Publications
Partners
- AI often relies on large human workforces to be able to operate.
- Workers often perform tasks which can include moderating harmful and explicit content, and can be psychologically damaging to them.
- For more stories, visit the Tech and Trends homepage.
We are missing a big part of the AI story and that is the fact that it often relies on a large human workforce.
This is what Sonam Jindal, the programme lead for AI, labour, and the economy at the Partnership on AI, a US-based artificial intelligence research institute, told NBC News earlier this month.
She said that a lot of the discussion around AI is very congratulatory, but a Partnership on AI web post said that the "data enrichment work" required to make machine learning tools function optimally was "often overlooked, [and] performed by faraway workers under precarious conditions".
Humans provide feedback to AI tools which is used to provide better and more accurate information for users of the service.
Humans perform tasks such as labeling what they see in an image or providing feedback on whether certain content is explicit or unsafe, which the AI can take on board and use to provide safer and more accurate responses in the future.
Without some human input, it is very difficult for AI tools to know whether the responses it is providing are accurate or whether the material is potentially harmful.
So big technology companies often reach out to firms that have a large army of workers available to train AI models.
Training ChatGPT
One of the organisations involved in data enrichment work is Sama, a US-based machine learning training institution that hires workers from Kenya, Uganda and India to label data for clients like Google, Meta and Microsoft.
An investigation by Time that was published earlier this year found that OpenAI, the owners of ChatGPT, also contracted Sama to train ChatGPT, and that workers received between $1.32 and $2 per hour for the work.
OpenAI does not disclose the names of its data enrichment partners so it is not known if it worked with other data enrichment firms to train ChatGPT.
OpenAI contracted Sama in 2021 to label textual descriptions of sexual abuse, hate speech, and violence. Time spoke to four Sama employees, who wished to remain anonymous, all of whom reported that they were "mentally scarred" by the work.
Mental health sessions were made available to the employees by Sama, but the employees that Time spoke to said they were "unhelpful" and "rare".
A Sama spokesperson told the publication that "professionally-trained and licenced mental health therapists" were available any time for group or individual sessions for the employees.
Sama canceled its contract with ChatGPT in February 2022, eight months earlier than the contract stipulated, owing to the traumatic nature of the work.
An OpenAI spokesperson said in a statement to Time that OpenAI's mission was to "ensure artificial general intelligence benefits all of humanity".
READ MORE | OpenAI chief goes before US Congress to propose licenses for building AI
The statement said that the company worked hard to ensure that AI systems were useful and safe and that bias and harmful content was limited.
"Classifying and filtering harmful [text and images] is a necessary step in minimising the amount of violent and sexual content included in training data and creating tools that can detect harmful content."
Time said that the labour, which is often sourced from the global south, required for AI to function safely could often be "damaging and exploitative".
"These invisible workers remain on the margins even as their work contributes to billion-dollar industries."
Amazon Mechanical Turk
Another tool that can be used to train AI is Amazon’s Mechanical Turk.
The crowdsourcing marketplace allows businesses to capture data from an international workforce, which Amazon says is 500 000 people strong.
Companies post what are known as "human intelligence tasks" that freelance workers, known as turkers, compete to complete. These tasks include labeling photographs, completing surveys, or moderating content.
Some tasks only pay $0.01 or about R0.19.
A study cited by the New York Times found that the median "turkers" hourly wage was $1.77. Only 4% of turkers earned above minimum wage.
Mechanical Turk can be used for tasks that contain adult content, but that content must be labeled as such and cannot contain illegal material.
Amazon said that crowdsourcing workers to complete microtasks was better than the traditional ways of completing work like this, which required hiring a temporary workforce.
"While technology continues to improve, there are still many things that human beings can do much more effectively than computers, such as moderating content, performing data deduplication, or research," said the company.
