In 2021, a few isolated cases emerged in India where gig workers said they were locked out of their accounts when the platform’s facial recognition system failed to recognize them. Uber routinely asks drivers to take a selfie to check if their face matches the image that the company has on file through a program called “Real-Time ID Check.” It was rolled out in India in 2017 and can request drivers’ selfies as often as multiple times a day.
In early 2021 an Uber driver in the southern Indian state of Telangana claimed that the system did not recognize him after he shaved his head. Another food delivery worker on the Indian food delivery platform Swiggy claimed the same issue of being locked out of his account after the platform’s facial recognition technology did not recognize him following a shaved head and face.
As a nonprofit journalism organization, we depend on your support to fund more than 170 reporting projects every year on critical global and local issues. Donate any amount today to become a Pulitzer Center Champion and receive exclusive benefits!
Both of these cases intrigued me. At the same time, Uber drivers in the U.K. had accused the platform’s facial recognition system of failing to identify people of color. At least one union in the U.K. pursued legal action against Uber; Uber settled this case out of court.
These incidents made me wonder if there was an unreported story of more Uber drivers in India facing similar issues: where a possibly imperfect—worse, inaccurate—facial recognition system was mediating access to work for an already precarious working class. That’s what prompted me to investigate the impact of facial recognition on Uber drivers in India.
As part of my investigation, I conducted a survey of 150 Uber drivers across different parts of India to find out how many of them had been locked out of their accounts—either temporarily or permanently—due to issues related to facial recognition. This investigative effort prompted the gig workers' union to start collecting their own data to petition the platforms.
Hypothesis and data collection
What was my hypothesis?
With people of color in the U.K. and drivers in India getting locked out of their Uber accounts due to a false mismatch of the face ID system, my hypothesis was that—as with most facial recognition systems—there wasn’t ample training data of diverse faces, leading to issues for people of color whenever there was a slight change in appearance or lighting conditions.
What prompted me to come up with the survey?
I started my reporting by just reaching out to some gig worker unions. I was hoping to connect with distressed drivers who were locked out of their accounts due to a false image mismatch. After weeks of failed attempts, I realized I’d have to go a step further and create my own data set to truly understand the scale of the problem.
As soon as I realized that I’d have to create my own dataset, I knew the only way to do it was to conduct a survey of a random sample of drivers across some of the top cities in India.
How did you design your survey and how did you disseminate it?
I wanted to ensure the survey was short, straightforward, and clear. I started with basic demographic questions and questions about the frequency of being asked for a selfie for verification. The goal was to get an idea on whether this was an arbitrary phenomenon or something consistent across many drivers. Then I moved on to the central question of whether they had ever been locked out of their account due to a selfie mismatch. If they said no, they would automatically exit and submit the survey. If they said yes, they were directed to the second part of the survey which had detailed questions to understand what exactly happened to them.
Challenges and changing strategies
- Language: Once the survey was drafted and ready in English, the next task was to translate it into multiple languages and share it with drivers across different cities. I translated it into Hindi, Telugu, Kannada, and Punjabi.
- Mass dissemination: After this, I felt the easiest way to disseminate the survey would be to share them in WhatsApp and Telegram groups of gig workers through the different unions in each city. But I was wrong. Most of these groups are typically flooded with messages, making it difficult to command their attention in a sea of messages. Each group has thousands of members. I didn’t get more than 1-2 responses over the next several days.
- On-the-ground survey: I had to immediately switch to the next strategy. The first step was to target the city I was present in: Bangalore. I drove to the city’s airport which has designated Uber Zones where one can easily find hundreds of drivers stationed, waiting for their next ride. I thought I hit the jackpot, but Uber supervisors there played spoilsport. As I approached a few drivers leaning on their respective cars on a hot summer afternoon in 2022, a couple of Uber representatives did not let me speak with the drivers. I managed to get a few phone numbers of some of these drivers to follow up later. That’s how I got my survey started.
- Partnering with others: Next, I decided to find journalism students in different cities to help me conduct the survey on the ground. The key was to ensure I found people I could trust to avoid all kinds of bad data. I also reapproached the union groups differently by contacting union heads and insisting they share the survey link with their members. This time, I was able to get multiple responses through them. I also called up many Uber drivers, asked the survey questions myself, and filled out the form on their behalf. In the end, I had roughly 250 responses.
- Bad data: Sifting through the responses, I removed bad and irrelevant data. This meant omitting responses from drivers who did not work for Uber (85 of them worked for Indian local cab-hailing platform Ola); duplicate entries; entries that said yes to the main question but the description that followed did not involve facial recognition, et cetera. I was left with 150 usable entries in the survey. Lam Thuy Vo, another AI Accountability Fellow from New York, helped me by discussing the survey approach and ways in which I could clean up bad data once the survey was completed.
Result of the survey & impact
A survey of 150 Uber drivers in India from the top five cities showed that roughly 50% of them had been locked out of their accounts due to a selfie mismatch/facial recognition issue. Many said this was due to a change in appearance such as a change in facial hair or a haircut; others said it was due to bad lighting conditions. A uniquely Indian problem that many complained about was that the front cameras on their low-cost Android phones were usually scratched up, which made selfie-taking difficult.
One of the best things that happened after this survey and story is that it prompted Indian gig worker union Telangana Gig and Platform Workers Union (TGPWU), in collaboration with Biju Mathew, co-founder of New York Taxi Workers Alliance, to collect their own data on how many drivers are being locked out of their accounts without reason. Their plan is to use the data to petition Uber so they cannot block drivers without giving a proper reason.
The founder of TGPWU Shaik Salauddin, who is regularly invited to speak at conferences and meetings to discuss the working conditions of gig workers in India, has frequently presented data from this story and the article itself to explain the situation on the ground. Meanwhile, without help from companies, gig workers are now leaning on the government. Salauddin and gig worker unions in India had a major win recently when Rajasthan became the first state in India to pass a law for gig worker welfare. Salauddin is now working towards other states doing the same, including his home state Telangana.