Training artificial intelligence systems requires large volumes of text, images, videos and code. Models such as ChatGPT and Gemini use this material to learn patterns, perform tasks and interact with users. However, the supply of high-quality human-created data is not growing as quickly as the models, prompting technology companies to explore increasingly broad ways of obtaining information.
The data may come from the internet, public or licensed databases, books, news reports, scientific articles, commercial partners, specialists, user interactions permitted by service policies and synthetic material generated by other AI systems. The process usually includes pre-training with large collections, fine-tuning with selected datasets, and alignment based on human feedback. Workers evaluate answers, classify content and conduct safety tests, while user prompts, conversations and uploaded files may also be used depending on each company’s settings and policies.
The choices made during training influence what a model can do, the errors it may produce and the biases it may reproduce. For robotics systems, companies also collect first-person recordings made with cameras, mobile phones or wearable devices while workers perform tasks such as folding clothes, washing dishes and organizing objects. Tesla, Figure AI and Agility Robotics are investing in this area, while startups including Micro1 and Objectways recruit people to record tasks. Scale AI has reportedly collected more than 100,000 hours of this type of material.
These recordings can capture homes, bystanders and other details in the environment, creating questions about privacy, storage and the subsequent use of the data. They also create a labor paradox: workers act as human sensors to teach machines to perform tasks that could eventually replace their own jobs.
The shortage of high-quality material has also intensified disputes over copyrighted works. Anthropic downloaded and stored more than 482,000 books obtained from pirate digital libraries to train Claude, leading to a reported $1.5 billion settlement. In the Kadrey v. Meta Platforms case, Meta used books by authors including Sarah Silverman and Richard Kadrey from the LibGen pirate library to train LLaMA. According to attorney Diogo Coletta, the company reportedly stopped licensing negotiations after discovering that the pirate library contained many of the titles under discussion with publishers.
Attorney Daniel Failla says that an innovative purpose does not by itself make the way data was obtained lawful. Technical access, he argues, should not be confused with legal authorization.
Human labor remains another essential part of the AI supply chain. So-called ghost workers classify images, assess responses and moderate content. Investigations cited in the original report found that workers in India reviewed extreme violence and abuse for long periods, while contractors involved in Google Gemini testing in the United States reported exposure to disturbing material without adequate warning, training or psychological support.
Productivity targets can add to the pressure. Some testers and data labelers work with 10- to 15-minute timers for complex tasks and quotas of up to 12 reviews per hour. The reported consequences include anxiety attacks and panic attacks. Highly qualified professionals, including people with master’s and doctoral degrees in journalism, law and physics, have also turned to AI training work as temporary or alternative employment, often for lower pay and without benefits or stability.
In Brazil, the microtask market includes labeling images for autonomous-vehicle systems, reviewing medical exams and recording repetitive phrases for virtual assistants such as Alexa. The study Microtrabalho no Brasil estimates that AI trainers can earn an average of between R$538 and R$583 per month, or approximately R$9.36 per hour. The amount is less than half of Brazil’s minimum wage. Professor Dora Kaufman of PUC-SP describes this structure as “Digital Taylorism,” in which workers perform small parts of a task without necessarily knowing the larger system they support.
Companies have also been accused of trying to bypass technical restrictions imposed by websites. Cloudflare identified what it described as hidden crawlers used by Perplexity, along with changing IP addresses to appear as ordinary browsers. The method was allegedly intended to circumvent robots.txt directives and blocks applied by sites that did not authorize automated collection. Separately, concerns have emerged over the collection of personal and biometric information, including social-media profiles and data that may expose connections between individuals and information about minors.
The search for human-created material has extended to old and out-of-print books. An investigation by 404 Media reported that Amazon operates a facility in Las Vegas where rare, old and out-of-catalog books are purchased, sometimes through intermediaries. The books are reportedly cut and scanned at high speed. Anthropic documents presented in court also referred to Project Panama, an initiative aimed at destructively scanning books. After the pages are digitized, the remaining material is sent for recycling.
The practice is linked to concerns that excessive use of AI-generated content could degrade future models. Human-created material produced before the widespread adoption of generative AI is therefore considered valuable for maintaining training quality. In Bartz v. Anthropic, a United States court found that legally purchasing a physical book and destroying it after creating a single digital copy for training may qualify as fair use, depending on the circumstances.
Book dealers and historians have warned that digitization cannot reproduce the historical value of some physical works. Marginal notes, handmade bindings and other features may be lost, while the digital files remain in private corporate databases.
Companies continue to pursue large and varied data sources because computing power cannot compensate for poor, incomplete or outdated information. This is especially important for specialized systems used in medicine, law, finance and industry, where inaccurate data can contribute to flawed decisions. At the same time, high-quality information may be protected by confidentiality, contracts, copyright or regulatory requirements.
The economic incentives are also significant. Computing capacity can be purchased, but large collections of high-quality human text are limited and cannot be expanded as quickly as demand. Licensing every work from publishers, record companies, newspapers and authors can be expensive and slow. According to Coletta, some companies may therefore judge it economically preferable to move ahead and deal with legal disputes later. The uncertainty is reinforced by the lack of definitive decisions on several issues and by differences between countries, while authors may struggle to prove that their works were included in a model’s training data.
Sources: Canaltech.