It was a remote work opportunity, with no fixed hours, and it paid in dollars. In a moment of extreme economic crisis playing out in Brazil, the offer seemed irresistible. Felipe, who was unemployed, took the job working for TikTok, the social media giant.
Felipe, who asked The Intercept to use a pseudonym to protect his future employment prospects, worked alongside other Portuguese speakers. The transcription job was simple: Listen to the audio from TikTok videos and write down what was said. The transcribed text, a manager on the transcription project told the workers at one point, would be used to build up ByteDance’s artificial intelligence capabilities.
For Felipe and other transcribers, compensation was linked to performance, and the best transcribers would be recognized with a bonus at the end of the job period. It was to be a three-month gig, from February to April 2021.
For Felipe, the plan to make a little quick money became a hellish experience.
Felipe didn’t last a week.
He quit the same way he’d been given the job: through a WhatsApp message. He had neither a contract nor any documents regulating his employment.
For Felipe, the plan to make a little quick money became a hellish experience. With TikTok’s short-form video format, much of the audio that needed transcription was only a few seconds long. The payment, made in U.S. dollars, was supposed to be $14 for every hour of audio transcribed. Amassing the secondslong clips into an hour of transcribed audio took Felipe about 20 hours. That worked out to only about 70 cents per hour — or 3.85 Brazilian reals, about three-quarters of Brazil’s minimum wage.
The minimum wage, however, did not apply to the TikTok transcribers — like many other workers, the transcription job used the gig economy model, a favorite of tech firms. Gig economy workers are not protected by some labor laws; they are considered independent contractors rather than employees or even wage earners. In the case of the TikTok transcribers, who did not even have formal contracts, pay was based on how much transcribing they did rather than the hours they worked.
The production level needed to make decent money was staggering. The work was so overwhelming that Felipe ended up leaving before his agreed-upon three months were up — without receiving any compensation.
In 2020, TikTok was the most downloaded app in the world. For ByteDance, the Chinese company that owns TikTok, the popularity meant huge amounts of cash: ByteDance doubled its profits last year, making $34 billion, a 111 percent increase over the previous year. The company has been valued at more than $400 billion for private equity investors.
TikTok exploded in Brazil in particular. This past April, people in Brazil downloaded the app more than any other nation. The company is expanding its operations in the country: TikTok has hired local executives and listed jobs for managerial positions. To keep the app running, the company hired workers to transcribe the high volume of content.
In several instances examined by The Intercept, transcription workers said they had not received promised payments.
The Intercept interviewed four transcription workers in Brazil and had access to WhatsApp groups of temporary workers hired through subcontractors to transcribe for ByteDance. The Intercept reviewed documents used to manage the transcription projects, such as spreadsheets logging productivity, and videos of meetings in which managers addressed transcription workers.
The workers were separated during the recruitment process into small group chats on WhatsApp to streamline communications. In these chats, dozens of temporary transcribers complained about the workload. In several instances examined by The Intercept, transcription workers said they had not received promised payments.
The Brazilian office of ByteDance declined to comment for this story and referred The Intercept to the company’s U.S. offices. The U.S. branch of ByteDance did not respond to multiple requests for comment.
None of the transcription workers were hired directly by ByteDance. Instead, intermediaries offered subcontracted transcription services to the Chinese social media giant. The chain of subcontractors wound its way through Pakistan before ending up in Brazil.
The subcontractors searched for potential workers on social media networks. The first point of contact for many interested applicants was a WhatsApp account belonging to Natasha De Rose, a clinical psychologist based in Rio de Janeiro who functioned as a de facto recruiter despite having no direct link to the company.
On March 1, a call for applications was published on De Rose’s Facebook page. “Whoever needs freelance work and is willing to commit to a home office job, I am recruiting people to work on transcription pt-br,” the post said, using a common abbreviation for Brazilian Portuguese. “More information ONLY ON MY WHATSAPP.” (De Rose did not respond to a request for comment from The Intercept.)
After being retained, workers were separated into the small WhatsApp groups. The titles of the chat groups all listed ByteDance’s name, in addition to a number. Felipe was in group 6.0.
Felipe’s work was completely virtual. Before beginning, he watched a short training video online. In a recorded virtual meeting that was made available to new recruits, recently hired workers learned to cut audio, transcribe selected segments, and navigate the platform used for the service; they also took a quick class on Portuguese and the rules of transcription.
Managers said the work was ultimately being done for ByteDance. The transcription service was done through an app downloaded from a link whose URL began with “Bytelemon”; a banner on the page said, “This app is only for ByteDancers and Teams related to ByteDance.” One of the ways to log in was with ByteDance credentials.
According to videos of meetings and conversation logs reviewed by The Intercept, money to fund the project came from abroad and was transferred from an account belonging to Maria Clara Alarcão, a project manager responsible for the training.
Neither Alarcão nor De Rose are ByteDance employees. Alarcão’s LinkedIn profile, which is otherwise almost bereft of content, describes her as a “project manager”; the profile follows just one company, Transcribe Guru, a transcription services platform. In a message to the text group, Alarcão said she was retained by ByteDance: “As I have explained in meetings, we were contracted by a Chinese company, ByteDance, to provide transcription services for a client,” she wrote on April 3 to the WhatsApp group. (Alarcão did not respond to requests for comment from The Intercept.)
Transcribe Guru, co-founded by Pakistani businessperson Izhar Roghani, provides transcription and translation services. The company’s website shows the ByteDance logo among a list of its transcription clients. “We own a global Workforce of more than 500 employees with majority from Afghanistan, Nepal, Sri Lanka, Libya, Brazil, Pakistan, Portugal, China, and Palestine,” the company boasts on its site.
In a WhatsApp audio message obtained by The Intercept, one of the managers for the Brazilian transcription project, Leandra Narciso, explains that she and others working with her are part of an extended chain of outsourcing. “Izhar doesn’t even have a contract with ByteDance, they’re outsourced,” Narciso says. She adds that though ByteDance is a “good company,” the chain of subcontractors means that there is less money to go toward the end of the chain. “By the time it reaches us,” she says of the money from the contracts, “the value has already decreased a lot.”
In response to an inquiry from The Intercept, Roghani confirmed that Transcribe Guru does not have a direct contract with ByteDance. Instead, he said, the TikTok work came as a subcontractor for another Pakistani company called Little Walto Technologies. Roghani declined to provide contact information for Little Walto, and the company has no website or publicly available contact information beyond an address at a commercial center in Islamabad. Roghani said he could not show The Intercept his contract with Little Walto because it was confidential.
Asked about his subcontractors in Brazil, Roghani said they did not have formal contracts. “I don’t have one with Maria Clara Alarcão,” he said. “We work on trust.”
Treated as gig worker-style independent contractors, transcription workers at the bottom of the subcontracting chain were paid on production: Rather than an hourly wage based on how much time they worked, workers were paid based on the total time of video content that they transcribed. Each transcription — of a small clip of audio, usually a few seconds long and internally called a “task” — took an average of one minute to complete, according to internal documents. Felipe confirmed the figure.
The pay scale also divided workers into different levels of production. Those who completed 300 or more transcriptions per day would receive $14 for each hour of TikTok content they transcribed. Those below the daily threshold were paid $10 per hour of content.
To reach $14 of total pay — by completing transcriptions for one total hour of video content — a worker would need to work around 20 uninterrupted hours.
Based on the internal document’s estimate of how long a task takes, to produce one hour of transcribed content would require around 1,200 tasks, or 1,200 minutes. In a March WhatsApp message sent to employees, De Rose confirmed that 600 transcriptions would produce, on average, a half-hour of transcribed content. In other words, to reach $14 of total pay — by completing transcriptions for one total hour of video content — a transcriber would need to work around 20 uninterrupted hours.
Compensation packages were managed by Alarcão. On April 11, Alarcão recorded a virtual meeting with about 50 workers to explain how the payments would be made. The Intercept reviewed the full video. Payments were supposed to cover the period from February 25 to March 24, but on the day of the meeting, Alarcão warned that she was still waiting for the cash to arrive in her account.
With a spreadsheet open on her computer, Alarcão explained to the transcription workers how to calculate their expected pay and, if necessary, how to dispute the amount they would receive. The document, which was reviewed by The Intercept, logs the amount of time, in seconds, of video content transcribed by each worker. To figure out their pay total, each worker would divide the number of seconds by 3,600, the number of seconds in an hour, Alarcão explained.
According to the document, the transcription worker who received the most money for work in March 2021 transcribed 18,577 seconds of content – just over five hours. Based on the estimate of 20 hours of work to translate one hour of content, the worker received $72 for at least 100 hours of work. It’s unclear if the worker held other employment, but if this was their only source of income, the compensation would be far below the government’s established monthly minimum wage of about $200: In all, the worker earned about $75 that month.
The pay fell well below what was initially advertised by the project’s managerial team. In a long WhatsApp message demanding more productivity from workers, De Rose offered some motivation: “I’m certain that if you knew how to take advantage of this opportunity it would be a great way to get past the global [financial] suffocation that we are currently experiencing.”
There are indications that the document in which managers kept track of workers’ transcription hours seems not to have originated in Brazil. The spreadsheet contains five tabs, four of which are in Chinese. During one videotaped virtual meeting, Alarcão said, with a laugh, “There are a number of tabs here. Everything’s in Chinese, wonderful.” Other documents related to the project were also in Chinese.
“There are a number of tabs here. Everything’s in Chinese, wonderful.”
On WhatsApp chats, Alarcão and De Rose had their own bosses to answer to. A man identified only as “the Turk” pressed them for results. In one message to the transcribers, De Rose translated an English order from the Turk as a means of motivating the workers to be more productive: “Please replace those (workers) who have low transcription numbers or a low percentage of hits with trained people.”
More than once, the company insinuated that workers were not productive enough, and in certain cases, De Rose shared the Turk’s disappointment with her team: “He said he won’t put more people on the team until you produce more,” she said to the WhatsApp group ByteDance 6.0, which Felipe belonged to.
Just over two weeks after The Intercept requested comment from ByteDance on its TikTok transcription processes and policies, Alarcão sent a despairing message to the group of workers: “Unfortunately we are shutting down the project.” She said the “payment source” — an apparent reference to one of the levels of subcontractors in the chain — had refused to make a full payment, instead only paying for what were deemed to be “acceptable” transcriptions. (Roghani, of Transcribe Guru, said payments made to him were based on “client approval” and that, on the Portuguese transcription project, he made all the payments he was responsible for “based on the reports that I received.”)
Alarcão said bonuses for the top 10 transcribers could not be paid. “Tomorrow I will shut down this group because in truth we have not the slightest hope of receiving the bonus payment,” she wrote to the transcription workers on the WhatsApp group. (Roghani said he was unaware of any bonus structure for transcription workers.)
On July 19, the WhatsApp group was shut down.
Though the Portuguese-language project was stumbling, the TikTok transcription project in Brazil was ramping up with other languages. In June, De Rose posted a message to her Facebook page: “Guys!! I’m recruiting people to work with transcription in: SPANISH, ITALIAN and FRENCH!! More info ONLY ON MY WHATSAPP.”
The structure of the work remained the same: Workers were brought together in WhatsApp groups and closely monitored by team managers. This time, however, the project presentation meeting — a video of which was reviewed by The Intercept — was led by a man named Diogo Macedo. “The audio we’re going to transcribe, the audio you’re going to transcribe, is TikTok audio,” he explained. ByteDance, he said, was the parent company: “Their field of work, in addition to other issues, is the field of artificial intelligence. This work we do is to improve voice recognition applications. They collect this material that we send and insert it into their system to improve the applications.”
According to Macedo, the work would serve to “feed artificial intelligence.” He noted that transcription boxes would already have text in them — the AI’s first pass at transcribing the audio — but that the text would be filled with errors. He said that gradually, as the transcriptions were corrected, the AI would improve, something he said had occurred on the Portuguese project as well. “The machine, poor thing, it wasn’t intelligent,” he said of the Portuguese project. “At the end of the project, the machine was even including accents.” (In response to an inquiry from The Intercept, Macedo declined to comment.)
In the Italian transcription project, performance standards were set high. Management in this case, though, was done directly by Roghani. In the WhatsApp group chat with workers, which The Intercept reviewed, Roghani sent the performance spreadsheets and messages demanding more productivity. “Italians still don’t follow instructions,” he said at one point in July. “We’ve talked to them hundreds of times. We talk to them every day. They don’t give shit.”
The group that did the Italian transcriptions was promised better pay — $32 per hour of transcribed audio delivered — than the Portuguese transcribers. Full payments, though, continued to be a problem, one transcription worker said.
Alessandra Zanotto, an Italian teacher living in Brazil who had signed up as a transcription worker for extra income, said she never got paid for some of the work she did. She said all her work for her first two weeks transcribing — more than 5,000 tasks — went unpaid. The pay, she said, should have been around $130.
Zanotto filed a complaint against Alarcão to the Public Labor Prosecution Office, alleging that she did not receive full payment. The Public Labor Prosecution Office declined to investigate Zanotto’s claim because it was outside the office’s scope of responsibility. The agency referred her complaint to a department in the Ministry of Labor and Social Security, which did not immediately respond to a request for comment.
After more than two months of delay, Zanotto redoubled her attempts to get payment. She wrote to the project manager who had recruited her, Narciso, as well as to Alarcão, to no avail. (Roghani confirmed that there was a delay in payments because most of the month’s work had been rejected by the client for “low quality.” He said, “I stopped paying out of my own pocket” — and insisted that he did not owe any more money on the project.)
Zanotto’s last-ditch effort to get paid was to write directly to ByteDance. She sent the email to every address she could find on ByteDance’s website, as well as on TCS, the program used by the company for internal communications. “I know it’s a form of outsourced work and that these people may not be directly related to ByteDance, but we work on a platform of yours,” Zanotto wrote. “Furthermore, they introduce themselves as ByteDancers” — how ByteDance employees refer to themselves — “which damages the company’s image.” She wrote in the email that she was one among many of the transcription workers hired to work for TikTok who did not receive their full, promised payment.
Zanotto never heard back.