Photo Credit: Grab
Scientific Frontline: Extended "At a Glance" Summary: AI Tools and the Disabled Worker Wage Gap
The Core Concept: Simple artificial intelligence text-to-speech tools used by deaf and hard-of-hearing delivery drivers measurably improved communication with customers, closing a significant portion of the wage and performance gap with their non-disabled peers.
Key Distinction/Mechanism: Rather than utilizing complex large language models or displacing human labor, the intervention utilized low-cost AI to address specific communication bottlenecks during the "last mile" of food delivery.
Origin/History: The findings are based on a 2026 working paper from the National Bureau of Economic Research, analyzing data from deaf and hard-of-hearing drivers working for a major Chinese food delivery platform.
Major Frameworks/Components:
- Personnel Economics: The study applied principles of labor economics to evaluate how workplace accommodations affect marginalized workers within a specific corporate environment.
- Efficiency vs. Labor Supply: Researchers found that prior to the AI tool, disabled workers experienced lower efficiency (slower deliveries, more negative ratings) but compensated with a higher overall labor supply (more hours worked, lower quit rates).
- Wage Gap Reduction: Implementation of the text-to-speech outbound calling tool eliminated roughly one-third of the hourly wage gap and reduced negative customer ratings by two-thirds.
- Disability Severity Correlation: The data indicated that profoundly deaf workers benefited more from the AI intervention than those who were hard-of-hearing.
Branch of Science: Artificial Intelligence, Economics, and Human-Computer Interaction.
Future Application: The framework suggests low-cost AI could provide targeted assistance for other disabilities in the workforce, such as screen readers for low-vision employees or reading/spelling aids for those with dyslexia, allowing them to compete equitably in roles where their disability presents only isolated challenges.
Why It Matters: The research provides empirical evidence that AI can be leveraged to augment and support traditionally disadvantaged human workers, rather than replacing them, while simultaneously proving profitable for the employing platform.
Researchers at UC Santa Barbara, the University of Toronto, and Zhejiang University have found that using artificial intelligence-based tools can help some disabled workers improve their performance and productivity, potentially putting them on par with their nondisabled counterparts.
“There are a lot of jobs where your disability wouldn’t really matter, except for one or two areas where it might matter a lot,” said UCSB economist Mitch Hoffman, co-author of a working paper for the National Bureau of Economic Research. In these areas, according to the researchers, AI could be an effective tool for closing the gap.
Such has been the case for deaf and hard-of-hearing (DHH) workers in China working for a major food delivery platform. Using AI to better communicate with customers, DHH workers were able to measurably improve their outcomes, increasing their income while profiting the company—a win-win.
Improving Communication and Efficiency
People with disabilities tend to face challenges in the working world. Built mainly around the able-bodied, the circumstances surrounding labor often present obstacles that are unfavorable to a person with a disability, and these obstacles often translate into less pay or fewer opportunities to work. On a larger scale, workers with disabilities have lower rates of employment and face discrimination. Companies, according to the researchers, may choose to hire fewer disabled workers or hire them for lower-paying jobs, seeing the investment in workplace accommodations as an increased cost and the hiring of disabled workers as a legal risk. In places like China, where social insurance—such as disability pay, workers’ compensation, healthcare, and retirement plans—is less developed, the challenges, according to the paper, “are especially acute.”
So it was with great interest that one of Hoffman’s Chinese colleagues found themselves interacting on the phone with an AI tool on one of the country’s major food delivery platforms. Not only did it spark questions about the effectiveness of the tool and its benefit to the worker, but the size of the platform also promised an ideal setting for the researchers’ inquiry.
“It’s a very large platform, and it’s exciting to do research that has the scope to benefit a lot of people,” said Hoffman, whose research interests lie in the realm of personnel economics, which is the application of labor to issues inside companies. “There’s growing interest in personnel economics and understanding how different human resource policies affect workers with different levels of disadvantage,” he said.
Additionally, according to the paper, the platform follows an open hiring policy, taking all who meet the basic requirements for the job, which in this case is food delivery. Wages are based on deliveries made and customer satisfaction, and this type of work has become a key source of employment for disabled workers in China. All these factors lent themselves well to a robust experiment to answer the researchers’ questions regarding how disabled and nondisabled workers compare when working the same job, and how AI tools can affect outcomes for workers with disabilities.
“We thought that this was an interesting situation where AI was being used in a different way, where it wasn’t replacing someone’s job but helping a type of worker who was traditionally disadvantaged benefit.”
The researchers focused on DHH workers, who, like their nondisabled counterparts, were required to accept orders, pick up the food, and deliver it to the customers, mediated by the platform.
“You’d think of food delivery as a job that doesn’t seem like it has a lot of communication, and that’s true, but it does involve communication often at some critical points on some orders,” Hoffman said. In particular, the delivery step of the task—the so-called “last mile”—could become fraught with confusion and delay as the drivers contend with finding the specific location to drop off the food. “For example, if you’re going into a giant high-rise and you can’t get in the door, it’s really important to be able to contact your customer,” he explained. “Or if someone says, ‘leave it here.’ What does ‘here’ mean?” The ability of the worker to solve those puzzles can mean the difference between a good customer rating and a bad rating for late, damaged, or failed deliveries and a subsequent monetary penalty.
Prior to the AI tool, which is a text-to-speech outbound phone calling system, the DHH workers were slightly slower than nondisabled workers, particularly during the crucial food drop-off phase. They experienced 9% more late deliveries and received 31% more negative customer ratings.
“But they do work more hours, and they’re less likely to quit,” Hoffman added. “So in that sense, they’re actually profitable workers for the platform pre-AI.” Indeed, according to the paper, “the higher labor supply of DHH workers more than offsets their small efficiency gap, with DHH workers completing more orders per week and generating higher weekly profits for the platform than non-disabled workers.” However, because the DHH workers were slower and received more bad reviews, their wages were about 10% lower than those of nondisabled workers.
With the introduction of the AI tool, which allowed the DHH workers to call customers and communicate with them using a realistic voice, there was a significant improvement in their outcomes.
“We found that this AI tool substantially reduces some of the gaps we saw pre-AI,” Hoffman said, noting that it eliminated a third of the hourly wage gap between DHH and nondisabled workers and reduced the bad customer rating gap by about two-thirds.
Over the course of the study, the researchers also found that the tool benefited profoundly disabled workers (severely deaf) more than nonprofoundly disabled (hard-of-hearing) workers, and that it had larger effects on hours, income, and profits in stronger (lower unemployment) local labor markets than in weaker (higher unemployment) local labor markets.
The tool itself was also relatively inexpensive and straightforward, according to the paper, meaning both the cost and the effort spent to implement it were small relative to the benefit to both the workers and the company. Rather than being a large language model, it was a simpler tool that converted text to speech and allowed for automatic prompts, which saved time during delivery. It didn’t solve all the communication-related problems that could arise, but it handled the most common issues.
While the study concentrated on DHH workers, there are implications for workers with other disabilities as well—workers who may have the know-how for the job but could use a little assistance now and then.
“If you’re a low-vision person, you might be able to do most of a knowledge work job just fine, but having an AI tool that could tell you what’s on the screen might be very useful,” Hoffman said. “Or if you have dyslexia, you might be totally fine for almost everything, but AI could be useful for reading, spelling, and a few things like that.”
Given the current interest in the negative potential of AI and concerns over widespread job loss, environmental problems, and other deleterious consequences, this research highlights something different: AI and its ability to help human workers perform better and get more out of their work.
“We thought that this was an interesting situation where AI was being used in a different way, where it wasn’t replacing someone’s job but helping a type of worker who was traditionally disadvantaged benefit,” said Hoffman.
Funding: The research is supported by Schmidt Sciences’ AI at Work program.
Published in journal: National Bureau of Economic Research
Title: Empowering Inclusive Work
Authors: Yanyou Chen, Mitchell Hoffman, Huilan Xu, and Zhe Yuan
Source/Credit: University of California, Santa Barbara | Sonia Fernandez
Edited by: Scientific Frontline
Reference Number: ai091426_02
