Kenyan employees denounce the “torture” of their work to reduce the toxicity of AI

Warning, this article contains explicit descriptions of violence, rape and sexual assault.

The rise of artificial intelligence (AI) and especially their new, unprecedented accessibility to the general public have had a lasting impact on the year 2022. Among the tenors of the genre, it is the OpenAI company that has made the most noise. It was recently valued at $29 billion and could soon receive a $10 billion investment from giant Microsoft.

Two tools developed by OpenAI stood out. First, Dall-E, which is able to generate images from descriptions, and more recently, ChatGPT, a text generation tool whose “infinite” possibilities frighten as much as they fascinate. From a whimsical and personalized cooking recipe to the rewriting of miserable as if sung by Jul, going through student essays, nothing seems to stop the beast. During its launch week last November, ChatGPT had already amassed a million users.

Before ChatGPT, there was a beta version called GPT-3. The latter, still incomplete, had in particular shown limits and had shown itself capable of explicitly expressing acts of violence. In 2020, the French start-up Nala used the GPT-3 model (against OpenAI’s advice) to design its own medical chatbot. During the test phase, the latter had “advised” a patient to commit suicide.

See also on Konbini

Even today, this is one of the greatest limits of artificial intelligence: its mimicry of human behavior and reactions is based on what it “observes” on the Internet. Unfortunately, the Web is not known to be a perfect safe place. Thus, the violence of the words and actions found there can be integrated into the responses of the AIs.

This is precisely the promise of ChatGPT: a total absence of toxicity or explicit violence in the texts generated. It is clear that the OpenAI tool has won its bet and has not repeated the errors of its predecessors. Except that these virtues do not develop in a snap of the fingers, they have a financial cost but especially human.

Behind the AIs, there is still humans

To fight against toxic and violent content with artificial intelligence, there are not 1,000 solutions for the moment: you have to use the human mind. OpenAI has followed the example of social networking giants, such as Facebook, before it by relying on “data annotation”. The AIs thus receive images, videos, texts containing violent descriptions so that they can understand what must be labeled as such – and therefore prohibited.

A major survey of Time thus looked at the OpenAI partners, responsible for identifying dangerous content. To obtain this data annotation, OpenAI subcontracted this task to Sama, a firm based in San Francisco but whose workers are in Kenya. Created in 2021, this “ethical AI” company has already worked for major clients such as Google, Meta and Microsoft.

Several tens of thousands of texts have thus been sent to Sama so that they can be read, analyzed and annotated by its Kenyan employees. Some of them were particularly explicit in graphic detail on subjects such as child sexual abuse, murder, suicide, torture, self-harm or even incest.

The Time was able to access hundreds of pages of internal Sama and OpenAI documents, including employee payslips. Four of them were also interviewed, on condition of anonymity. Three say they had to read and annotate between 150 and 250 passages of text per nine-hour shift, excerpts ranging from around 100 words to more than 1,000.

The four employees interviewed by the Time said they had been mentally affected by their work. One Sama employee in particular explains that he suffered from “recurring visions after reading a graphic depiction of a man having sex with a dog in the presence of a young child”.

“It was torture […]. You read a number of statements like that throughout the week. By the time Friday rolls around, you are disturbed thinking back to those images.”

An OpenAI spokesperson told the Time that psychological counselors were present for any Sama employee who requested them. In fact, no individual interview seems to have been conducted, which OpenAI finds “unacceptable” returning the ball to his subcontractor.

The traumatic nature of the work ultimately led Sama to cancel all of his work for OpenAI in February 2022, eight months ahead of schedule.

A prematurely terminated contract

According to documents recovered by the TimeOpenAI has signed three contracts with Sama for a value of $200,000. At the end of the day, the sixty Kenyan employees only received an hourly wage oscillating between 1.32 and 2 dollars net, knowing that this maximum value could only be reached on the condition of a certain number of objectives. quantities completed. The majority of employees on this mission thus received a monthly salary of 21,000 Kenyan shillings (about 160 euros), knowing that the minimum wage is currently 15,120 shillings – about 124 euros.

From February 2022, Sama reportedly began to gradually move away from OpenAI. In a few weeks, the company would have broken several contracts, and this, eight months before the term and the delivery. In a statement, the outsourcing company said the agreement for “image collection” for OpenAI did not contain any reference to illegal content. It was only after signing that OpenAI would have requested the collection of images with illegal content – ​​in particular child pornography. A Sama spokesperson told Time:

“Sama immediately terminated the image classification project and announced that she would cancel all [projets] remaining with OpenAI.”

In a statement, OpenAI confirmed that it had received 1,400 images from Sama that included all possible categories of explicit violence, but said it never wanted employees to have to go looking for images with child pornography content – ​​and therefore illegal. OpenAI points a “communication problems” with his former partner.

On January 10, Sama went further by announcing that it was canceling all of its projects including sensitive content. The company would also not renew its content moderation contract (up to 3.9 million dollars) with Facebook, resulting in the loss of some 200 jobs in Nairobi.

“After many discussions with our team globally, Sama has made the strategic decision to leave all work [de traitement du langage naturel] and content moderation to focus on computer vision data annotation solutions.”

However, whether it’s Sama or others, this need for sensitive data annotation remains essential for all AI-based tools. The human and psychological cost of the small hands behind the tools of the big tech giants nevertheless remains marginalized on the eve of major multi-billion fundraisers.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *