Levelling the playing field – the role of AI in employment fairness
A new study finds that generative AI, such as ChatGPT, is helping non-native English-speaking freelancers on digital platforms compete better against native speakers.
Wong Wei Chen
22 December 2025
Technology has often been hailed as the “great equaliser”. For example, when the automobile first made its debut, driving was a rare privilege accessible only to the wealthy. Now, the proliferation and evolution of assembly lines have transformed cars into a mass consumption good.
The advent of generative AI – exemplified by large language models such as ChatGPT and Gemini – promises to level the playing field even further. In their working paper “AI for Fairness? The Role of Generative Tools in Shaping Freelancer Success”, Cao and Li (2025) investigate how ChatGPT eroded structural advantages enjoyed by native speakers of English, helping non-native speakers make deeper inroads into the freelancing industry.
The underlying premise is that with English as the global lingua franca, native speakers have a head start over those grappling with the language. However, the idea that generative AI can bridge this linguistic divide is attractive, and findings from the study corroborate this intuition.
Background
Cao and Li’s paper focuses on Upwork, a leading international freelancing platform where transactions are conducted primarily in English. The platform offers an ideal setting for this natural experiment due to its massive global footprint: at the time of the study, Upwork operated in over 220 countries with more than 800,000 active clients, and in 2023 alone, more than US$4.1 billion was transacted on the platform.
Against this backdrop, how did the arrival of ChatGPT alter the competitive dynamics between native and non-native freelancers?
Empirical design
To isolate the causal impact of the technology, the researchers employed a rigorous empirical design spanning a 22-month observation window. This period is symmetrically divided around the release of ChatGPT in November 2022, covering the 11 months immediately preceding the launch and the 11 months following it.
The study utilises a Difference-in-Differences (DiD) methodology, which relies on a comparison between two distinct cohorts. Non-native speakers served as the treatment group, as they stand to gain the most from an AI tool that automates linguistic fluency.
Conversely, native speakers formed the control group. The underlying assumption is that for individuals who already possess 100% fluency, the marginal benefit of ChatGPT’s linguistic capabilities is negligible, thereby rendering the tool redundant. Thus, the native speaker cohort provides the counterfactual scenario: a baseline representing how the market trajectory would have evolved had ChatGPT not been introduced.
Findings
Amid an overall decline in the employment market, findings revealed that non-native freelancers fared better than native speakers. They suffered 1.5% fewer job losses, experienced 4% smaller reductions in earnings per job, and achieved a 5.5% relative improvement in total earnings.
Beyond improved earnings, the distribution of work also shifted. The study found that non-native speakers successfully moved up the value chain, securing higher-budget contracts and taking on tasks with longer completion times, which supports the idea that they were taking on more complex or demanding tasks.
Finally, generative AI appeared to have reduced the friction of cross-border trade. Following the adoption of ChatGPT, fluent non-native speakers formed significantly more connections with clients from high-income countries, which typically offer higher-paying jobs. This suggests that by smoothing out linguistic friction, ChatGPT had effectively unlocked access to wealthier markets – e.g. USA, Australia, UK and Canada – that were previously insulated by a "native fluency" premium.
Robustness test
To further support the hypothesis that ChatGPT had been instrumental in improving employment outcomes for non-native speakers on Upwork, Cao and Li conducted a “placebo test” by shifting the intervention timeline back by one year. In other words, instead of using the original 22-month observation window (January 2022 to October 2023), the regression models were applied to the period January 2021 to October 2022, assuming a hypothetical release of ChatGPT in November 2021 (instead of actual release in November 2022).
The estimated impact of this hypothetical shock was significantly weaker, yielding estimates of 1.3% for the number of jobs, 3.2% for total earnings, and 2.2% for average earnings. This confirms that ChatGPT acted as a powerful accelerator that significantly amplified these gains for non-native speakers.
The statistical significance of these placebo estimates, however, points to a valuable nuance: a structural convergence between native and non-native speakers was likely already underway. This implies that the pre-shock trends might not be perfectly parallel – which is a standard prerequisite for the DiD methodology – and also suggests that the study's main findings could, as the researchers themselves pointed out, “overstate the causal effect of GenAI technologies in reducing language barriers and improving job performance for non-native English speakers”.
Conclusion
Cao and Li’s study provides new empirical evidence on how generative AI technologies, such as ChatGPT, are reshaping outcomes in digital labour platforms. Notwithstanding longstanding disparities in job access and earnings between native and non-native speakers, the findings suggest that GenAI can help narrow these gaps.
Specifically, the technology enabled non-native speakers to mitigate job losses and access higher-paying employers, particularly in complex roles. However, the researchers note that this impact is not uniform. While technical fields like engineering saw clear gains in efficiency and compensation, service-oriented roles (such as administrative support and customer service) experienced mixed feedback.
The study concludes that if appropriately integrated, GenAI has the potential to democratise economic opportunity. Yet, the mixed outcomes and potential displacement in certain quarters highlight a critical need for platforms and policymakers to monitor quality outcomes and ensure inclusive adoption going forward.
Cao, Jingman is a PhD candidate at the Department of Economics, National University of Singapore.
Li, Zhonglin is an assistant professor in the Department of Real Estate, NUS Business School, National University of Singapore.