Platform workers in Bangladesh quietly build and sustain AI systems, performing essential but invisible work under precarious conditions; they cope through identity masking, pirated tools and peer networks while platforms and policy fail to recognize or protect them.

Main Finding

Platform workers in Bangladesh materially undergird the production and maintenance of AI systems—through annotation, moderation, design, and other hidden tasks—while their contributions remain systematically invisible, undervalued, and precarious. The authors coin "Ghostcrafting AI" to capture how these workers both enable AI and are erased from recognition; they survive via situated, communal learning and tactical repertoires (e.g., identity masking, pirated tools), revealing important economic externalities and governance gaps in global AI supply chains.

Key Points

• Concept: "Ghostcrafting AI" extends Gray & Suri’s Ghost Work by emphasizing (a) the situated, communal practices that produce work in the Global South and (b) the tactical survival strategies workers adopt to cope with precarity and erasure.
• Worker motivations: Platform work is framed locally as prestigious and a pathway to socioeconomic mobility despite low pay and instability.
• Learning & skill formation: Skills are acquired informally—renting cyber-café time, copying top-seller gig templates, following foreign-language tutorials, and relying on WhatsApp peer networks—rather than via formal training.
• Forms of invisibility/erasure: NDAs, bans on showcasing work, platform policies, and client practices deny authorship and recognition; algorithmic visibility and rating systems penalize downtime and create long-term precarity.
• Tactical repertoires: Workers use VPNs and borrowed IDs to appear Western, embed contact info in images to avoid platform fees, use pirated “safe” software, accept many low-pay tasks to boost metrics, and rely on communal troubleshooting.
• Structural harms: Exploitative wages, unreliable payments, biased task allocation, and opaque governance concentrate power with platforms/clients and externalize costs onto low-wage workers.
• Proposed remedies (from the paper): peer learning infrastructures, accountable/transparent algorithms, recognition mechanisms (portable credentials), regional training funds, emergency income supports, and redistributive obligations for platforms and AI system purchasers.

Data & Methods

• Study design: Eight-month ethnography (three field periods: May–Aug 2024; Dec 2024–Jan 2025; May–Aug 2025) across seven sites in Dhaka and Jashore, Bangladesh.
• Methods: 18 observation sessions (~60 hours), 34 semi-structured interviews (45–60 min each; 32 audio-recorded), three biographical sketches, fieldnotes and photos.
• Participants: N = 34 platform-affiliated workers and stakeholders (28 male, 6 female), age 19–39 (median 27). Roles included data annotation (15), graphic/UI-UX design (14), content writing/translation (10), web development (14), data entry (8), trainers (7), plus other roles.
• Platforms represented: Upwork (24), Fiverr (23), Freelancer.com (15), Crowdgen (10), 99designs (9), Toloka (6), PeoplePerHour (5).
• Education: distributed from primary to master’s degrees (most in secondary/higher secondary); variable digital literacy.
• Analysis: Transcription and translation (Bengali→English), open coding and thematic analysis; codebook and IRB-approved procedures; reflexivity noted (authors are Bangladeshi).
• Limitations acknowledged: small, purposive sample; Bangladesh-specific context; qualitative, not nationally representative; reliance on self-report and ethnographic observation.

Implications for AI Economics

• Hidden labor costs and externalities: Economies of AI underprice the labor that creates training/label data. Suppliers (platforms/clients) externalize labor costs (low wages, no benefits), which biases the true cost structure of AI systems and skews investment/price signals in markets.
• Global wage arbitrage and monopsonistic pressure: Platforms mediate cross-border labor markets, enabling wage arbitrage that depresses bargaining power in origin countries. This can sustain AI development at lower reported labor cost but amplifies inequality and may reduce overall labor quality and retention.
• Quality vs. cost trade-offs: Reliance on precarious, undertrained workers may affect data quality, consistency, and long-term reliability of AI outputs. Short-term cost savings could increase downstream model error, maintenance costs, and reputational risk—economic trade-offs that buyers underestimate.
• Market failures & information asymmetries: Opaque platform algorithms, NDAs, and invisible authorship generate information asymmetries that hinder labor-market efficiency (e.g., skill signaling, credible credentials, contracting).
• Policy and procurement levers with economic impact:
  • Wage/contract standards or procurement requirements for “responsible AI” that internalize labor costs (e.g., require fair-pay evidence for data-labeling contracts).
  • Portable credentials and verified skill registries to reduce frictions and improve labor-market matching and wage signaling.
  • Regional training funds/subsidies to correct underinvestment in human capital where private incentives are weak.
  • Transparency mandates and audits (algorithmic task allocation, payment reliability) to reduce monopsony power and improve market outcomes.
  • Social insurance or emergency income supports to stabilize worker supply and reduce turnover/externalities.
• Measurement and research priorities for AI economics:
  • Quantify the share of model development cost attributable to paid human labeling/cleaning and the distribution of that spend across geographies and worker incomes.
  • Estimate welfare losses from underpaid labor and erasure (e.g., lifetime earnings, health/stress costs) to inform true social cost of AI services.
  • Study how buyer-side procurement choices (price, time pressure, NDA terms) influence data quality and downstream model performance and costs.
  • Analyze market structure (platform concentration, buyer-side power) and its effects on wages, entry, and innovation.
• Corporate governance & liability: Firms procuring data/annotators should account for labor practices in their ESG and risk models; failing to do so creates reputational and regulatory risk that has economic consequences (litigation, consumer backlash, regulation).
• Long-run labor market dynamics: Without interventions, “ghostcrafting” may lock a segment of the Global South into precarious, low‑skill work, reducing opportunities for productive upgrading—this has macroeconomic implications for human capital accumulation and inclusive growth.

Suggested actionable economic policy experiments (aligned with the paper): - Pilot procurement requirements where buyers pay a certified “fair wage” for labeling tasks and track effects on cost, quality, and worker outcomes. - Implement regionally funded training + portable credential pilots to test whether credentialing improves wages/matching and reduces need for tactical (and potentially illegal) survival strategies. - Mandate disclosure of task allocation algorithms for large platforms in selected jurisdictions and measure impacts on distribution of work and worker earnings.

Limitations and open questions for economists: - How generalizable are the findings beyond Bangladesh and the sampled platforms? - What is the causal effect of improved labor standards on AI data quality, model performance, and buyer costs? - How to design enforceable cross-border labor standards without imposing prohibitive transaction costs?

Overall, this paper documents an undercounted component of AI supply chains—cheap, hidden, communal labor—that has measurable economic consequences. Addressing “ghostcrafting” requires aligning procurement incentives, improving transparency, and investing in worker recognition and training to internalize the human costs of AI production.