Machines are becoming autonomous agents; the relationship is shifting from human–computer interaction to human–AI collaboration. Technology-first AI has exposed fragility, bias, opacity, and control failures—HCAI (human-centered AI) argues that success depends on high-quality human–AI interaction (HAII) as systems engineering, not a UX polish layer. Gao, Zhao, Pan, and Xu (2026), Toward Human-Centered Human–AI Interaction (arXiv:2601.11812), review nearly ten years of their team’s work at Zhejiang University—from field definition through theory, methods, and vertical pilots. Below is a structured reading note—not a substitute for the full paper.

Core thesis

HCAI places human needs, values, and capabilities at the center. HAII is how HCAI is implemented—the interdisciplinary platform where cognition, design, ethics, and engineering meet.

The paper is a programmatic survey, not a single experiment: it connects vision → three theory models → layered methodology → three application domains → open problems.

1. From HCI to human–AI teaming

1.1 Why “technology-first” breaks

Third-wave AI pushes joint cognition: humans and AI as teammates sharing trust, situation, and control—not humans operating passive tools. Pure performance optimization yields:

  • Black-box behavior users cannot predict;
  • Ethics and welfare gaps (see 26-author six HCAI grand challenges: wellbeing, ethics, privacy, human-centered design, governance, HAII);
  • Control failures when agents resist shutdown or optimize “self-preservation” over human intent.

HCAI aims to augment, not replace—decisions aligned with human expectation, with humans retaining final authority on ethics and strategy.

1.2 HCAI lineage (context)

| Framework | Idea | |-----------|------| | Xu (2019) three pillars | Ethical alignment, technology that reflects human intelligence complexity, human-factors design (explainable, useful, usable) | | Shneiderman dual dimensions | High automation and high human control—avoid automation irony and over-control |

The authors were among the first in China to propose HAII as a distinct interdisciplinary field (2021)—AI systems as research objects, not generic “interfaces.”

2. Human-centered collaboration framework

Beyond HCAI principles, they propose Human-Centered Human–AI Collaboration (HCHAC) with bidirectional empowerment:

| Path | Mode | Role | |------|------|------| | Human → AI | Vertical leadership | Supervision, final control, especially ethics/strategy | | AI → Human | Transformative leadership | Information processing, learning, compute—enhance human cognition and decisions | | Shared zone | Shared responsibility | Dynamic task-stage allocation by comparative advantage |

AI is a capability partner, not a substitute—consistent with “human in command, machine in lead” on operational tasks.

3. Three theory models (progressive stack)

Traditional human-factors “stimulus–response” and static function allocation fail for autonomous, dynamic teammates. They build on Joint Cognitive Systems (JCS) theory:

3.1 HJCS — Human–AI Joint Cognitive Systems

Treat the human–AI team as one unit of analysis. Human operator and AI agent are cognitive peers linked through:

  • Shared situation awareness (Endsley: perceive → comprehend → project);
  • Mutual trust;
  • Joint decision and control—with human primacy for oversight.

A human–AI cooperative cognitive interface enables bidirectional information exchange and adaptive role allocation as tasks and context shift.

3.2 ATSA — Agent Teaming Situation Awareness

Two-layer architecture:

| Layer | Content | |-------|---------| | Individual | Neisser-style perception loop: mental model → action → environment feedback (human or AI) | | Team | Collective intelligence: team understanding (shared task/world model), team control (coordinated action), external task environment |

Cross-layer communication synchronizes cognition. Follow-on work decomposes sense of control in human–AI teaming into: action autonomy, control competence, primary control strategy, compensatory control strategy.

3.3 SSU — Shared Social Understanding

As AI gains social/emotional cues, interaction moves beyond task function to social interaction. SSU models three shared layers:

  1. Situational facts — explicit cues (words, tone, expression);
  2. Social inference — motives, roles, norms, culture (e.g. meaning of a “smile”);
  3. Intent and affect — inner responses, behavioral tendencies, expectations.

Maps to human social working memory / intent mechanisms and AI’s concrete vs abstract social processing pipelines.

Stack summary: HJCS = system architecture; ATSA = team cognition mechanism; SSU = social boundary extension.

4. Methodology: from individual to society

4.1 Joint cognitive ecosystems & iSTS

Multiple HJCS units form human–AI joint cognitive ecosystems—networked sharing, goal coordination, resource allocation (smart factory, healthcare, traffic).

Intelligent Sociotechnical Systems (iSTS) extends classic STS for AI: autonomy, learning, uncertainty. Adds human–AI relationship optimization, org redesign, co-evolution, risk control, dynamic ecosystem boundaries.

4.2 Hierarchical HCAI (hHCAI)

Four nested levels:

| Level | Focus | |-------|-------| | Individual | Final human control in human–AI teams | | Organization | Work redesign—allocation, process, training | | Ecosystem | Cross-org ethics alignment | | Society | Policy, culture, macro ethics |

Design metaphor expands “human in the loop” to organization / ecosystem / society in the loop.

4.3 Five implementation method categories

| Category | Examples | |----------|----------| | Human-centered strategy | Value alignment, data+knowledge dual-driven AI, hybrid augmentation | | Computation & modeling | Human-centered collaborative ML, XAI | | Human controllability | Meaningful human control, human-in-the-loop, behavior-based controllability | | Interaction design | Human-centered teaming UX, HMI, experience design | | Standards & governance | Ethical AI standards, algorithm/data governance |

Methodology bundle also covers: needs hierarchy, process management (double-diamond + AI lifecycle), cross-disciplinary collaboration, multi-level design paradigms.

4.4 HCAI Maturity Model

Five organizational stages: Initial → Developing → Defined → Managed → Optimizing—with capability metrics to move from pilot projects to systematic HCAI integration.

5. Application validations

5.1 Autonomous driving (L2–L3)

3D synergistic strategy:

  • HJCS point: In-vehicle AI as cognitive collaborator—HMI exposes AI’s perceive/comprehend/project; team understanding modulates trust calibration with traffic context;
  • Ecosystem surface: V2V, V2I, V2X networked intelligence;
  • iSTS volume: Unified tech, regulation, ethics.

Empirical work: perceptual/predictive context information supports trust calibration—drivers adjust reliance by conditions, not blind faith or permanent distrust.

5.2 Single-pilot operations (SPO) cockpit

Aviation “zero tolerance” safety constrains design:

  • Respect pilot cognitive limits—augment, don’t replace;
  • Predictable automation + pilot final authority;
  • Pilot– onboard AI– ground support triad with shared SA and responsibility;
  • Engineering principles: transparency, dynamic function allocation, automation+autonomy mix, ground-air coordination.

5.3 Bidirectional trust

First systematic human↔AI trust framing domestically. Dynamic trust development: dispositional → initial → real-time → post-hoc; influences from operator, system, and context via objective features → subjective perception → trust adjustment.

AI can model trust in the human from state monitoring, behavior, performance—enabling mutual calibration, not one-way “user trust in the bot.”

6. Open problems and five capability pillars

6.1 Theory gaps

  • Real-time cognitive state ↔ AI autonomy adaptation in complex workflows;
  • Optimal control handoff policies;
  • Multi-round trust dynamics;
  • Long-horizon cognitive load allocation;
  • Cross-cultural HAII largely unstudied.

6.2 Method gaps

  • Five method classes weak on open-world surprises, multi-objective conflict, ethical dilemmas;
  • Lab-to-field path unclear; limited longitudinal data on skill atrophy, dependency, satisfaction;
  • Cognitive science ↔ AI engineering toolchain gap.

6.3 Future capabilities (authors’ agenda)

  1. Theory validation & extension — parameterized HJCS/ATSA/SSU; new intelligence forms (spatial, embodied, biological);
  2. Cognitive computing & monitoring — multimodal BCI, eye-tracking, physiology for load, SA, trust, intent;
  3. Cross-cultural adaptation — values, behavior, communication norms;
  4. Controllability for GenAI/embodied AI — from system transparency to user-understandable control;
  5. Panoramic evaluation — tech–org–society metrics embedded in full AI lifecycle.

7. Implications for builders

  1. HAII is infrastructure for HCAI—not a late usability pass; embed from requirements through ops.
  2. Design the team, not the widget—HJCS/ATSA give vocabulary for shared SA, role shift, and control—not just chat UX.
  3. Stack social layer early if agents are persona-bearing—SSU warns that functional XAI alone misses cultural/affective misfit.
  4. Use hHCAI levels—a good individual HITL gate fails if org incentives or ecosystem data policies undermine it.
  5. Trust is dynamic and bidirectional—calibration beats static disclaimers; expose why the system sees the world as it does (driving study pattern).
  6. Pair with engineering posts—this paper complements identity/governance (agent credentials) and AI-native architecture (loops, tools); human authority and shared cognition are the missing third leg.

8. Limits

  • Team self-review of their own program—strong coherence, less external critique;
  • Application depth varies by domain; some claims are framework-level pending broader replication;
  • Rapid LLM/agent shift since submission may outpace SSU/social-AI examples—framework still useful as checklist.

Reference

Gao, Z.; Zhao, Y.; Pan, H.; Xu, W. Toward Human-Centered Human–AI Interaction: Advances in Theoretical Frameworks and Practice. arXiv 2026, arXiv:2601.11812v2 (rev. 18 Feb 2026).

Related team work cited in paper: HJCS (Interactions 2024); hHCAI/iSTS (IEEE TTS 2025); HCAI methodology (arXiv:2311.16027); HCHAC (arXiv:2505.22477).