A TIME investigation reveals that OpenAI used a contractor, Sama, to add a filter to ChatGPT — to prevent it from producing violent responses. But, the working conditions of the people hired by Sama were traumatic, given the content they had to moderate.
ChatGPT impresses as a technological gem and has been in the news for several months now. But behind this success hides a much less rosy side. Extensive TIME magazine survey, published on January 18, 2023, reveals how OpenAI, the company behind the program, recruited Kenyan workers paid $2 an hour via a subcontractor. In working conditions, what is more, traumatic.
If OpenAI called on these workers, it was to make ChatGPT less “ toxic », That is to say, to avoid that he can answer you with racist, homophobic, violent texts. Their role was to moderate this risk upstream. But at what cost ?
Data from “the darkest corners of the internet” to moderate ChatGPT
If ChatGPT’s artificial intelligence stands out, it’s thanks to the GPT-3.5 model. Based on an artificial neural network, it contains hundreds of millions of parameters. Its training is so powerful that, taking the form of a chatbot with which one can converse, it effectively answers a multitude of questions, in a language that is often quite natural. The program is designed as a service interface, to access user requests.
But, to get there, GPT-3 was trained from massive data mined… from the internet. For better or for worse: the web is not always a very healthy place and humanity does not always reveal its best side there. Result, GPT-3 had initially integrated racist, sexist, homophobic biases. The algorithm could provide, at one time, violent and discriminating answers. However, it is apparently difficult at this stage for OpenAI to eliminate this prior toxic data from the database. It is therefore downstream, rather than upstream, that the company has implemented a mechanism to limit the toxicity of its algorithm.
The security system developed by OpenAI consists of adding a filter to ChatGPT, teaching it to identify and label problematic content. Clearly, ChatGPT is exposed to these origin biases, but is supposed to be able to “block” them upstream so that they don’t get to the user. This is, in part, what some social networks do, like Facebook.
As TIME relates, OpenAI called on a subcontracting company called Sama. Very mobilized in Silicon Valley, it specializes in the labeling of sensitive data and employs people in Kenya, India and Uganda. For ChatGPT labels, workers were in charge of dealing with tens of thousands of snippets of text from ” from the darkest corners of the internet “. Among these contents, with very graphic descriptions, we found ” child sexual abuse, bestiality, murder, suicide, torture, self-harm and incest “.
“It was torture”
OpenAI and Sama have awarded three contracts worth a total of $200,000. In its center in Kenya, the company has set up three groups, of a dozen people each, divided into three main themes: sexual abuse, hate speech, violence. These employees were paid between 1.32 and 2 dollars per hour for this labeling. They had to analyze about 150 to 250 snippets of text, 100 to 1,000 words per snippet, per nine-hour day.
Four employees, with anonymized testimony, tell TIME about a job of a traumatic nature. A man explains to the magazine that the graphic description of a certain scene, particularly atrocious, led him to suffer from recurring nightmares durably. ” It was torture “, he describes. Employees had access to “wellness” therapists to help them; but these sessions were “rare”, were limited to group sessions and did not compensate for the productivity requirement otherwise. A witness even indicates that he was repeatedly refused an individual interview with one of these therapists.
Sama partially denied the accusations of these testimonies to TIME. The company indicates for its part that the therapeutic sessions were individualized and could be requested at any time; that the number of texts to be labeled was 70 per day and not 150 to 250; and that pay ranged from $1.46 to $3.74 per hour.
Nevertheless, the TIME investigation reveals that the ” traumatic nature of the work eventually led Sama to end his OpenAI collaboration in February 2022, “ eight months ahead of schedule “.
“A necessary step to reduce the amount of violent and sexual content” according to OpenAI
OpenAI also provided an answer from its side, first confirming having used Sama to make ChatGPT less toxic: “(…) We work hard to build safe and helpful AI systems that limit bias and harmful content “. According to the spokesperson, classification and filtering of [textes et images] harmful is a necessary step to reduce the amount of violent and sexual content included in training data and create tools that can detect harmful content “. As for working conditions, OpenAI says it did not indicate productivity targets and that all management fell to Sama as a subcontractor.
OpenAI adds: “ We take the mental health of our employees and that of our contractors very seriously. We understood that [chez Sama] wellness programs and individual counseling were offered, that workers could refuse certain tasks without being penalized, that exposure to explicit content was limited, and that sensitive information was handled by specifically trained workers. »
TIME also highlights the difficulty of Sama employees in labeling certain ambiguous content. The survey as a whole therefore points to the way in which the issue of toxic biases, within AI, resulting from incorporated massive data, poses serious problems for managing them.
Subscribe to Numerama on Google News to not miss any news!
