Work without the worker - Phil Jones
- Victor Hugo Germano
- Jan 29
- 6 min read
During a Computer Vision activity in my master's degree, we were discussing how we would annotate images to be used in a pattern recognition model. To my surprise, a colleague seemed to not believe that we should manually annotate all the images so that we could train the model.
"Obviously, big companies do this automatically, it doesn't make sense to do it manually, it's a huge waste of time!"
With a little study, all the utopian appeal of using autonomous tools disappears, and we gradually realize what is behind the advances: The magic of machine learning is the human rush of data classification.
The dream of General Artificial Intelligence hides an army of forgotten workers, in precarious conditions of near survival, who become dependent on platforms with the sole objective of extracting the most from them.
In addition to being a great way to speculate with startups: Create a company posing as AI, when in reality, through the availability of an API, humans perform all the work. Waving to the Venture Capital market's desire for AI solutions when in fact everything is done through Microwork. Amazon itself defended for years that stores with Walk Out technology were all automated using AI, when in fact more than 70% of the operations were confirmed by humans !
The great advances in Generative AI and Deep Learning in the world come largely from the work of Fei Fei Li , a Princeton professor, who built some of the first image datasets with enough annotations to boost the performance of the algorithms. How? By hiring the Amazon Mechanical Turk service, which "employs" people to perform atomic tasks, to map the 14 million images from ImageNet , the dataset that sparked the AI race this summer.
Machine Learning and Deep Learning depend on this annotation data to function, and mainly need new data so that the models can be re-trained as they encounter new scenarios: whether it is image recognition or text generation, it is necessary to constantly update the training information. Who are the best agents to indicate information to the algorithm: us, humans.
“Behind search engines, apps, and smart devices are workers, often those banished to the margins of our global system, who, for lack of other options, are forced to clean data and oversee algorithms for little more than a few pennies. Facebook and Twitter feeds may appear to weed out violent content with automated precision, but decisions about what constitutes pornography or hate speech are not made by algorithms. A facial recognition camera seems to willingly pick out a face in a crowd, a self-driving truck seems to drive without human involvement. But in reality, the magic of machine learning is the hard work of data labeling.”
The only way AI works is by not paying the real cost of producing and operating the technology. Phil Jones writes a tough book that takes a long time to digest. The Work without the Worker is a necessary book to understand the current moment.
This is a quick but intense book, and it asks us whether it is really worth using AI platforms, while tens of millions of precarious people, without enough to survive, are exposed to a new type of exploitation: Microwork . Snatched around the world by companies whose speech is of social transformation, when in fact we are just facing another stage in the long downfall of capitalism.
The central idea of the book revolves around Microwork, and how platforms exploit people in vulnerable situations - to perform digital tasks for mere pennies. Microwork is an activity that is hard to believe, while we are bombarded with dreams of a utopian future and abundance brought about by artificial intelligence and robotics.
“Microwork represents not the phoenix of the South, but a new twist in our planetary labor crisis. Microwork is the sum of the same processes of slow growth, proletarianization, and declining labor demand that have inflated the informal sectors of countries like India, Venezuela, and Kenya.”
What began as Mechanical Turk on AWS is now a form of automation and distribution of tasks among individuals around the world in a process called arbitration. None of these individuals have much say in the type of work that gets done, or any ability to negotiate better working conditions. In the most informal way, Data Annotators are recruited by companies like Scale and Mighty AI, mostly in places ravaged by war and poverty, not despite their desperate circumstances, but because of them.
The services are accessed using APIs and can be integrated into systems using code: Workers as computing infrastructure. It is currently unclear exactly how many people work for these companies, and under what conditions, but it is estimated that hundreds of millions of people spend their days performing all sorts of tasks. The Chinese platform Zhubajie alone has 12 million registered workers – making it by far the largest employer in the world.
Microwork comes without any rights, security or routine, paying a pittance - just enough to keep the person alive and socially paralyzed.
This is the biggest secret of the AI market: it is not known exactly how many people work in this market, nor how many companies in total. Everything is governed by NDAs by the main contractors. Everything is done to distance workers from the result of the work: distributed, fragmented and without guarantees, making it even easier to reduce working conditions.
Being paid by task, it is common for these people to spend a lot of time without payment while looking for new tasks, in the same way that Uber drivers need to work more than 14 hours a day to earn a living from the platform and cover their personal demands. The population pushed out of the formal job market finds itself forced into this type of work on the edge of survival, earning less than $2/hour, when they have a task to perform.
Platforms decide which task is delivered, how long it should take (some in as little as 30 seconds), what payment will be made and whether the note-taker will be paid . It is common for platforms to pay in Amazon gift cards or cryptocurrencies! The total informality of the employment relationship and global arbitration create an ideal space for payment reduction and abuse.
Obviously, the number of people on these platforms reduces the pay per task, which is the business model of these companies: crowds of people desperate to keep working, accepting a few cents to ensure survival in shifts of up to 80 hours per week. This is absolute digital precariousness, the sweat shops of today.
“This is a significant but under-reported feature of platform capitalism: the workers who turn vast amounts of data into the valuable information that sustains the system are paid only in the loosest sense of the word. Microwork sites allow large platforms to hide this reality, or at least make it seem acceptable. The workforces of Google and Microsoft exist behind a marketing mirage that sustains the idea that microwork is not really work, and the microworker is not really a worker. Requesters are often given all the work without having to pay for it. Only to those who actually do the work does the sentiment ring hollow: getting paid for not doing much really means doing a lot for not getting paid.”
Cloud Capital and Platform Capital have created “Artificial AI” (Jeff Bezos’s term) – or human labor, designed to destroy the bargaining power of employees by atomizing labor into tasks that are distributed across a network of workers. All these people see is a task, all the relationship between the company and the person is done through the platform, and all the payment is based on execution – welcome to the high-tech version of Chaplin’s screw tightener – Microwork is the warped child of late capitalism that takes everything from the human being for the demands of the algorithm.
The web of relationships between Taskers (or Annotators, or Microworkers) and the platforms that route tasks is impressive:
A company uses Mechanical Turk to perform a task to train its own AI
Amazon collects data on task execution, your actions, and content performance to determine a job profile.
The annotated data is used to train AWS’s own services. Additionally, the execution information is used to establish new performance benchmarks for workers in Amazon’s warehouses.
Smaller platforms act as brokers. Playment hires microworkers and has agreements with Facebook to deliver workers’ data, for example.
Playment itself accesses this information to identify in the worker's network of contacts who are the most likely people to also be hired as Taskers.
Workers' behavioral information is also shared with Facebook, which is interested in expanding its network of influence.
This worker is exploited on several fronts: his ability to perform, his personal data and network of connections, his behavior during the execution of the task, and in his choice of what type of task can or cannot be paid.
This behavior and work data is then used for the platform’s own performance algorithms, increasing the platforms’ ability to reduce payout and arbitrage tasks to higher ROI regions.
The reality and future presented by the author are quite shocking. The solutions are few.
A necessary book. I recommend it.
A new social polarization is emerging – probably already happening here – between those with a single stable career, and those forced to walk dogs in the morning, clean houses in the afternoon and act as a hired friend in the evening, before searching for online tasks in the evening.
댓글