Nearly half of US companies using ChatGPT say it has already replaced workers

U.S. companies that adopted generative AI tools after November 2022 began cutting spending on freelance labor platforms almost immediately, redirecting those budgets toward AI model providers. That spending shift, tracked at the firm level in a new research paper, offers the first transaction-based evidence that businesses are treating chatbots as direct substitutes for human workers in writing, coding, and data tasks. The finding lands at a moment when hiring managers, freelancers, and policymakers are all trying to separate real displacement from hype.

Spending Shifts After ChatGPT’s Launch Signal Real Displacement

The core tension behind this story is speed. Within weeks of ChatGPT becoming publicly available, firms did not simply experiment with the tool. They changed how they allocated money. The research paper, available as a peer‑review preprint, tracks firm-level spending on online labor marketplaces against spending on AI model providers around the time of ChatGPT’s release. The data shows a measurable drop in demand for contract workers on platforms where companies had previously sourced writing, coding, and data-processing help.

That pattern matters because it converts an abstract debate about automation into observable financial behavior. When a company stops posting jobs on a freelance marketplace and starts paying for API access to a large language model, the substitution is not theoretical. It shows up in transaction records. The hypothesis worth testing against future data is straightforward: firms that cut freelance-platform spending the most after November 2022 should also show above-average declines in new junior-hire postings in their SEC filings or state unemployment insurance records. No public dataset has confirmed that second link yet, but the spending-side evidence is already in hand.

Early adopters did not wait for formal corporate AI policies before making staffing changes. Reporting from a major newspaper documented how some companies began replacing workers with chatbot-generated output even as developers warned that the technology should not be relied on for high-stakes tasks. That gap between corporate caution in public statements and rapid adoption in private budgets is one of the defining contradictions of the current AI rollout.

What “Payrolls to Prompts” Actually Measured

The paper uses a research design built around a natural experiment: ChatGPT’s public release created a sudden, exogenous shock to the availability of generative AI. By comparing firm spending on online labor marketplaces before and after that shock, the authors can isolate changes in demand for human labor that coincide with the adoption of AI tools. The study focuses on transaction-level data rather than surveys, which avoids the self-reporting bias that plagues most estimates of AI-driven job loss.

This distinction is worth spelling out. When a headline says that a large share of companies using generative AI believe it has replaced workers, the underlying evidence often comes from executive surveys where respondents have incentives to overstate or understate adoption. The “Payrolls to Prompts” approach sidesteps that problem by looking at what companies actually spent, not what they told pollsters. The result is a narrower but more reliable picture: firms did redirect money away from human freelancers and toward AI providers, and they did so quickly.

The study’s scope, however, is limited to online labor marketplaces. It does not capture changes in full-time payroll, internal hiring freezes, or layoffs at companies that never used platforms like Upwork in the first place. That means the spending shift it documents is a floor estimate of substitution, not a ceiling. Companies that relied on in-house staff for the same tasks may have made similar changes without leaving a trace in marketplace data.

There are also important nuances in what kinds of work were affected. The sharpest declines in spending appear in tasks that map cleanly onto current generative AI capabilities: drafting marketing copy, summarizing documents, writing simple code, and performing routine data cleaning. More complex projects that require domain expertise, coordination across teams, or direct interaction with customers show smaller and more uneven changes. This pattern suggests that, at least in the short term, AI is displacing the most modular and easily specifiable forms of freelance work.

How Corporate Narratives Diverge From Internal Use

Public messaging about AI has emphasized experimentation, pilot programs, and responsible deployment. Yet the combination of marketplace spending data and inside accounts paints a more aggressive reality. The New York Times, drawing on leaked conversations with early testers and employees, described companies racing to integrate chatbots into workflows even as basic questions about accuracy and safety remained unresolved.

This divergence matters for workers trying to plan their careers. Official statements may reassure employees that AI will “augment, not replace” their roles, while line managers quietly reduce freelance budgets and reassign tasks to automated systems. Freelancers, who lack the protections and visibility of full-time staff, feel the impact first: fewer postings, lower rates, and shorter contracts for work that can be partially automated.

At the same time, some companies are using AI to stretch existing teams rather than cutting headcount outright. In those cases, the displacement is more subtle. Junior roles that once served as entry points-such as content drafting, basic QA testing, or routine data entry-may shrink or disappear, even if overall staff numbers remain stable. The long-term effect is a thinner pipeline of early-career opportunities.

Gaps in the Evidence and What to Watch Next

Several pieces of the puzzle are still missing. No publicly available dataset links the freelance-spending decline directly to headcount reductions in corporate payroll systems. The Bureau of Labor Statistics has not published occupation-level employment figures that isolate the effect of generative AI from broader hiring trends. State unemployment insurance records, which could confirm whether displaced freelancers filed claims, have not been matched to the firms in the study.

The New York Times reporting that documented early corporate adoption relied on internal company sources rather than public filings or verified payroll data. That means the strongest public claims about worker replacement rest on two types of evidence: anonymous accounts from inside companies and transaction-level spending data from labor marketplaces. Neither source, on its own, proves permanent job loss. Together, they build a case that is directionally clear but not yet definitive.

The arXiv study also ends relatively soon after ChatGPT’s launch, so it captures the initial spending response but not the longer-term trajectory. Did companies that cut freelance spending in late 2022 and early 2023 sustain those cuts, or did they return to human labor after discovering the limits of AI-generated output? That question requires follow-up data covering at least four to six additional quarters.

Another open question is how quickly workers can adapt. Some freelancers have begun marketing themselves as “AI copilots,” offering to combine model output with human judgment. If that hybrid approach proves valuable, spending patterns could stabilize at a new equilibrium where AI reduces the volume of work per project but not the number of projects overall. Alternatively, if clients become comfortable interacting directly with AI tools, intermediated human labor could erode further.

Policy and Market Implications

For policymakers, the main lesson is that displacement can show up in financial data before it appears in headline employment statistics. Monitoring transaction-level spending-where privacy rules permit-may offer earlier warning signs than traditional labor surveys. It can also help distinguish between sectors where AI is primarily a productivity tool and those where it is functioning as a labor substitute.

For companies, the study underscores the need to align public messaging with internal practice. If firms are in fact replacing certain categories of work with AI, acknowledging that reality can support more honest planning around retraining, redeployment, and severance. Quiet substitution may minimize short-term controversy but leaves workers with little time to adjust.

For workers and freelancers, the evidence so far suggests that tasks most exposed to generative AI-formulaic writing, basic coding, and routine data work-are already under pressure. Skills that involve client management, complex problem definition, or high-stakes decision-making remain more resilient. The challenge is to move up that value chain quickly enough to stay ahead of the tools.

The transition from payrolls to prompts is still in its early stages, and the data remains incomplete. Yet the combination of marketplace spending records and on-the-ground reporting points in a consistent direction: companies are not just talking about AI; they are reorganizing budgets around it. Whether that shift ultimately produces broader prosperity or deeper precarity will depend on how quickly institutions, and the people who work within them, respond to what the numbers already show.

More from Morning Overview

*This article was researched with the help of AI, with human editors creating the final content.