HAIC: Human-AI Collaboration Overview
- HAIC is human-AI collaboration, a structured partnership where human expertise and AI capabilities jointly pursue common goals through adaptive decision-making.
- Evaluation frameworks for HAIC emphasize three modes—Human-Centric, AI-Centric, and Symbiotic—with metrics like user satisfaction, system accuracy, and dynamic task allocation.
- HAIC applications span fields such as healthcare, manufacturing, software engineering, and robotics, addressing domain-specific challenges via guided deferral, knowledge transfer, and iterative co-design.
HAIC commonly denotes Human-AI Collaboration: a cooperative partnership between humans and AI systems working toward common goals, characterized by joint decision-making, mutual learning, adaptation, and shared outcomes (Fragiadakis et al., 2024). In this sense, HAIC is not merely human use of an AI tool; it is a dynamic, reciprocal, and context-dependent arrangement whose performance depends on the human contribution, the AI contribution, and the joint performance of the combined system. The term also appears in adjacent research programs on human-centered AI and, in other subfields, as an unrelated acronym for distinct datasets and control frameworks, so its meaning is partly domain-specific (Xu et al., 2021).
1. Conceptual scope and defining properties
Within the human-AI collaboration literature, HAIC is distinguished from conventional human-machine interaction and from narrow automation benchmarks by its emphasis on reciprocity, adaptation, and shared outcomes (Fragiadakis et al., 2024). Conventional HMI evaluation often prioritizes task completion time, error rate, and user satisfaction; those measures remain useful, but they do not by themselves capture trust calibration, quality of shared decision-making, how humans and AI adapt to one another, whether task allocation shifts appropriately over time, or whether transparency, fairness, accountability, and bias mitigation are handled within the collaboration itself. This distinction is foundational: the object of evaluation is not only a tool or model, but an effective collaborative arrangement.
The literature also frames HAIC as a socio-technical and cognitive partnership. In organizational knowledge-retention work, AI is treated not only as a decision aid but as a knowledge-bearing collaborator that can function as a trainer for novices, preserving task-specific expert knowledge when subject matter experts depart (Spitzer et al., 2022). In software engineering, HAIC is described as a human-AI team-level concept within behavioral software engineering, with AI acting as a collaborative partner or cognitive partner with roles and responsibilities rather than as a mere utility (Rani, 13 Apr 2025). In requirements engineering, HAIC is explicitly defined as bidirectional interactions where both human expertise and AI capabilities contribute to outcomes (Rani et al., 3 Nov 2025).
A recurring misconception is that HAIC is equivalent either to full AI automation or to passive AI validation. The reviewed literature rejects both reductions. AI-only decision pipelines exclude the human contribution altogether, whereas passive validation leaves AI in a confirmatory role rather than an active collaborative one. HAIC instead denotes settings in which role allocation, information asymmetry, and joint performance are all first-class design variables (Leitão et al., 2022).
2. Collaboration modes, evaluation factors, and metric selection
A central methodological formulation treats HAIC as varying along three collaboration modes—Human-Centric, AI-Centric, and Symbiotic—distinguished mainly by task allocation and decision authority (Fragiadakis et al., 2024). In the Human-Centric mode, humans retain primary decision authority and AI augments human capability; in the AI-Centric mode, AI is the dominant agent and evaluation emphasizes autonomous performance, efficiency, and uncertainty handling; in the Symbiotic mode, humans and AI act as interdependent partners, with two-way interaction, shared decision-making, continuous feedback exchange, and dynamic reassignment of roles.
| Mode | Decision authority and interaction | Indicative metrics |
|---|---|---|
| Human-Centric | Humans retain primary decision authority; AI is supportive | Clarity of Communication, Ease of Use, User Satisfaction, Teaching Efficiency |
| AI-Centric | AI is the dominant agent; interaction is more one-directional | Learning Curve, Model Improvement Rate, Response Time, System Accuracy |
| Symbiotic | Two-way interaction and dynamic task reassignment | Adaptability Score, Dynamic Task Allocation, Expertise Utilization, Decision Effectiveness |
The same framework organizes evaluation around three primary factors: Goals, Interaction, and Task Allocation. Goals divide into Individual Goals, such as Learning Curve, Prediction Accuracy, and Teaching Efficiency, and Collective Goals, such as Overall System Accuracy and Objective Fulfillment Rate. Interaction divides into Communication Methods, Feedback Mechanisms, Adaptability, and Trust and Safety, with metrics including Clarity of Communication, Feedback Quality, Adaptability Score, Trust Score, Safety Incidents, Error Reduction Rate, and Confidence. Task Allocation divides into Complementarity, Flexibility, Efficiency, Responsiveness, Collaborative Decision Making, Continuous Learning, Mutual Support, and Robustness, with metrics such as Query Efficiency, Dynamic Task Allocation, Model Improvement Rate, Response Time, Impact of Corrections, Knowledge Retention, Task Completion Time, Adversarial Robustness, Domain Generalization, and System Reliability.
The framework’s operational core is a decision tree. Evaluators first identify the collaboration mode, then answer mode-specific questions, then select metrics from the factor families. This produces a tailored but structured assessment rather than a fixed benchmark. The paper also proposes, but does not formalize, an overall weighted scoring mechanism in which AI-Centric systems may weight accuracy and efficiency more heavily, Human-Centric systems may weight ease of use and trust more heavily, and Symbiotic systems may distribute weights more evenly across collaboration-oriented metrics. The weighting scheme is explicitly described as requiring future empirical validation.
Quantitative, qualitative, and mixed-methods approaches all appear in the literature. Quantitative approaches dominate and use sensitivity, specificity, precision, recall, F1-score, detection rates, response times, and process-performance measures. Qualitative approaches rely on interviews, focus groups, case studies, and thematic analysis, especially in healthcare ethics, design, creativity, and mental-health applications. Mixed-methods approaches are treated as especially promising because many collaboration properties—communication quality, trust, perceived fairness, and workflow fit—cannot be fully captured by performance numbers alone.
3. Decision delegation, knowledge transfer, and iterative co-design
One major HAIC strand concerns decision routing: who should decide which instances, under what uncertainty conditions, and with what information handoff. A prominent formalism is Learning to Defer (L2D), in which a system learns a deferral variable that routes an instance either to the model or to a human reviewer while optimizing combined system loss (Leitão et al., 2022). The critique is not that L2D is mathematically uninteresting, but that it often assumes human predictions for every training instance, stable i.i.d. data, fully observed outcomes, and static reviewer capacity. The paper argues that realistic HAIC is constrained by missing human predictions, non-random assignment, selective labels, multiple heterogeneous experts, finite human capacity, fairness concerns, and dynamic environments. This pushes HAIC beyond supervised abstention toward a broader socio-technical resource-allocation problem.
In healthcare, guided deferral extends this logic by changing not only when the AI defers but how it defers. A guided deferral system for lumbar spine MRI report parsing with open-source LLMs produces a structured explanation, a probability, a top reason for, a top reason against, and a final conclusion; uncertainty is quantified by distance to the decision boundary, with confidence defined as (Strong et al., 2024). In the reported pilot study, all participants improved with guidance, and guided humans outperformed both unguided humans and the LLM on deferred uncertain cases. A core technical result is that hidden-state-derived prediction was better calibrated than verbalized confidence and yielded the best deferral performance.
A second strand treats HAIC as knowledge transfer. In a training-oriented framework, HAIC has two stages: Stage , SME-AI collaboration for knowledge induction into AI, and Stage , AI-novice collaboration for novice training through HAIC (Spitzer et al., 2022). The AI is trained with SME-provided labeled data until it can perform the task sufficiently, then collaborates with novices by revealing predictions and explanations through XAI. The central hypotheses are that training novices with no prior knowledge of a particular task can be conducted through HAIC, and that AI predictions supplemented with explanations enhance the novice’s comprehension of the task and thus advance the training process. The paper is explicit that task-specific expert knowledge is not replicated but rather transferred, and that tacit knowledge may be only partially captured.
A third strand frames HAIC as iterative co-design rather than static assistance. In clinical prediction modeling, HACHI alternates between an AI agent rapidly exploring and evaluating candidate concepts in clinical notes and clinical and domain experts providing feedback to improve the CPM learning process (Feng et al., 14 Jan 2026). Concepts are defined as simple yes-no questions used in linear models, making the resulting CPM fully interpretable. The human team does not merely validate outputs; it changes prompts, concept granularity, weighting schemes, dataset inclusion criteria, and even objective-function behavior when it detects leakage, bias, poor generalization, or clinically implausible coefficient directions.
4. Domain applications and organizational forms
The evaluation literature repeatedly stresses that HAIC is domain-specific in its risks and values. In manufacturing, HAIC is presented mainly as Symbiotic; the principal challenges are safety, worker trust, context-sensitive adaptation, and maintaining productivity, so evaluation emphasizes Adaptability Score, Error Reduction Rate, Confidence, and Task Completion Time (Fragiadakis et al., 2024). In healthcare, collaboration typically occurs between AI diagnostic support and human clinicians, and priorities include System Accuracy, Prediction Accuracy, Response Time, Clarity of Communication, and Decision Effectiveness. In finance, the collaboration is largely Symbiotic, with AI surfacing patterns or fraud risks and human experts interpreting edge cases and making final judgments; highlighted concerns include risk mitigation, trust, security, transparency, and Error Reduction Rate. In education, HAIC blends Human-Centric and Symbiotic modes, with Teaching Efficiency, Clarity of Communication, Ease of Use, Task Completion Time, and Learning Curve emphasized; the Lumilo example is used to illustrate real-time teacher-AI complementarity.
Requirements engineering offers one of the clearest industry-adoption snapshots. A survey of 55 practitioners across elicitation, analysis, specification, and validation reports that HAIC accounts for 54.4% of all RE techniques, whereas full AI automation is 5.4% and passive AI validation is 4.4–6.2% (Rani et al., 3 Nov 2025). By phase, HAIC is reported at 49.2% in Elicitation, 60.5% in Analysis, 54.1% in Specification, and 53.6% in Validation. The interpretation is that practitioners value AI’s active support over passive oversight and prefer AI as a collaborator guided by human expertise rather than as a replacement.
Software engineering literature extends HAIC from task assistance to team cognition and organizational behavior. One PhD-positioning paper argues that behavioral software engineering should explicitly include the human-AI team as a unit of analysis, focusing on decision-making at individual, team, and organizational levels (Rani, 13 Apr 2025). A later qualitative study with 10 practitioners argues that software developers view AI models as intellectual teammates rather than social partners, and that the apparent socio-emotional gap is better understood as a functional gap in collaborative capabilities (Rani et al., 27 Jan 2026). It introduces functional equivalents—internal cognition, contextual intelligence, adaptive learning, and collaborative intelligence—as technical capabilities intended to produce collaborative outcomes comparable to human socio-emotional traits without reproducing human emotion.
At the organizational-strategy level, HAIC use cases have been proposed as devices for prospective sensemaking. In a case study of a large automotive manufacturer, 63 potential HAIC use cases identified through 14 executive interviews were mapped onto Porter’s value chain and McGrath’s group task circumplex, producing a structured view of where AI may support work and what kinds of group tasks it is currently most suited for (Sudeeptha et al., 2024). The study finds concentration in support activities such as firm infrastructure, human resource management, and technology development, and a heavy skew toward choose and execute tasks rather than negotiate tasks.
5. Governance, limitations, and frontier forms
HAIC research repeatedly treats governance, safety, and human authority as constitutive rather than peripheral. In autonomous materials synthesis, a human-AI collaborative workflow combines RAG-grounded LLM hypothesis generation, Bayesian optimization, and an autonomous pulsed laser deposition laboratory, but the LLM is intentionally kept offline from instrument control because direct LLM control in high-stakes synthesis is treated as too risky (Haque et al., 14 Nov 2025). The authors distinguish their approach from conventional human-in-the-loop systems by locating collaboration across hypothesis generation, experimental planning, execution, interpretation, and policy revision between autonomous batches. A central lesson is that scalar optimization targets and analysis pipelines may need repeated human-AI revision during the campaign itself.
In spacecraft operations, SpaceHMchat proposes an open-source HAIC framework for all-in-loop health management of spacecraft power systems in the satellite mega-constellation era, organized around work condition recognition, anomaly detection, fault localization, and maintenance decision-making (Di et al., 19 Jan 2026). The system is explicitly designed as a copilot-style, human-controlled framework rather than full automation, motivated by high reliability requirements, non-transferable responsibility, and absolute human control authority in aerospace. The paper introduces the AUC principle—aligning underlying capabilities—which maps each subtask’s intrinsic nature to corresponding human and AI capabilities: logical reasoning to prompt-based reasoning, tool-dependent anomaly detection to function calling and MCP, learning-and-approximation fault localization to fine-tuning, and knowledge-intensive maintenance decisions to retrieval-augmented generation.
The literature is also candid about unresolved limitations. The broad evaluation framework for HAIC is theoretical and not yet validated in real deployments; behavioral factors are intentionally excluded as a primary category because they are difficult to standardize; ethical considerations, although emphasized in the literature review, are not fully operationalized in the framework itself; and the proposed weighted HAIC score is only conceptual (Fragiadakis et al., 2024). For LLM-based systems and generative AI in the arts, the same review argues that explanation quality, bias mitigation, evolving human trust, co-creativity, authorship, originality, inspiration, and process quality require further methodological development. This suggests that current HAIC frameworks are starting points rather than settled standards.
6. Acronymal reuse and unrelated meanings
Although HAIC most commonly refers to Human-AI Collaboration in the cited literature, the acronym is also reused for unrelated technical artifacts. In multimodal video understanding, “HAIC: Improving Human Action Understanding and Generation with Better Captions for Multi-modal LLMs” introduces HAICTrain, a dataset of 126K video-caption pairs, and HAICBench, a benchmark with 500 manually annotated video-caption pairs and 1,400 multiple-choice QA pairs, centered on person-disambiguated, chronologically structured action captions (Wang et al., 28 Feb 2025). There, HAIC denotes a dataset-and-annotation framework for human action and interaction comprehension rather than collaboration between humans and AI.
In robotics, “HAIC: Humanoid Agile Object Interaction Control via Dynamics-Aware World Model” denotes a proprioception-centric framework for humanoid whole-body control under partial observability, especially for underactuated objects such as carts and skateboards (Li et al., 12 Feb 2026). Its key technical elements are a dynamics predictor for high-order object states, projection onto a canonical object point cloud, and asymmetric fine-tuning of a world model and student policy. In the wireless sensing literature, the phrase integrated human activity sensing and communications has also been used for an ISAC subfield focused on human activity recognition, with deployment taxonomies spanning monostatic, bistatic, and distributed systems (Li et al., 2022).
Accordingly, HAIC is best treated as a context-dependent acronym. In current arXiv literature, its dominant encyclopedic sense is Human-AI Collaboration: a structured, evaluable, and increasingly domain-specialized form of cooperative work in which humans and AI systems jointly pursue tasks, share information, and adapt to one another under explicit constraints of trust, safety, interpretability, and organizational fit (Fragiadakis et al., 2024).