By Qiu Bin and Zhang Qun

Qiu Bin is a Professor at the School of Economics and Management, Southeast University.

Zhang Qun is a Tenure-Track Associate Professor at the School of Economics, Zhongnan University of Economics and Law.

Originally published in Teahouse for Economists, Issue 2, 2025.

Introduction

On August 7, 2025, OpenAI released GPT-5, sparking an immediate backlash from established GPT-4o subscribers who launched public petitions demanding the restoration of the earlier model. The reaction surprised OpenAI CEO Sam Altman. The release had been framed as a technical milestone: GPT-5 matched panels of domain PhDs, reduced hallucinations, and eliminated the sycophantic tendencies of prior releases.

Yet many paying users had treated GPT-4o as a conversational confidant and therapeutic sounding board. When the updated model adopted an analytical, objective, and distant persona, OpenAI faced a public relations challenge.

By SOLOMOAT Editorial Team

The episode highlights how model capabilities diverge from initial institutional expectations. In an April 2025 interview with Bloomberg News, Altman acknowledged two outcomes he had not anticipated: first, generative AI directly challenging Google’s core search franchise; second, models partially absorbing diagnostic workflows in healthcare.

The displacement of legacy keyword search is already evident. Users querying ChatGPT, DeepSeek, Doubao, or Kimi rely on direct multi-step reasoning, real-time web retrieval, and structured text synthesis rather than scanning index links.

In healthcare, particularly in regions constrained by scarce clinical appointments and insurance gaps, patients input lab metrics and symptom logs to receive diagnostic evaluations that rival professional consultations.

These developments raise core economic questions:

Which specific labor segments and occupational roles are being substituted?

What underlying economic functions are models displacing beyond routine manual tasks?

What structural reorganizations and institutional shifts will this substitution trigger?

Who Gets Replaced?

Technological substitution operates as a continuous transition rather than an instantaneous event.

Forecasts from institutional consultancies point to broad occupational exposure. McKinsey projects that generative AI platforms will automate roughly 8% of labor tasks across economic sectors; combined with traditional industrial robotics, aggregate workplace automation could reach 30%, potentially impacting 300 million roles globally over a five-year horizon.

Analyses from the United Kingdom identify ten highly vulnerable occupations: data entry clerks, telemarketers, customer service representatives, cashiers, copy editors, paralegals, bookkeepers, frontline food service staff, warehouse logistics workers, and junior business analysts.

Conversely, Microsoft research indicates that roles with low substitution risk concentrate in physical healthcare and non-routine manual trades: surgical assistants, registered nurses, physical therapists, ophthalmic technicians, dredge operators, and commercial roofers.

Standardized occupational taxonomies illustrate the breadth of this exposure. The International Standard Classification of Occupations (ISCO-08) identifies 10 major groups, 43 sub-major groups, 130 minor groups, and 436 unit groups across management, professional services, and manufacturing.

In China, the Occupational Classification Code of the People's Republic of China (2022 Edition) catalogs 8 major categories, 79 medium groups, 449 minor groups, and 1,636 detailed unit roles—spanning legacy trades alongside emergent titles like AI model trainers, live-stream sales operators, and commercial drone pilots.

What Gets Replaced?

In 2049: Possibilities for the Next 10,000 Days, Kevin Kelly and Wu Chen argue that generative models first absorb structured, repetitive workflows—such as call center interactions and corporate expense processing.

At the executive level, C-suite responsibilities change very little; frontline labor speeds up; the primary structural shock hits middle management. If a manager's primary function is relaying directives downward and compiling status reports upward, automated systems execute that coordination with zero operational friction.

This restructuring mirrors earlier enterprise transformations. While the deployment of Business Process Reengineering (BPR) via SAP systems in the 1990s flattened operational processes, generative models directly eliminate entire managerial tiers.

This dynamic reflects Thomas Friedman’s flat-world thesis: the first wave of globalization was driven by nation-states, the second by multinational corporations, and the modern phase by software-empowered individuals.

From an analytical standpoint, artificial intelligence substitutes for specific functional utilities and discrete tasks rather than whole job titles.

In a June 2025 National Bureau of Economic Research (NBER) working paper titled Expertise, David Autor demonstrates that when sub-tasks within a role are automated, the economic value of the remaining labor depends on whether the removal of those tasks raises or lowers the skill requirements of the unautomated work:

This mechanism is visible across modern enterprise deployments.

When Unitree humanoid robots replaced human performers during the 2025 CCTV Spring Festival Gala, direct demand for traditional stage dancers fell, while demand for robotic choreographers, audio engineers, and hardware technicians expanded.

Similarly, in advanced semiconductor fabrication facilities, end-to-end automated control systems execute core processing routines, leaving production oversight to a small group of specialized systems engineers.

Can AI Resolve Baumol’s Cost Disease?

In our 2025 analysis published in Management World, we evaluated labor substitution across the digital economy, specifically addressing whether generative AI alleviates or worsens Baumol's Cost Disease.

William Baumol established that labor productivity expands more slowly in services than in manufacturing. As manufacturing productivity rises, market-wide wage gains push up the price of service deliverables. Because demand for essential services (such as healthcare and education) exhibits low price elasticity and high income elasticity, aggregate demand persists.

Because service productivity remains stagnant, meeting this demand requires absorbing a larger share of the labor force, eventually dragging down economy-wide productivity growth.

A common assumption is that deploying AI across services—from digital workflows to finance, medicine, and education—will lift sector productivity and resolve Baumol's disease.

Yet Baumol's disease is driven by inter-sectoral productivity differentials. The issue is not whether service productivity improves in absolute terms, but whether it matches or exceeds the rate of expansion in manufacturing.

Data from the China Digital Economy Development Report (2024) shows that the productivity gains driven by digital inputs in manufacturing have outpaced those in services since 2021. Digital tools have made modest progress in overcoming the non-tradable, non-storable nature of traditional services, which can widen the productivity gap between industrial production, digital services, and traditional consumer services.

Evaluating this dynamic through the lens of AI adoption reveals two distinct sectors: AI-Intensive Sectors and AI-Lagging Sectors.

Expanding productivity spreads between these two segments can create a modern variant of Baumol's disease, where high-adoption sectors displace labor faster than traditional sectors can productively absorb it. Under these conditions, the classical assumption of market-wide full employment breaks down.

Resolving this structural imbalance requires directing capital and model architectures toward specialized, domain-specific service applications rather than deploying generic horizontal tooling.

Because foundation models carry high fixed capital costs and near-zero marginal inference costs, unconstrained market mechanisms will generate deep structural displacement and transitional unemployment across misaligned skill brackets.

Institutional Governance and Macroeconomic Risks

Evaluating labor substitution requires examining the institutional incentives of market participants.

As Joseph Stiglitz noted in Globalization and Its Discontents, international economic integration was largely directed by multinational enterprises and institutional financiers to maximize capital extraction, often generating structural domestic dislocations.

An analogous dynamic is visible across today's concentrated technology conglomerates. Driven by return-on-equity mandates, these firms possess both the capital scale and operational incentive to substitute software for human labor without internalizing the broader social costs.

Human capital development and systemic stability must take precedence over pure capital efficiency. Labor laws must update to provide flexible protections for freelance, contract, and platform-mediated workers.

Furthermore, capital inflows into artificial intelligence can decouple from underlying productivity contributions, inflating asset valuations and misallocating capital into oligopolistic platforms. This dynamic raises barriers to entry, reinforces data monopolies, and slows technology diffusion across mid-market enterprises.

Proactive policy frameworks across competition enforcement, cross-platform data interoperability, and intellectual property licensing are essential to prevent technological substitution from worsening structural inequality and macroeconomic instability.

Frequently Asked Questions

What does AI usually replace first?

AI first compresses repeatable information work and standardized execution steps, not accountability or strategic judgment.

How should operators respond?

Repackage expertise around decisions, context, quality control, and customer outcomes that do not reduce to a generic task.