They sort through humanity's horrors so we don’t have to
The human cost of LLMs
Imagine sifting through the darkest corners of the internet every day. This was the reality for Facebook content moderators, whose job it was to look at and label graphic violence and hate speech, leading to severe mental health issues like PTSD. In 2020, over 11,000 US Facebook content moderators (former employees and contractors from Accenture, Cognizant, Genpact, and ProUnlimited) filed a class-action lawsuit against Facebook, collectively winning $52 million to seek treatment for PTSD. As AI platforms consume vast amounts of internet content, a similar problem emerges.
The role of data annotation
The development of large language models (LLMs) relies heavily on data annotation, where human workers label vast amounts of data to train AI systems. This work is crucial for teaching AI models to recognize patterns and generate accurate responses. Tasks range from labelling images for self-driving cars to categorizing text snippets for language models. Without human inputs, LLMs wouldn't be able to perform at the level we expect.
For instance, when using an image generation AI called Leonardo, I asked it to create an image of a mystical tarot reader. It produced three similar cartoonish images of a scantily clad woman, with one labeled as sexually explicit. Two of the images were deemed by the AI as fine for a child to see, and one of them was labelled as sexually explicit. They all seem equally explicit to me.
This is where data annotators come in, exercising judgment to teach AI where to draw the line.
Working conditions and mental health impact
In 2021, Kenyan workers contracted by San Francisco firm Sama were tasked with labelling toxic content, a job involving graphic descriptions of violence, hate speech, and sexual abuse. These workers, paid between $1.32 and $2 per hour, faced significant psychological strain, resulting in PTSD symptoms for many.
Despite the mental toll, the support systems in place at companies like Sama are often inadequate. The high productivity demands and pressure to meet performance targets further exacerbate the situation, leaving workers without the necessary support to cope with their jobs' emotional strain.
Economic disparities and wages
The disparity in wages paid to data annotators is stark. While AI companies in Silicon Valley rake in billions, workers in the global south who label data often earn less than $2 per hour. Companies like Sama (and by association, OpenAI) are clearly leveraging low wages in developing countries to maximize profits. Cost of living adjustments determining these wages often fail to consider the true value of the labour provided and the impact the work has on the people doing it.
Testimonials from workers reveal the harsh economic realities they face. Despite being integral to the development of cutting-edge AI, many data annotators struggle to make ends meet. As one Kenyan worker put it in Josh Dzeiza’s 2023 piece for The Verge, “I really am wasting my life here if I made somebody a billionaire and I’m earning a couple of bucks a week”.
The precarious nature of these jobs adds another layer of concern. Data annotators face uncertainty and inconsistency in their work, with projects often ending abruptly. As AI continues to evolve, the future of these jobs remains uncertain. Will AI eventually automate these roles, or will there always be a need for human oversight?
Ethical implications and responsibilities
The ethical responsibilities of AI companies toward their workers are significant. Outsourcing labour to low-wage countries raises serious moral questions about fairness and exploitation. Companies need to consider the human cost of their technological advancements and take steps to ensure their practices are ethical. AI companies must implement fair wages in developing economies and provide robust mental health support for workers dealing with troubling content.
Greater transparency and recognition of the labour force behind AI are crucial. By acknowledging the contributions of data annotators and ensuring they’re treated fairly, companies can take meaningful steps toward addressing the human cost of AI.
Weekly disruptions
Apple, Nvidia, Anthropic Used Thousands of Swiped YouTube Videos to Train AI (Proof News) An investigation by Proof News reveals that major AI companies, including Apple, Nvidia, and Anthropic, have used transcripts from thousands of YouTube videos to train their AI models without creators' consent, violating YouTube's policies. The dataset has sparked controversy over the unauthorized use of creators' content, raising concerns about compensation and potential misuse.
Tinder’s new AI tool will curate your dating profile pictures for you (CNBC) Tinder is rolling out an AI-powered tool, Photo Selector, to help users choose the best photos for their dating profiles. By taking a selfie and granting access to their smartphone photos, users receive recommendations for up to 27 images designed to make a great first impression. According to Tinder, 68% of users find this feature helpful.
The ‘godmother of AI’ has a new startup already worth $1 billion (The Verge) Fei-Fei Li, known as the "godmother of AI," has launched World Labs, a startup now valued at over $1 billion just four months in. World Labs aims to develop AI with advanced reasoning capabilities through human-like visual data processing. The venture could revolutionize fields like robotics, AR, and VR.