Structured, 'logical' engagement with ChatGPT raises both originality and usefulness of short marketing copy, while cognitive prompts boost novelty and informational prompts improve usefulness; mid-educated users benefit most from AI assistance, with lower-educated users showing the greatest potential uplift.

Main Finding

Logical, structured human–AI interactions boost both novelty and usefulness of creative outputs; cognitive engagement with AI increases novelty only, while informative (high-volume, relevant input) interaction increases usefulness only. Creatively, medium-skill participants produced the best outcomes overall, and low-skill participants showed the greatest potential gains from higher-quality human–AI interactions.

Key Points

• Interaction dimensions: the authors define three measurable interaction styles with AI:
  • Cognitive — depth of thought/mental modeling in the dialogue.
  • Logical — structured, well-organized, rule-based exchanges.
  • Informative — volume and relevance of factual/contextual information supplied by the human.
• Creativity measured along two standard dimensions:
  • Novelty (originality/distinctiveness).
  • Usefulness (practical value/relevance).
• Empirical associations:
  • Logical interactions → positive association with both novelty and usefulness.
  • Cognitive interactions → positive association with novelty only.
  • Informative interactions → positive association with usefulness only.
• Skill heterogeneity:
  • Participants grouped by education (ISCED proxy) into low-, medium-, and high-skill.
  • Medium-skill participants achieved highest creative performance on average.
  • Low-skill participants showed the most room for improvement when they engaged in higher-quality AI interactions.
• Study framing: propositions linking AI benefits to individual skill differences and to the quality/nature of human–AI interaction; interaction effects are hypothesized to be moderated by skill level.

Data & Methods

• Design: Online quasi-experimental creative task using ChatGPT (OpenAI GPT-4o). Writers were free to use the model; all human–AI dialogues were recorded.
• Primary creative sample (writers): 90 UK-based participants recruited via Prolific, stratified into three educational skill groups (30 low-, 30 medium-, 30 high-skill). Each produced two short social-media posts (180 posts total).
  • Demographics (aggregate): mean age ≈ 42.7; 62.2% female (group-specific breakdowns reported).
• Outcome measurement:
  • External raters: independent sample of 281 Prolific workers provided blind ratings (novelty and usefulness) on 100-point scales; overall 554 valid ratings; each post rated ≈3 times.
  • Reliability: novelty α = 0.92; usefulness α = 0.90.
• Human–AI interaction coding:
  • Two trained research assistants manually coded conversational transcripts for cognitive, logical, and informative dimensions using a 1–7 scale and a structured coding manual.
  • Interrater agreement: cognitive 0.71; informative 0.76; logical 0.78.
• Controls included age, gender, prior AI experience, and job creativity intensity.
• Notes on reporting: the manuscript is an unedited “in press” version and contains some inconsistent participant-count references (e.g., an early reference to 245 participants vs. the 90 writers reported in methods); primary analyses center on the 90 writers / 180 posts dataset.
• Limitations acknowledged by authors (and implied by design):
  • Quasi-experimental (not randomized) — causality limited.
  • Skill proxied by education (ISCED); may miss domain-specific expertise.
  • Single task domain (marketing/social-media posts) and a single AI model (GPT-4o) — limits generalizability.
  • Manual coding of interaction dimensions introduces subjectivity despite acceptable interrater agreement.

Implications for AI Economics

• Skill-biased technological change:
  • Findings nuance the standard “skill-biased” framing: AI does not uniformly favor high-skilled workers. Medium-skilled users gained most in this creative task, and low-skilled users can make notable gains if interaction quality improves. Models of labor demand should account for interaction quality and task-specific complementarities, not just skill level.
• Human capital and complementarities:
  • Returns to AI depend on human ability to structure inputs and cognitively engage with AI. Investments in training that develop logical structuring and effective information provision (prompt-engineering skills) could raise productivity and creative output across skill groups, especially for low- and medium-skill workers.
• Product design and firm strategy:
  • AI product features that promote logical structuring (templates, scaffolds, step-by-step prompts) may simultaneously boost novelty and usefulness, improving adoption value for a wider user base.
  • Interfaces that encourage informative inputs (context forms, data upload, guided prompts) can increase usefulness; features that stimulate cognitive engagement (explainable outputs, justification requests) can enhance novelty.
  • Firms can segment offerings: lightweight informative features for users seeking usefulness; scaffolding and cognitive prompts for innovation-focused tasks.
• Inequality and policy:
  • Policymakers should consider that AI’s effect on wage/productivity inequality will depend on whether workers can access training/support to engage effectively with AI. Subsidized upskilling in “interaction” skills (prompt design, AI literacy) could reduce unequal gains.
• Measurement and evaluation in economics research:
  • Studies of AI’s economic effects should measure not only access/use but interaction quality (structure, information content, cognitive engagement). Linguistic/dialogue-based metrics (automated or coded) could be integrated into field data and firm-level surveys.
• Future empirical work recommended for economists:
  • Randomized controlled trials varying interface features (logical scaffolds, information prompts) to estimate causal effects on productivity/creativity and heterogeneous impacts by skill.
  • Cross-sectoral studies to assess whether interaction-style effects generalize beyond creative marketing tasks (e.g., programming, scientific writing, policy drafting).
  • Cost–benefit analyses on training vs. AI customization: quantify returns to training workers in interaction skills compared with investing in model features that reduce user-side cognitive/structuring burdens.

If useful, I can extract specific recommended interventions (training curricula, interface design elements, or experimental designs) tailored for economists, firms, or policymakers based on these results.