A platform-built logo-AI reshapes demand: EPWK logo-design posts request a broader mix of skills after the tool’s launch, and heightened competition among freelancers explains part of the change.

Main Finding

The platform-embedded logo-generative AI (Xiaowei LOGO) introduced on EPWK (Oct 2018) caused a statistically significant increase in the diversity of skills requested in logo-design job posts relative to other design categories. The effect is large and robust (DID TWFE estimates: +0.224 centroid dispersion; +0.286 pairwise dispersion) and is partly mediated by intensified competition among freelancers.

Key Points

• Research question: How does platform-embedded generative AI reshape task-level skill diversity in online labour markets, and through what mechanisms?
• Context: EPWK (Chinese design-focused OLM); Xiaowei LOGO — an employer-facing logo generator launched Oct 2018. Dataset covers Oct 2017–Oct 2019.
• Main empirical result: Logo jobs show higher semantic dispersion of requested skills after the embedded logo-AI rollout compared with other design jobs (difference-in-differences with client and time fixed effects).
• Mechanism: Part of the increase in requested skill diversity appears to operate via increased competition among freelancers, which correlates with employers requesting broader skill bundles.
• Contribution: Distinguishes effects of platform-embedded GenAI (task-specific, bounded shock) from general-purpose GenAI and shifts attention from aggregate outcomes (demand, price) to compositional changes in required skills.

Data & Methods

• Data: ≈9,184 design-related task postings from EPWK, one year pre- and post-launch (Oct 2017–Oct 2019). Task metadata includes title, description, views, and final price.
• Treatment/control: Treated = client-months containing only logo-design tasks; Control = client-months with other design categories (visual identity, poster, slide, font, etc.). Post indicator = months after Oct 2018.
• Skill extraction:
  • Built a 117-label skill vocabulary (58 professional, 34 tool-related, 25 soft skills) using GPT-4.1 plus manual verification.
  • Prompted GPT-4.1 to extract up to six required skills per task (pilot-tested and refined on 100 tasks).
• Skill embedding & diversity measures:
  • Each skill label embedded with OpenAI text-embedding-3-small (1,536-d vectors).
  • Two dispersion-based skill diversity metrics computed per task: centroid breadth (mean cosine distance from task centroid) and pairwise breadth (mean pairwise cosine distances). Both normalized between 0 and 1.
• Estimation:
  • Difference-in-differences using two-way fixed effects (client cluster and month fixed effects). Outcome aggregated at client-month level (average skill diversity of that client’s postings in month).
  • Standardisation of variables; parallel-trends test conducted and pre-treatment coefficients were generally indistinguishable from zero.
• Key numeric results:
  • DID coefficient Logo-AI × Post: centroid = 0.224 (SE 0.041) ; pairwise = 0.286 (SE 0.040) .
  • Observations / clusters reported: 1,112 (R² reported: 0.065 for centroid, 0.190 for pairwise).
• Robustness / validity:
  • Parallel-trends visually/tests support DID assumptions.
  • Pilot and human verification used in prompt design to limit hallucination and noise in skill extraction.

Implications for AI Economics

• Microcomposition matters: Embedded GenAI can change not only aggregate demand/price but the composition and breadth of human skills employers request. Economic models of AI impact should incorporate skill-bundle reconfiguration and semantic dissimilarity of required capabilities.
• Platform design choice matters: Platform-embedded, task-specific GenAI (versus broad general-purpose releases) can produce localized, measurable reallocation effects within categories — suggesting platforms can steer labor-market skill demand by embedding tailored automation tools.
• Labor market sorting and competition: The finding that intensified competition partially mediates higher requested skill diversity implies embedded GenAI can alter competitive dynamics (e.g., push employers to ask for broader/higher-differentiated bundles), with potential welfare implications for freelancers (reskilling needs, reallocation to complex tasks).
• Policy & managerial implications:
  • Platforms should consider complementary supports (training, certification, matchmaking changes) when embedding GenAI to mitigate adverse transitions and help freelancers adapt to new skill bundles.
  • Economists and policymakers assessing AI disruption should track task-level skill composition (not only employment counts/prices) to understand revaluation of human capabilities.
• Methodological implication: LLM-based extraction + semantic embeddings offers a scalable way to measure skill diversity from text, useful for future empirical work on task composition and AI impacts.

Limitations & next steps (as noted by authors): single platform and domain (logo/design), early-stage (2018) embedded tool may differ from post-2022 general-purpose diffusion; planned extensions include inductive Computationally Intensive Theory Construction (CITC) combining LLMs, topic modeling, human coding, and social media to build a dynamic process model of how requested skill diversity evolves.