewa Deelman

cs.AIOct 29, 2026

Human--AI Interactions as First-Class Data: A Formal Interaction Model and Runtime for Explainable AI Collaboration

Authors: Michela Taufer, Kin Ng, Ian Lumsden, Aashish Pandey, Rajasi Gore, Angelina Ju, Sukanya Moudgalya, Ewa Deelman,

Organizations: University of Tennessee, Knoxville, USA · University of Southern California, Marina del Rey, CA, USA

Abstract

AI assistants are increasingly embedded in computational notebooks, scientific workflows, software engineering tools, and learning systems, yet human-AI interactions are often treated as transient prompt-response exchanges rather than durable data products, hindering reproducibility, explainability, replay, behavioral analysis, and adaptive AI support. We argue that these interactions form a new data class requiring formal representation and introduce the Human-AI Interaction Description Language (HAIDL), a formal interaction model with a Backus-Naur Form-inspired concrete syntax. HAIDL encodes each interaction as a structured object capturing computational context, user request, prompt payload, AI response, feedback, problem context, provenance, and model metadata. We integrate this vision into HAIDL-X, a middleware for JupyterLab that instantiates HAIDL by capturing user-bot state, building contextualized prompts, interfacing with AI providers, collecting feedback, and storing interactions in a repository. We add a learning layer over HAIDL that maps interaction objects and notebook execution traces into reproducible measures of user behavior, including engagement, performance, help-seeking, error-oriented use, and learning personas. HAIDL-X provides a foundation for treating human-AI collaboration as structured, queryable, replayable, analyzable data.

Figures & tables

Fig. 1: Lifecycle of a human–AI interaction. A user request arises from a computational state, the AI returns guidance, and the user evaluates or acts on it. HAIDL represents the full lifecycle as a structured artifact whose outcome seeds the next interaction.

Fig. 2: Conceptual organization of a HAIDL interaction object: a single structured object with seven semantic components, forming the atomic unit of human–AI collaboration and the root production of the HAIDL grammar.

	Interaction Events				Learning Profiles
Assign.	Error	Stuck/Next	Concept.	Valid.	HPHE	HPLE	LPHE	LPLE
03	44	67	2	21	19	0	0	16
04	106	0	163	14	18	0	0	17
07	31	37	5	0	22	0	8	5
08	150	13	119	3	19	0	6	10

TABLE I: Derived interaction events and notebook-derived learning profiles by assignment.

	Exposure States			Behavioral States
Assignment	Exp. User	Exp. Non-User	Control	Self-suff.	Diseng.	Scaffolded	Expl. Over.	Cross-Cond.
03	9	8	12	4	4	8	1	6
04	9	9	7	7	3	8	0	10
07	7	10	14	5	5	7	0	4
08	8	10	11	7	5	6	0	6

TABLE II: HAIDL-X exposure states and derived behavioral states by assignment.

Panel of Figure 5 — Fig. 3: Per-model overhead distributions across concurrent clients and three providers (Azure, vLLM, Sambanova), showing consistently low, provider-agnostic overhead.

Measure	Median	P95
HAIDL object size (KB)	4.53	61.79
Persistence latency (ms)	1.40	1.97
Retrieval: student–assignment (ms)	0.399	0.581
Retrieval: error-containing (ms)	0.399	0.711
Retrieval: ordered trajectory (ms)	0.401	0.652

TABLE III: Persistence and retrieval characteristics of HAIDL objects and queries.

Captured information	OTel	PROV-	Prog-	HAIDL
	GenAI	AGENT	Snap2
Prompt, response, model metadata	✓	✓	—	✓
Computational state motivating a request	—	✓ ‡	✓ †	✓
User assessment of the response	—	—	—	✓
User actions after the response	—	—	✓ †	✓
Interaction as an explicit state transition	—	—	—	✓

TABLE IV: What existing representations capture vs. HAIDL. Columns: AI observability (OTel GenAI [ 3 ] ), agent-centric provenance (PROV-AGENT [ 2 ] ), and programming-process data (ProgSnap2 [ 6 ] ).

Human-AI Interaction

Human-AI interaction studies how people and AI systems communicate, collaborate, make decisions, and shape one another over time.

See all 333 papers

HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
Sep 10, 2025Benjamin Sturgeon, Daniel Samuelson, Jacob Haimes +1LLM Agent Evaluation Human-AI Interaction
Apart Research · AI Safety Cape Town · University of Chicago +2
RECAP: An End-to-End Platform for Capturing, Replaying, and Analyzing AI-Assisted Programming Interactions
May 1, 2026Keyu He, Qianou Ma, Valerie Chen +2Human-AI Interaction AI-Assisted Software Development
Carnegie Mellon University
Transfiver: Human-AI Co-Inference through a Shared Editable State
Sep 3, 2026Minji Park, Seunghyun Yoon, Hyuk LimHuman-in-the-Loop AI Human-AI Interaction
Korea Institute of Energy Technology (KENTECH)

Sep 10, 2025J/K move · Enter open · S save

HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants

Benjamin Sturgeon, Daniel Samuelson, Jacob Haimes, Jacy Reese Anthis

Apart Research · AI Safety Cape Town · University of Chicago+2

LLM Agent Evaluation Human-AI Interaction AI Agent Benchmarks AI Ethics

Open paper AskPDF arXiv

As humans delegate more tasks and decisions to artificial intelligence (AI), we risk losing control of our individual and collective futures. Relatively simple algorithmic systems already steer human decision-making, such as social media feed algorithms that lead people to unintentionally and absent-mindedly scroll through engagement-optimized content. In this paper, we develop the idea of human agency by integrating philosophical and scientific theories of agency with AI-assisted evaluation methods: using large language models (LLMs) to simulate and validate user queries and to evaluate AI responses. We develop HumanAgencyBench (HAB), a scalable and adaptive diagnostic tool for six behaviors related to human agency. HAB measures the tendency of an AI assistant to Ask Clarifying Questions, Avoid Value Manipulation, Correct Misinformation, Defer Important Decisions, Encourage Learning, and Maintain Social Boundaries. We find low-to-moderate agency support in contemporary LLM-based assistants, with substantial variation across system developers and behaviors. For example, while Anthropic LLMs most support human agency overall, they are the least supportive LLMs in terms of Avoid Value Manipulation. These behaviors do not appear to consistently result from increasing LLM capabilities or instruction-following (e.g., RLHF); we encourage further study of these behaviors so that developers and users can better understand the complexities of modern human-AI interaction.

Abstract

Human-AI Interaction

Explore similar work

HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants

RECAP: An End-to-End Platform for Capturing, Replaying, and Analyzing AI-Assisted Programming Interactions

Transfiver: Human-AI Co-Inference through a Shared Editable State