ChatGPT could steal social network moderator’s job – 10/03/2023 – Tech

The chief executive of OpenAI (the company behind ChatGPT), Sam Altman, used to respond that content moderators on social networks would be the first to be replaced by artificial intelligence, when asked about the jobs the technology would take.

He spoke of workers who delete violent publications or publications about murders, pedophilia and other absurdities from the so-called “sewer of social networks”. This service was voted as the worst online gig by Brazilian freelance workers interviewed in a survey by DiPLab (a European laboratory that studies work on the internet). In the industry, cases of post-traumatic stress are frequent due to the strong content of the images, texts and audios that need to be reported.

The speculation proved to be a plan in August, when OpenAI announced that ChatGPT, or rather the GPT-4 language model behind the internet’s most famous chatbot, was capable of moderating content.

Today, in general, this work is divided into two parts: on the one hand, experts in ethics and law put together complex moderation protocols based on samples of behavior on social networks; on the other, workers, especially from the global south, carry out the steps elaborated by the first group, in direct contact with the most absurd publications on the internet.

Frontline moderators apply for these positions on remote work management platforms, which work like an Uber of digital jobs — filling out spreadsheets, transcribing audio, translating texts and reporting abusive publications.

As this activity does not require physical presence, people from countries with cheap currencies can perform these jobs, which makes the service cheaper for big techs based in the United States. Preferred candidates come from English-speaking countries with low per capita income, such as India, Nigeria, Kenya and Pakistan.

The research by Latraps (a European laboratory focused on research on work on the internet) “Who are the workers behind artificial intelligence?” shows that there are Brazilians moderating content and demonstrating dissatisfaction. The study was conducted in Brazil by the professor of work psychology at UEMG (State University of Minas Gerais), Matheus Viana Braz.

One of these workers found by the report was Paulo (pseudonym), 24, who has been evaluating since 2019 whether publications on Facebook, Instagram and other social networks are in accordance with the platforms’ policies. “I’ve seen posts with obscene content, videos and photos of accidents, murders and even pedophilia,” he says.

He asked to have his identity preserved due to a confidentiality agreement he signed when registering with Appen, a website that mediates, among other things, between content moderators and the social networks that hire them. Paulo cannot talk about the services he performs or the projects he is on.

On condition of anonymity, Paulo says the work was disgusting. “However, seeing that it was about reporting and banning the post or the user from the platforms, I was relieved to have seen those things.”

Additionally, he uses the extra income at Appen to secure his early-stage career in film production. He graduated in 2022 from the audiovisual course at a higher education institution in São Paulo.

Among workers on Brazilian online gig platforms, the proportion of people with completed higher education is above the average for the Brazilian population: 43% versus 24%.

This activity at the end would be the part carried out by GPT-4. Artificial intelligence is capable of interpreting rules and nuances during the moderation exercise, and readily adapting to changes in norms. This, according to a publication on the OpenAI blog, which denied interview requests from Sheet.

Even though it is critical to automate human work without compensation, Matheus Viana Braz states that this application of artificial intelligence can bring benefits. “If generative AI, like ChatGPT, reduces this demand, even in its initial phase, we will have a considerable advance.”

OpenAI itself, however, states that GPT-4 will not remove everyone from the content moderation production line. “Like any and all AI applications, human monitoring and verification of the entire process will be required.”

Ronaldo Lemos, columnist for Sheet and chief scientist at the Rio de Janeiro Institute of Technology and Society (ITS-Rio), goes further: “artificial intelligence will not replace content moderation but rather create more demand for these services.”

Lemos remembers that OpenAI itself has hired people in various parts of the world, African countries and Brazil included, to classify, train content, to look for responses that are infringing on issues of rights violations and so on.

“These people hired by AI companies are not called content moderators, but they basically do the same work that platforms do in search of content that violates rights on their platforms,” says Lemos. OpenAI has outsourced this work to Kenya, for example, with pay of less than $2 per hour (low by American standards but above the local minimum wage).

“AI requires gigantic human work. It needs to be trained, retrained, in search of moderation of its own content. So, it seems to me that the premise [de Sam Altman]again, is wrong”, says Lemos.

On the other hand, companies like Appen itself, which hires the services of audiovisual producer Paulo, says on its blog that it uses artificial intelligence to speed up moderation work on social networks. However, like OpenAI, the freelancer platform reiterates the importance of keeping humans in the verification link.

Appen was also approached by Sheet by email since the 27th and did not respond to the report’s messages.

The demand for this service, in fact, is older than ChatGPT itself. Research from the Oxford Internet Institute shows that remote freelance platforms had 163 million users in 2021.

The word corrector on cell phones and computer programs, for example, is trained from people who fill out diagrams with all the possible errors linked to a word, called a “rhizome” due to its branching format like a root.

In this sense, artificial intelligence is used to boost the recruitment of content moderators with advertising production, says Italian sociologist Antonio Cassili, from the Paris Polytechnic Institute.

As shown by Sheet, online freelance sites pay, on average, around R$9 per hour in Brazil and require candidates to undergo unpaid training and communicate in English. Often there is no offer to work 40 hours a month.

These platforms advertise themselves as forms of extra income, but 33.5% of those interviewed in the Latraps survey in Brazil say they have no other source of resources.

“GPT-4 is a robot that automates the hiring of precarious workers who perform the same tasks as higher-paid humans,” says Cassili.

For Ronaldo Lemos, the treatment of “social media waste” is an area that will grow more and more, even due to legal requirements. “For example, the DSA, European legislation that inspires the Brazilian fake news PL, requires that there be initiatives on the part of platforms in the area of ​​content moderation.”