Translate page with Google

Journalist Resource December 18, 2024

How I Investigated the Gig Work Feeding the Global Surveillance Industry

Country:

Author:
Drone
English

AI surveillance tools rely on low-paid workers to label and organize data.

SECTIONS
Tasks on the Toloka platform for NTechLab, asking gig workers to identify the actions of people walking past the camera. Videos courtesy of TBIJ.

As a tech reporter at The Bureau of Investigative Journalism, I worked on our Digital Sweatshops series, reporting on the precarious, low-paid workers powering the AI boom. We investigated working conditions for moderators in Colombia, Amazon video reviewers training surveillance systems in India and Costa Rica, and the workers in Honduras, India, and Guatemala tasked with making dating apps a safe place for users.

After we revealed that Amazon relied on Indian and Costa Rican workers to train the computer vision systems it used for stock management and surveillance of warehouse workers, I began to think more about the role that these kinds of low-paid data workers played in the global surveillance and defense industry unwittingly training systems of repression.

I was interested in the intersection of two harms: working conditions for data workers who deal with precarity, punishing targets, long hours, and low pay, and people who were subject to excessive surveillance by companies or law enforcement. Gig workers are given little if any information about who they are ultimately working for and how their inputs are used, which was another motivation to look into this area.

Initial reporting

I became interested in Toloka, a gig work platform founded in Russia with its headquarters in the Netherlands, due to the Russian government’s extensive use of facial recognition. When I found an initial link between Toloka and NTechLab, a Russian facial recognition company used widely at home and abroad, I knew we had the beginning of a great story. 

Open-source data posted by gig workers ended up being the key evidence base for this piece. Digital gig workers obviously don’t have physical workplaces in which to gather, so there are very active communities online where workers ask each other for help or complain about the stresses of the job. We found these communities by keyword searching “Toloka” across platforms such as YouTube, Reddit, and Facebook.

We trawled these online communities for references to Toloka tasks requested by NTechLab and Tevian, another Russian surveillance company. We found examples which included drawing boxes around “human bodies” in a security camera still of a girls’ dance class, taking a selfie and uploading it alongside a printed photo of your face or studying CCTV footage of people walking down the street to identify what they were doing. A helpful quirk of the Toloka platform meant we could see which company had commissioned each task.

We built up a spreadsheet with additional information about each task, recording the client who had requested it, any descriptions of the task, where it had been posted, who it had been posted by and where they appeared to be located, as well as the posting date. 


Image courtesy of Niamh McIntyre.

What we revealed

Our reporting found that surveillance companies sanctioned for their role in detaining protesters and activists in Russia recruited gig workers around the world via Toloka to train their facial recognition systems.

Toloka recruited workers to perform data training tasks—such as labelling videos taken from street cameras or uploading selfies—for NTechLab and Tevian, another Russian surveillance company. Both were sanctioned under the EU's human rights regime in July 2023 for contributing to the oppression and detention of protestors in Russia.

We found numerous tasks posted on the Toloka platform requested by both Tevian and NTechLab, including after the invasion of Ukraine. We found that workers in India, Turkey, Pakistan, and Bangladesh had done tasks for the two companies, either providing photos to build their image libraries or labelling footage to improve automatic recognition of people and actions, since 2019. The reporting took around four months in total.

The date parameter we had recorded in our spreadsheet turned out to be the most important. Our reporting suggested Toloka continued to work with NTechLab, and potentially Tevian, after they were sanctioned by the EU in July 2023. Screen recordings of gig workers performing the same NTechLab task were posted to YouTube within a day of each other in September 2023, and in one the phone used to complete the task briefly showed news headlines from September 4.

Human stories

It was really important for us to foreground the harms of facial recognition systems used in Moscow by including the voices of the people who had been targeted. For this reason we started our piece with the story of Alexey Gusev, who had fled Russia after being detained and charged for attending a protest in support of the jailed opposition leader Alexei Navalny. Police said his image had been captured on a surveillance camera, according to Gusev.

We also spoke with Julia Scherbakova, who told us she was detained by policemen on her way to drop her daughter off at nursery. They also accused her of attending a pro-Navalny protest, though Julia says this was a case of mistaken identity. 

We decided we would only interview people who’d left Russia after they were detained via facial recognition for safety reasons. We assumed most of the people who would have been targeted by Moscow’s facial recognition would have remained in Russia, so that narrowed the pool in terms of people we could speak to. However, both Alexey and Julia were former local politicians with a public profile in Russia, so it was helpful that they felt able to go on the record with their stories. 

Interviews with workers helped corroborate the open-source information and allowed us to gain a better understanding of what it was like working for the platform. We contacted hundreds of workers through LinkedIn, Reddit, Facebook, and Youtube.

We also spoke with a number of gig workers and included Santosh* and Khalil*’s perspective. Khalil had been concerned after completing a task which involved sending in photos of his face, while Santosh was deeply concerned about the potential for misuse of facial recognition technology.

Challenges

The most complex parts of the story related to EU sanctions law and mapping the corporate structures of Toloka, Tevian, and NTechLab. After the two facial recognition companies were sanctioned by the EU, EU-registered companies should not have done business with them.

Toloka was registered as three companies, in the Netherlands, Switzerland, and Russia, raising questions about a possible sanctions breach. Something which was not immediately clear was that, because the companies were sanctioned under the EU’s human rights regime, and not the Russia sanctions programme, Switzerland does not implement these sanctions.

A further complexity was that Yandex, Toloka’s parent company, was in the midst of separating its Russian arm from its international business. Yandex Russia said they had begun migrating all Russian clients onto a new platform, Yandex Tasks, in December 2023. 

In the transition period before they had been moved onto Yandex Tasks, Tevian and NTechLab used the Toloka platform, but Yandex Russia said “the volume of work offered through the platform…by the two companies was immaterial, both financially and in terms of the number of tasks.” Toloka said Yandex NV (which has now taken over the international business) had never “received any payments from NTechLab or Tevian” and that the companies had dealt with Toloka’s Russian entity.  

We overcame these challenges by seeking advice from a number of different sanctions experts, and carefully documenting the relevant Toloka tasks and corporate structures involved.

Key takeaways for replicating our work

  1. Gig work platforms are completely opaque for journalists and researchers, but journalists have a way in via gig workers, who do have access to the platform. If possible, ask for screenshots, screen recordings, or contracts (in a way that doesn’t compromise worker’s safety or employment).
  2. However, you will need to send hundreds of messages to reach a decent number of interviews, so be persistent. Gig workers will likely be hesitant about speaking to a journalist: It helps to explain clearly what your story is about, what you hope to achieve by publishing it, and lay out steps you’ll take to safeguard their identity.
  3. There’s also a wealth of great material posted by gig workers online, most of which is untapped. This will give you insight into their working conditions, what kind of tasks they’re doing, and which clients they’re working for. Spend time seeking out the most relevant online communities, set up systems to monitor them, and document what you find methodically.
  4. Don’t neglect the human stories that will bring your investigation to life.

RELATED INITIATIVES

Logo: The AI Accountability Network

Initiative

AI Accountability Network

AI Accountability Network
teal halftone illustration of two hands shaking and a scale holding dollar bills

Initiative

Transparency and Governance

Transparency and Governance

RELATED TOPICS

an orange halftone illustration of a hand underneath a drone

Topic

AI Accountability

AI Accountability
technology and society

Topic

Technology and Society

Technology and Society
teal halftone illustration of a construction worker holding a helmet under their arm

Topic

Labor Rights

Labor Rights
Criminal Justice

Topic

Criminal Justice

Criminal Justice
a pink halftone illustration of a woman speaking a microphone while raising a fist

Topic

Democracy and Authoritarianism

Democracy and Authoritarianism