What do you mean by human-AI collaboration: Prerequisite functions and the affordances needed to achieve it

Synthesis: Asks what is gained and lost when 'collaboration' is applied freely to human-AI interaction. Argues true collaboration requires symmetric/negotiated relationship, shared goals, low and shifting division of labor, interactive exchange, and mutual modeling. Introduces a 5-level diagnostic taxonomy: Transactional, Situational, Operational, Praxical, and Synergistic. Only Synergistic satisfies full collaborative conditions. Most current human-AI interaction is consultation, AI Governance, delegation, or instruction.

• Introduces a 5-level taxonomy of human-AI teaming; most current AI interaction is consultation/delegation, not true collaboration.

Key Findings

• The chapter reconstructs the requirements that a situation, an interaction, and a set of cognitive processes have historically had to meet before being called collaborative, returning to longstanding accounts of Collaborative Learning.
• True human-AI collaboration requires a partly symmetric and negotiated relationship, shared and negotiable goals, a low and shifting division of labor, interactive and synchronous exchange, and mutual modeling, grounding, and socially shared AI Regulation in Education.
• Reviewing process-sensitive empirical studies of writing and problem solving, the author shows that most current human-AI interaction is better described as consultation, governance, delegation, or instruction rather than collaboration.
• The chapter introduces a five-level diagnostic taxonomy of human-AI teaming — transactional, situational, operational, praxical, and synergistic — defined by the affordances an AI system exhibits, with only the highest level beginning to satisfy the conditions the tradition places on collaboration.
• The author argues that most of the prerequisite functions an AI must possess for collaboration are present-day engineering choices rather than capabilities to be awaited.

The Five-Level Taxonomy

The taxonomy runs from transactional interactions (discrete, affordance-poor exchanges) up through situational, operational, and praxical teaming to synergistic collaboration, where evidence of collaboration should be sought in the trajectory of the human-AI unit rather than in output quality or user satisfaction alone. The high bar is not intended as a stick with which to beat existing systems but as a target that lets the field name what AI systems already do well without pretending it is the collaboration they do not yet achieve.

What this means for practice

• Instructors. Check which affordance you are actually claiming before calling a tool a collaborator: the levels are coded from observable behavior ("does the system do X?"), so a one-shot lesson-plan chatbot is transactional while a system that adapts to your enacted practice is praxical — different interaction types with different evaluative criteria.
• Instructors. Escalate deliberately rather than by default, and warn against inflating the construct into a slogan: co-reasoning costs attention, time, and trust, and a randomized controlled trial of situational-level real-time tutor guidance raised students' topic mastery by four percentage points overall and by nine for the students of lower-rated tutors — many routine tasks rightly belong at the transactional or operational level.
• Learners. Reserve "collaboration" for interactions in which the AI can offer reasons of its own and revise only when the argument warrants; a system that asks why you acted but always defers has the surface of reason-eliciting dialogue without revision authority.
• Designers. Treat the prerequisite functions — shared and negotiable goals, mutual modeling, grounding, and shared AI Regulation in Education — as present-day engineering choices rather than capabilities to be awaited, and use the taxonomy as a specification of design targets for AI tutoring and Collaborative Learning environments.
• Researchers. Measure collaboration in the trajectory of the human–AI unit rather than in output quality or user satisfaction: a systematic review of 105 empirical studies of AI-assisted decision-making found the observed interactions dominated by simplistic accept-or-reject paradigms.

Limitations

• This is a conceptual chapter built from longstanding learning-sciences accounts of collaboration, process-sensitive empirical studies of writing and problem solving, and prior systematic reviews; it reports no new sample, intervention, or dataset of its own.
• The five-level taxonomy is a proposal advanced with colleagues and defined by observable affordances, and the chapter reports no inter-rater reliability, external validation, or empirical test of the level boundaries it argues over.
• The taxonomy is explicitly a map of affordances rather than a quality ranking — higher is not uniformly better — so it cannot be used to score systems on a single quality scale.
• Claims about what current AI can do rest on published studies of particular systems and on observable behavior, because architectural claims about proprietary models often cannot be verified.

Citation

Mutlu Cukurova (2026). What do you mean by human-AI collaboration: Prerequisite functions and the affordances needed to achieve it. Handbook of AI and the Future of Education (forthcoming).