Most creative-sector tasks are headed toward hybrid human–AI workflows: a GPT‑4 semantic mapping of 593 Australian tasks finds only 2.7% clearly replaceable by AI, 11% effectively AI‑immune, and 86.3% best served by mixed human–AI configurations, with tacit knowledge limiting autonomy while codified structure plus novelty boosts AI feasibility.

Main Finding

The paper shows that creativity decomposed into three rule types — retention (codified), adoption (tacit), and origination (novel) — maps onto distinct patterns of human, AI, and hybrid task carrying. Using GPT-4 annotations of 593 tasks across 126 Australian occupations, the authors find (1) agent-rule specialization, (2) a "structured novelty effect" where AI autonomy is more feasible when codified cognitive rules co-occur with novel-rule creation, and (3) a "tacit knowledge boundary" where tacit rules reduce AI autonomy feasibility. Simulating combined efficacy (agent-rule matching) and efficiency (technical/effort gains) logics yields three task classes: mixed/hybrid (86.3%), limited replacement (2.7%), and AI-immune (11.0%).

Key Points

• Conceptual contribution: an evolutionary framing of creativity with three rule types:
  • Retention (codified rules) — explicit, formalizable knowledge.
  • Adoption (tacit rules) — embodied, context-dependent know-how.
  • Origination (novel rules) — genuine novelty/creation.
• Empirical approach: synthetic semantic annotation of task descriptions using GPT-4 to infer cognitive and behavioural rule content and likely task carriers (human, AI, hybrid).
• Two empirical mechanisms:
  • Structured novelty effect: AI autonomy feasibility is higher when tasks combine defined/codified cognitive rules with a requirement for novel rule creation.
  • Tacit knowledge boundary: presence of tacit/adoption rules is negatively associated with AI autonomy feasibility.
• Reconciliation of logics:
  • Efficacy logic: matching agent strengths to rule types (human/tacit, AI/codified, hybrid/novel combinations).
  • Efficiency logic: optimizing for technical feasibility and potential gains.
• Classification outcome: mixed/hybrid tasks dominate (86.3%); very few tasks are clear candidates for outright replacement (2.7%); a minority are AI-immune (11.0%).
• Two levels of complementarity:
  • Within-task complementarity across rule types (supports different creativity modes: combinatorial, exploratory, transformational).
  • Performance complementarity concerning efficiency potential when tasks are done (who can do them faster/cheaper).

Data & Methods

• Data:
  • Task-level descriptions: 593 tasks drawn from the Australian Skills Classification spanning 126 occupations in cultural and creative industries.
• Annotation method:
  • GPT-4 was used to generate synthetic annotations of the semantic content of each task description.
  • From these annotations the authors derived indicators for:
    • Cognitive rules (types and mix: codified, tacit, novel).
    • Behavioural rules.
    • Likely carriers: human, AI, or hybrid human-AI configurations.
    • AI autonomy feasibility and efficiency potential (GenAI exposure scenarios).
• Analysis:
  • Agent-rule matching analyses to identify specialization patterns.
  • Regression/associational tests (implicit in findings) linking rule mixes to AI autonomy feasibility.
  • Simulated classification combining efficacy (agent-rule match) and efficiency (technical/effort gains) logics to categorize tasks into mixed, limited replacement, and AI-immune.
• Scope & caveats in methods:
  • Reliance on generative model annotations (GPT-4) to operationalize latent semantic features.
  • Single-country occupational taxonomy (Australia) and a focus on cultural/creative sector tasks.

Implications for AI Economics

• Labor substitution vs complementarity:
  • Most tasks are likely to become hybrid rather than fully automated; policy and firm responses should prioritize augmenting human-AI complementarities rather than defending against wholesale displacement.
• Occupational/skill reconfiguration:
  • Training and skill policies should emphasize:
    • Strengthening tacit, contextual, interpersonal, and embodied skills that form the tacit boundary.
    • Enabling workers to operate in hybrid configurations that combine codified rule execution with novel rule generation.
• Measurement and forecasting of automation risk:
  • Task-level assessments should go beyond binary technical feasibility and incorporate rule-type mixes (codified/tacit/novel) and efficiency incentives to better predict actual adoption and impact.
• Productivity and diffusion:
  • The structured novelty effect suggests GenAI may raise autonomy feasibility in creative tasks when sufficient codified structure exists, potentially accelerating productivity gains in domains that marry structure and generativity.
• Firm strategy and product design:
  • Firms should design workflows to exploit two complementarity levels: allocate codified, repeatable components to AI; reserve tacit-sensitive components for humans; and structure roles that exploit AI-generated novelty with human curation.
• Policy/regulation:
  • Support for transition (retraining, portability of skill certifications) should account for dominant hybrid outcomes.
  • Intellectual property, accountability, and quality-control frameworks need adaptation for tasks where AI generates novel rules but humans retain judgment and adoption.
• Research agenda:
  • Validate synthetic annotations with human-coded labels and expand to other sectors/countries.
  • Link task classifications to wage, employment, and firm-level adoption data to quantify distributional and macroeconomic effects.

Limitations to keep in mind: the use of GPT-4 for annotations can introduce model-driven biases; task descriptions are static and may not capture on-the-job dynamism; findings are based on Australian task taxonomy within cultural and creative industries and may not generalize fully to other sectors or economies.