Student-AI Collaboration Preferences

• Student-AI Collaboration Preferences are defined by students’ desired roles, interaction modes, and automation levels in educational AI engagements.
• Key insights reveal that preferences vary with role specification, agency distribution, and task demands, guiding adaptive AI interface designs.
• Empirical and modeling studies show that cognitive, affective, and design factors inform student choices, emphasizing the need for transparent, customizable AI support.

Student-AI Collaboration Preferences encompass students’ expressed, revealed, and latent desires regarding the roles, behaviors, and affordances of artificial intelligence systems within educational contexts. This domain is defined by rigorous empirical evaluations and conceptual frameworks that address not only preferred modes of interaction and automation but also the cognitive, affective, and sociotechnical dimensions of human–AI teaming in learning environments. Preferences are shaped by task characteristics, learner dispositions, prompt engineering skill, and instructional design, with implications extending from one-shot feedback interactions to longitudinal, adaptive collaboration.

1. Dimensions of Student-AI Collaboration Preferences

Student preferences in AI-augmented learning are inherently multidimensional. Key dimensions include:

• Role Specification: Students distinguish and select among AI as tutor (didactic, explanation-focused), tool (procedural support), collaborator (strategy co-construction), or peer (motivational, reflective engagement) (Zhu et al., 8 Oct 2025, Muzumdar et al., 29 Nov 2025).
• Agency Distribution: Collaboration is stratified by perceived leadership—human-led, even contribution, AI-led—with most students favoring balanced or human-led approaches to safeguard sense of agency (Zhu et al., 2024).
• Level of Automation: Preferences span a spectrum from manual (no-AI), assistive (on-demand hinting), semi-autonomous (co-solution generation), to fully autonomous (AI-complete) workflows. Desired and observed automation levels show consistent, task-dependent gaps (Dan, 13 Jan 2026).
• Interaction Mode: Preferences include directive (stimulus–response), assistive (scaffolded explanation), dialogic (co-construction), and empathetic (personalized, affective support), often aligning with classical learning theories (Muzumdar et al., 29 Nov 2025).
• Scaffolded vs. Unconstrained Use: Many students appreciate guardrails (conceptual hints, partial support); others prefer “see solution” shortcuts, with usage patterns correlated to time pressure and academic standing (Kapoor et al., 15 Apr 2025).
• Temporal Evolution: Preferences are not static; they can shift with experience, exposure to different AI affordances, and integration with ongoing coursework (Lyu et al., 12 May 2025, Mehri et al., 6 Jan 2026).

2. Empirical Characterization of Student Preferences

Large-scale observational and experimental studies have produced several core findings:

Study Context	Most Preferred AI Mode	Notable Preference Patterns
Intro Physics Feedback (Sirnoorkar et al., 13 Aug 2025)	Structured prompt (PE + feedback theory)	61.4% chose feedback with explicit gap-closing; self-crafted lowest
CS Group Tutoring (Yang et al., 2024)	AI-enabled small group, synchronous editing	78% preferred with AI; demand for just-in-time, transparent hints
Pair Programming (Lyu et al., 12 May 2025)	Human+AI mix (PAI), not AI-only or solo AI	Attitudes ↑ w/ GenAI, but human ideation prized over pure AI help
Graduate CS Coursework (Dan, 13 Jan 2026)	High-competence automation, with control	Desired automation outpaces usage; transparency and control wanted
Math Modelling (Zhu et al., 8 Oct 2025)	Competent role (Tutor/Excellent/TA/Peer)	Weakest preference for “struggling” AI, DT learners like open TA

These preferences often depend on both immediate task demands and overarching learner traits (self-efficacy, design/cognitive/algorithmic thinking), with studies highlighting pronounced individual variation and the need for adaptive systems.

3. Behavioral Patterns and Interaction Taxonomies

Quantitative and qualitative analyses across multiple domains reveal stable taxonomies of student–AI engagement:

• Collaboration Type:

1. Human Leads: Student directs, AI supports
2. Even Contribution: Co-building
3. AI Leads: Student delegates, AI executes (Zhu et al., 2024)

• Interaction Mode:

1. Active Questioners: Frequent, substantive queries
2. Responsive Navigators: Progress-checks, off-task navigation
3. Silent Listeners: Passive content consumption (Hao et al., 3 Mar 2025)

• AI Scaffolding Preference:

1. Challengers: Minimal guidance, maximal challenge
2. Explorers: Stepwise, animation-driven support
3. Emerging Strategists: Unguided, reset-heavy trial (Vanacore et al., 3 Nov 2025)

• Prompt-Driven Feedback:
  • Engineered prompts that encode both formal structure and effective feedback principles are substantially favored for perceived usefulness, clarity, and actionable next steps (Sirnoorkar et al., 13 Aug 2025).
• Task-Critical Timing:
  • Students invoke AI at entry (“oracle” use), after own attempts (“debugging/clarification”), or rarely for verification (Amoozadeh et al., 2024). Those who balance independent effort and selective AI consulting report greater gains in self-efficacy.

4. Underlying Mechanisms: Cognitive, Affective, and Social Influences

Underlying student preferences are both affective and cognitive factors:

• Agency and Control:

High sense of negative agency (feeling like an “instrument” of AI) significantly predicts lower collaborative problem-solving outcomes (Zhu et al., 2024).

• Trust, Usefulness, and Genuineness:

Students trust and value AI-generated or co-produced feedback, but strictly AI-only outputs suffer reputational declines if revealed, especially in terms of genuineness and credibility (Zhang et al., 15 Apr 2025).

• Inequality Aversion:

Students exhibit preference for collaborative agents that enable meaningful human contribution and avoid large performance imbalances—a formalized “inequality aversion” effect (Mayer et al., 28 Feb 2025).

• Customization and Transparency:

Preferences are shaped by the ability to tune AI initiative, explanation granularity, and automation windows; students demand transparent reasoning, source traceability, and confidence communication (Dan, 13 Jan 2026).

• Skill- and Disposition-Driven Value:

More advanced prompt engineers and those with higher design/cognitive/algorithmic thinking exhibit nuanced, utility-maximizing AI use, particularly appreciating question-driven or collaborative scaffolding (Zhu et al., 8 Oct 2025, Hou et al., 2024).

5. Modeling and Optimization of Student Preferences

A principled approach to optimizing student–AI collaboration preferences combines multi-stage preference modeling, feedback loops, and explicit adaptation:

Model Phases and Key Methods

Modeling Phase	Representative Technique/Formula	Limitations
Pre-Interaction	Probabilistic persona modeling	Coarse-grained, inflexible to mid-task shifts
Mid-Interaction	Real-time adaptation via in-dialogue learning, SNCF	Compute-intensive, risk of misreading implicit cues
Post-Interaction	DPO/RLHF for reward-aligned fine-tuning	Relies on explicit/comparative user feedback, overfitting

Mathematical frameworks explicitly formalize learning and adaptation at each stage:

• Preference vector update: $$\theta_{t+1} = \theta_t - \alpha \nabla_\theta \ell_{\text{pref}}(\theta; \text{history}_t)$$ (Afzoon et al., 28 May 2025)
• Direct Preference Optimization (DPO) loss for policy fine-tuning: $$\mathcal{L}(\theta) = -\log \sigma[\beta(\log p_\theta(y_w|x) - \log p_\theta(y_l|x))]$$ (Afzoon et al., 28 May 2025)

Long-term retention of interaction preferences is further enhanced by memory-augmented architectures that embed per-user preference states, updated via reflection and aligned with task + memory objectives (Mehri et al., 6 Jan 2026).

6. Design Guidelines, Pedagogical Implications, and Future Directions

Empirical findings and modeling results combine to inform actionable design recommendations:

• AI Literacy and Prompt Engineering Training:

Instruction in foundational prompt engineering and effective feedback theory demonstrably enhances student ability to solicit useful, actionable AI support (Sirnoorkar et al., 13 Aug 2025, Hou et al., 2024).

• Dynamic, Profile-Driven Role Assignment:

Systems should switch among pre-defined AI personas/styles (Tutor, Peer, TA) according to detected learner profiles and in-situ performance indicators, with explicit notifications when role changes occur (Zhu et al., 8 Oct 2025).

• Preference-Centric AI Interfaces:

Controls for automation levels, explanation transparency, and reasoning traceability must be surfaced and readily adjusted by students (Dan, 13 Jan 2026, Muzumdar et al., 29 Nov 2025).

• Co-Production and Instructor-in-the-Loop Modes:

Hybrid workflows—where educators prompt, refine, or approve AI output—preserve trust and perceived value even when AI participation is fully disclosed (Zhang et al., 15 Apr 2025).

• Proactive, Adaptive Scaffolding:

Systems should detect off-task, overreliant, or performance-degraded usage and intervene dynamically with reflective or regulatory prompts (Kapoor et al., 15 Apr 2025).

• Collaborative Alignment and Mutual Contribution:

Agents designed to recognize, defer to, or explicitly blend with human initiative (via role partition, intentionality awareness, and contribution balancing) enhance both subjective and objective outcomes (Mayer et al., 28 Feb 2025).

Open Challenges

• Ensuring cognitive alignment and collaborative negotiation beyond “instruct–serve–repeat” patterns requires new research into proactive, socially aware, and epistemically transparent AI partner models (Saqr et al., 3 Aug 2025).
• Preference modeling must balance short-term satisfaction with long-term learning, dynamically mediate automation boundaries, and address the risk of negative agency or excessive delegation.
• Robust evaluation must combine performance, engagement, trust, and agency metrics, with longitudinal studies linking preference-sensitive interventions to durable learning gains (Mehri et al., 6 Jan 2026, Afzoon et al., 28 May 2025).

Future progress will rely on developing memory-equipped, role-adaptive, and preference-optimized student–AI collaboration systems, grounded in principled measurement and informed by empirically robust models of student agency, trust, and personalization.