Uber’s facial recognition is locking Indian drivers out of their accounts 

One aboriginal evening successful February past year, a 23-year-old Uber operator named Niradi Srikanth was getting acceptable to commencement different shift, ferrying passengers astir the southbound Indian metropolis of Hyderabad successful his midsize sedan. He pointed the telephone astatine his look to instrumentality a selfie to verify his identity. The process usually worked seamlessly. But this clip helium was incapable to log in.

It didn’t instrumentality agelong for Srikanth to travel up with a mentation arsenic to why. He had conscionable returned from visiting the Hindu Tirupati temple, 350 miles away, wherever helium had shaved his caput and prayed for a prosperous life. 

The Uber app prompted Srikanth to effort again, truthful helium waited a fewer minutes and took different picture. Rejected again. 

“I was disquieted astir bookings. We person regular targets wherever if we implicit a definite fig of bookings, we get incentives,” Srikanth says. “I was anxious to log successful and commencement driving, and not discarded immoderate time.” So helium tried erstwhile more. This clip helium utilized a 2nd telephone to propulsion up an representation of himself from earlier helium visited the temple. When helium took a representation of it, Uber informed him that his relationship had been blocked.

Srikanth is not alone. In a survey conducted by MIT Technology Review of 150 Uber drivers successful the country, astir fractional had been either temporarily oregon permanently locked retired of their accounts arsenic a effect of problems with their selfie. Many suspected that a alteration successful their appearance, specified arsenic facial hair, a shaved head, oregon a haircut, was to blame. Another 4th of them judge it was owed to debased lighting. 

Srikanth thinks the split-second determination to instrumentality a representation of different telephone outgo him his livelihood: helium went from earning implicit $500 a period to nothing. He spent months afterward trying to get his relationship reinstated, to nary avail. Eventually helium had to determination backmost to his hometown, wherever helium works a fewer antithetic jobs and makes hardly 10% of what helium utilized to.

Srikanth is acold from the lone idiosyncratic successful India who indispensable interact with facial designation technology. In summation to the country’s 600,000 Uber drivers, galore others enactment for the homegrown ride-sharing level Ola and for startups specified arsenic Swiggy, Zomato, and Urban Company. All inquire their level workers to upload selfies for logins oregon verifications. 

Niradi Srikanth, earlier and aft helium changed his facial hairsbreadth and hairsbreadth style.

COURTESY PHOTOS

In different markets, gig workers person fought backmost against facial recognition. In the UK, for example, astatine slightest 35 Uber drivers claimed past twelvemonth that their accounts were wrongly terminated. The Independent Workers’ Union of Great Britain has blamed a “racist algorithm.” Uber has faced astatine slightest 2 lawsuits successful the UK due to the fact that of the software. 

Some countries and regions person moved to supply amended protections for gig workers. The EU projected a directive past twelvemonth to amended moving conditions and supply algorithmic transparency. And successful September 2021, California tribunal struck down Proposition 22, a ballot inaugural that excluded gig workers from worker benefits nether authorities law. These regulations admit that algorithmic systems tin “negatively interaction the rights of workers,” says Divij Joshi, a lawyer and a PhD campaigner astatine University College London. But India presently has fewer ineligible protections successful spot for gig workers, Joshi says: “These aforesaid transparency efforts are not being seen successful India from a argumentation oregon regulatory lens.”

If problems persist—and protections stay limited—they could person an outsize effect, and not conscionable connected work. “Labor platforms successful India are starting to go a cardinal interface betwixt the worker, the market, and the government—they alteration loans for cars oregon adjacent recognition for larger household expenses,” says Aditi Surie, a elder researcher astatine the Indian Institute for Human Settlements, who has done probe connected gig enactment successful India. In a state wherever specified enactment tin catapult idiosyncratic from precarity to a middle-class beingness (especially erstwhile estimates suggest that the majority of people worldwide who fell into poorness during the pandemic unrecorded successful India), getting blocked from oregon kicked disconnected a level tin person devastating consequences.

Uber checks that a driver’s look matches what the institution has connected record done a programme called “Real-Time ID Check.” It was rolled retired successful the US successful 2016, successful India successful 2017, and past successful different markets. “This prevents fraud and protects drivers’ accounts from being compromised. It besides protects riders by gathering different furniture of accountability into the app to guarantee the close idiosyncratic is down the wheel,” Joe Sullivan, Uber’s main information officer, said successful a statement successful 2017.

But the company’s operator verification procedures are acold from seamless. Adnan Taqi, an Uber operator successful Mumbai, ran into occupation with it erstwhile the app prompted him to instrumentality a selfie astir dusk. He was locked retired for 48 hours, a large dent successful his enactment schedule—he says helium drives 18 hours straight, sometimes arsenic overmuch arsenic 24 hours, to beryllium capable to marque a living. Days later, helium took a selfie that locked him retired of his relationship again, this clip for a full week. That time, Taqi suspects, it came down to hair: “I hadn’t shaved for a fewer days and my hairsbreadth had besides grown retired a bit,” helium says. 

More than a twelve drivers interviewed for this communicative elaborate instances of having to find amended lighting to debar being locked retired of their Uber accounts. “Whenever Uber asks for a selfie successful the evenings oregon astatine night, I’ve had to propulsion implicit and spell nether a streetlight to click a wide picture—otherwise determination are chances of getting rejected,” said Santosh Kumar, an Uber operator from Hyderabad. 

Others person struggled with scratches connected their cameras and low-budget smartphones. The occupation isn’t unsocial to Uber. Drivers with Ola, which is backed by SoftBank, look akin issues. 

Some of these struggles tin beryllium explained by earthy limitations successful look designation technology. The bundle starts by converting your look into a acceptable of points, explains Jernej Kavka, an autarkic exertion advisor with entree to Microsoft’s Face API, which is what Uber uses to powerfulness Real-Time ID Check. 

Adnan Taqi holds up his telephone successful the driver's spot of his car. Variations successful lighting and facial hairsbreadth person apt caused him to suffer entree to the app.

SELVAPRAKASH LAKSHMANAN

“With excessive facial hair, the points alteration and it whitethorn not admit wherever the chin is,” Kavka says. The aforesaid happening happens erstwhile determination is debased lighting oregon the phone’s camera doesn’t person a bully contrast. “This makes it hard for the machine to observe edges,” helium explains.

But the bundle whitethorn beryllium particularly brittle successful India. In December 2021, tech argumentation researchers Smriti Parsheera (a chap with the CyberBRICS project) and Gaurav Jain (an economist with the International Finance Corporation) posted a preprint paper that audited 4 commercialized facial processing tools—Amazon’s Rekognition, Microsoft Azure’s Face, Face++, and FaceX—for their show connected Indian faces. When the bundle was applied to a database of 32,184 predetermination candidates, Microsoft’s Face failed to adjacent observe the beingness of a look successful much than 1,000 images, throwing an mistake complaint of much than 3%—the worst among the four. 

It could beryllium that the Uber app is failing drivers due to the fact that its bundle was not trained connected a divers scope of Indian faces, Parsheera says. But she says determination whitethorn beryllium different issues astatine play arsenic well. “There could beryllium a fig of different contributing factors similar lighting, angle, effects of aging, etc.,” she explained successful writing. “But the deficiency of transparency surrounding the usage of specified systems makes it hard to supply a much factual explanation.” 

Microsoft declined to remark successful effect to questions sent by MIT Technology Review.

The problems don’t extremity with the algorithm’s decision. Drivers accidental the grievance redress mechanics that Uber follows is tedious, time-consuming, frustrating, and mostly unhelpful. They accidental they sometimes walk weeks trying to get their issues resolved. “We person to support calling their assistance enactment incessantly earlier they unlock our accounts, perpetually telling america that the server is down,” said Taqi, with a code of frustration—but mostly a consciousness of defeat—in his voice. “It’s similar their server is ever down.”

Uber did not respond to a petition for comment. 

Srikanth visited the Uber halfway astatine slightest 3 times a week for 3 months earlier helium gave up and went backmost home. He stood successful queues with immoderate 80 to 100 different drivers. “The Uber radical kept telling maine my ID is permanently blocked and they can’t truly bash much,” helium recalled. “They said I could spell to the Bangalore [office] oregon conscionable deploy different operator to thrust my car.” 

Elizabeth Anne Watkins, an organizational sociologist from Princeton University who has extensively studied the interaction of facial designation connected Uber drivers successful the US, would apt find this signifier familiar. “Prone to malfunction successful adaptable conditions, the strategy places a dense load connected workers who are near with small organizational enactment erstwhile facial designation fails,” Hawkins, who is present a probe idiosyncratic astatine Intel Labs, wrote successful a 2020 paper. “Further, accountability for individuality verification is shifted to the workers, who carnivore the consequences for systemic failures.”

Samantha Dalal, who studies however workers recognize algorithmic systems, says determination could beryllium much transparency astir however the AI made a decision. “Including immoderate mentation that goes beyond ‘You are deactivated’” would help, says Dalal, a doctoral campaigner astatine the University of Colorado Boulder. “Such capabilities exist.”

Absent immoderate penetration into what the mercurial, non-human brag wants, gig workers effort a batch of proceedings and mistake portion interacting with the apps, Dalal says. In the lawsuit of Srikanth, she explains that since helium “couldn’t spell backmost successful clip to earlier helium had shaved his head, helium did the adjacent champion happening and showed a representation of himself.”

It’s been implicit a twelvemonth since Srikanth was locked retired of Uber. Despite everything, he’s not hostile toward the company. He simply wants his aged beingness back—one wherever helium was capable to marque a beingness for himself successful Hyderabad and physique up immoderate wealth. He can’t ideate returning to the metropolis unless helium tin get down the instrumentality again. 

Varsha Bansal is simply a freelance writer based successful Bangalore. Reporting for this communicative was supported by Pulitzer Center’s AI Accountability Network.