Think First, ChatGPT Later: Guiding Human–AI Collaboration for Learning Gains in Independent Human Creativity

Synthesis: Wong and Qiu (2026) ask whether ChatGPT's boost to creative performance translates into durable learning, or only inflates assisted output. University students (N = 196) solved a product-improvement task either independently (human-only), using ChatGPT freely (general-AI), or with a guided "think first, ChatGPT later" approach (regulated-AI) in which they first generated their own ideas, then collaborated with ChatGPT to improve, develop, and evaluate them. On the assisted task the general-AI group produced more creative solutions — but this advantage vanished on a later, harder invention task completed without ChatGPT, where their creativity fell back to human-only levels. In striking contrast, the regulated-AI group, despite no immediate gains, outperformed both other groups in independent creativity afterward. Process analyses showed the regulated-AI group used far more collaborative prompts aimed at improving their own initial ideas, and this greater use mediated their later advantage in independent originality.

Performance versus learning

The study foregrounds the distinction between performance (transient, assisted behavior) and learning (durable change that persists when assistance is removed). The authors argue this distinction is often neglected in generative AI research, which tends to measure students' performance on the assisted task rather than their later unassisted transfer. Freely using ChatGPT may raise immediate output while students merely appropriate its responses — outsourcing the thinking rather than developing it — a dynamic consistent with the Over-Reliance and Cognitive Offloading literatures.

The "think first, ChatGPT later" approach

To move students from editing AI output toward co-creating with it, the intervention structured the task into three phases:

1. Think first (Phase 1): students independently brainstorm and generate their own initial ideas.
2. Collaborate (Phase 2): students work with ChatGPT as a brainstorming and Feedback partner to improve, develop, and evaluate their own ideas, guided by sample prompts that scaffold metacognitive and creative processes (seeking refinements, elaborating, connecting ideas, identifying weaknesses, comparing, and refining).
3. Select (Phase 3): students independently refine and submit their single best solution.

This design operationalizes co-creation — the human sets the direction and contributes original ideas while AI amplifies them — rather than either dictating to or merely editing the machine.

Findings

A 3 × 2 mixed ANOVA revealed a striking crossover interaction:

• On the assisted product-improvement task, the general-AI group produced more original and useful solutions than both the human-only and regulated-AI groups. The regulated-AI group showed no immediate advantage — consistent with desirable difficulties, where effortful process can temporarily suppress performance.
• On the later unassisted product-invention task, the regulated-AI group outperformed both other groups in originality and usefulness. The general-AI group's creativity declined to human-only levels — its benefit was transient performance, not learning.
• Elaboration did not confound results: the regulated-AI group's solutions were less elaborate during the intervention, yet more creative afterward.

Process analysis of ChatGPT transcripts explained the mechanism. The regulated-AI group used collaborative prompts for 88.6% of its interactions (mostly "improve own initial ideas" and "develop selected ideas"), whereas 70.9% of the general-AI group's prompts were non-collaborative — most (59.6%) simply asking ChatGPT to generate ideas outright. Greater use of "improve own initial ideas" prompts was the only prompt type significantly correlated with later independent originality, and mediation analysis confirmed it accounted for the regulated-AI group's advantage.

Why free use fails: outsourcing divergent thinking

The general-AI group's pattern reflects a default of "metacognitive laziness": rather than engaging the divergent thinking that underpins original idea generation, students outsourced it wholesale to ChatGPT, investing less mental effort. Generating one's own ideas first forces the divergent phase of creative Problem Solving, and then using ChatGPT to improve those ideas exercises self-regulated refinement and evaluation. Notably, simply working alone (the human-only group) did not confer the same benefit — the human-only group declined on the harder invention task — so collaboration with ChatGPT added learning value beyond independent effort, consistent with redefinition effects of AI.

What this means for practice

• Instructors. Sequence AI use as "think first, ChatGPT later": require a phase of independent idea generation before students open ChatGPT, then a collaboration phase, then independent selection. The regulated-AI group prompted collaboratively in 88.6% of interactions and outperformed both other groups on the later unassisted task.
• Instructors. Supply sample prompts that position ChatGPT as a brainstorming and feedback partner — improve, develop, evaluate my ideas — rather than an answer provider: 59.6% of the general-AI group's prompts simply asked it to generate ideas outright, and only "improve own initial ideas" prompts correlated with later independent originality.
• Instructors. Do not judge the intervention by assisted-task output. The regulated-AI group showed no immediate advantage and produced less elaborate solutions during the intervention, yet was more original and useful afterward — the pattern expected from desirable difficulties, where effortful process temporarily suppresses performance.
• Designers. Design assessments that reach the unassisted moment. The general-AI group's advantage on the assisted task vanished on the later invention task, falling back to human-only levels, so a grade taken while ChatGPT is open measures performance, not the creativity students retain.
• Instructors. Add metacognitive feedback. Students often overestimate how much they "came up with" versus took from ChatGPT, so originality scoring or short reflection prompts can correct the illusion of authorship that sustains cognitive offloading.

Relationship to existing research

The findings sit alongside work on Over-Reliance and AI-assisted learning that shows free AI use can erode durable gains, and extend the Human AI Collaboration and co-creation literature with an experimental demonstration that guided collaboration beats both solo and free-AI conditions. It also provides direct experimental evidence for the "performance–learning" gap in generative AI education, and connects to AI literacy frameworks (e.g., ED-AI Lit's collaboration component) as a candidate practice for integration.

Limitations

• One brief lab session with 196 university students, run in groups of up to five: a 12-minute product-improvement intervention with a 3-minute independent-generation phase tests no semester-long course and no real classroom.
• Transfer was measured on a single unassisted product-invention task completed immediately after the intervention, so durability beyond the session — retention, or application to course work — is untested.
• The regulated-AI condition was not compared with a "regulated-human" group working the same sample prompts alone, so the design cannot fully separate the guided sequence from the specific contribution of collaborating with ChatGPT.
• The sample prompts were presented for illustration only and were not engineered to elicit ChatGPT's best output; the authors flag better prompt engineering and additional self-regulation prompts as untested refinements.

Citation

Wong, S. S. H., & Qiu, S. X. (2026). Think First, ChatGPT Later: Guiding Human–AI Collaboration for Learning Gains in Independent Human Creativity. Educational Psychology Review, 38(45).