PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteSome links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
AI depends on people as well as software. Workers collect and label data, transcribe audio, review content, assess AI-generated answers and give feedback that helps systems improve. Much of this work is routed through platforms, outsourcing firms and subcontractors, making it difficult to see who performs it, under what conditions, and for which technology companies. The evidence shows a varied workforce—not one job, one wage or one region—and does not establish a reliable global headcount.
What work do people do to build and operate AI?
People behind AI perform a range of tasks, from preparing training data to checking systems after deployment. The International Labour Organization (ILO) describes workers labeling, categorizing and annotating data for applications such as object recognition and text prediction. They may transcribe audio for virtual assistants, check whether a system understood a request, or help refine large language model responses and flag bias or harmful content.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass – An Urgent... | $15.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The OECD also identifies data classification, transcription, moderation, output verification and other small digital tasks as work that can support training or operating AI systems. These jobs are not interchangeable: a worker labeling images has a different assignment from a content moderator reviewing abuse or an evaluator judging a generated answer. Skills, working conditions and exposure to harmful material vary accordingly.
| Type of work | What the task can involve | Why it matters |
|---|---|---|
| Data collection and preparation | Gathering, labeling, classifying or annotating images, text and other data | Organizes examples that systems use in training or operation |
| Transcription and verification | Writing down audio or checking whether a system interpreted an input correctly | Helps develop and assess speech and assistant functions |
| Evaluation and feedback | Judging AI responses, identifying errors or harmful content, and helping refine outputs | Provides human feedback on system behavior |
| Content moderation | Reviewing material that may include violence, hate speech or exploitation | Can help identify harmful content, but may expose workers to disturbing material |
How does outsourcing make this labor hard to see?
Technology companies may obtain data work through digital platforms, business-process outsourcing (BPO) firms and subcontractors rather than hiring every worker directly. A task can pass through several organizations before reaching the person doing it. That fragmented chain can make it harder for workers, the public and sometimes clients to identify who is responsible for pay, protections, working conditions or support.
A 2026 report from Germany’s development agency, GIZ, describes opaque and fragmented supply chains in parts of South Asia, Southeast Asia and Sub-Saharan Africa, including informal arrangements and short or micro-contracts. It also discusses algorithmic management, such as real-time monitoring of keystrokes, errors and productivity. The findings describe conditions in the settings covered; they should not be taken to mean every worker or contract is alike.
In April 2025, Rest of World reported that the African Content Moderators Union and Personaldata.io mapped worker networks in 39 African countries and connections through outsourcing intermediaries to technology clients. That map is evidence of a particular network, not a census of all AI labor or a complete inventory of clients. The reporting notes that contracts may limit what intermediaries disclose. Sociologist Antonio Casilli, quoted in the article and not involved in the worker-network mapping it covered, described the result as “a circle of invisibility around this work.” The phrase points to limited visibility and disclosure, not to a lack of worker skill or agency.
What do the studies say about pay and job security?
There is no single representative wage for people doing AI-related data work. The ILO reported median earnings of about US$2 per hour among surveyed crowdworkers in developing countries. That figure belongs to the survey population; it is not a global estimate for every data annotator, evaluator or moderator. The ILO also reports limited social protection and risks of automated work rejection without clear explanations or a grievance channel.
Equidem’s 2025 investigation shows why broad averages can conceal different experiences. Equidem’s 2025 investigation interviewed 113 workers in Colombia, Ghana, Kenya and the Philippines, including workers at BPO firms subcontracted by Meta, OpenAI and ByteDance as well as workers serving other technology companies. Equidem explicitly says the sample is not statistically representative of the global data-enrichment workforce. It offers qualitative accounts of worker conditions, not a basis for estimating how common any one experience is worldwide.
In the United States, a 2023 study by the Communications Workers of America (CWA) and TechEquity surveyed 160 data workers and conducted 15 in-depth interviews. It reported financial worries, reliance on public assistance among some respondents and time spent waiting for tasks to become available. These findings complicate a simple North-versus-South account of insecurity, but the study’s sample should not be treated as representative of all U.S. data workers.
To understand a particular job, ask how it pays—hourly or per task—and whether waiting time, rejected work or bonuses affect take-home earnings. Also check whether a worker can challenge a rejection, how tasks are assigned and monitored, what the written terms say, and whether a grievance route or collective representation exists. The available evidence supports these as important comparison points; it does not provide a universal contract or pay model.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWhat are the health risks for workers reviewing sensitive material?
Some roles involve repeated exposure to disturbing content; many others do not. The ILO warns that workers moderating or training systems may encounter graphic violence, hate speech and child exploitation, with potential mental-health consequences. GIZ describes risks including anxiety, depression, insomnia and post-traumatic stress symptoms, and says trauma-informed occupational support is limited in the settings it examines.
A 2023 TIME report documented one specific case involving Kenyan workers contracted through Sama for a project connected to OpenAI. Workers told TIME they read and labeled graphic text, and said they were expected to process 150–250 passages in a nine-hour shift. Those figures and accounts apply to that reported project, not to standard workloads or conditions across AI data work. TIME also reproduced OpenAI’s statement that it engaged Sama as part of work to create safer AI systems and prevent harmful outputs. That is the company’s stated rationale, not an independent finding about intent or the project’s effects.
When evaluating a sensitive-content role, relevant questions include whether workers receive advance notice and suitable training, can report harmful exposure, have access to independent and trauma-informed support, and can raise concerns without relying solely on the company managing the task. Those ILO and GIZ accounts describe these as areas of concern; they do not establish that all employers provide—or fail to provide—the same safeguards.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How many people around the world work on AI?
Available sources do not establish a dependable worldwide headcount. GIZ uses a broad “millions” characterization, but the material does not provide a common definition and counting method that would support a precise total. Workers may be counted as platform contractors, BPO employees, subcontractors or people doing occasional tasks, and the same person may take on different projects over time. Equidem’s interviews, for example, are explicitly nonrepresentative. A precise global number would need a clear definition, date, population and transparent method.
What should readers take from the evidence?
AI labor is real, diverse and often mediated by organizations that stand between workers and technology companies. The evidence documents questions about pay, social protection, automated rejection, monitoring, disclosure and mental-health support. It does not show that every AI system relies on the same workforce, that all workers face the same risks, or that one country or region accounts for the work.
The most useful way to assess a specific claim about “the people behind AI” is to identify the task, location, hiring chain and source of the evidence. Distinguish institutional findings from worker testimony, mapped networks from complete counts, and company statements from independently established outcomes. That approach keeps attention on the labor without flattening different jobs and experiences into a single story.
Frequently Asked Questions
Does human feedback mean an AI system learns from every person who uses it?
No. Human feedback in this context can come from workers specifically assigned to evaluate or refine outputs. Whether a particular product uses customer interactions for training depends on that product’s policies and settings, which the evidence summarized here does not address.
Are AI data workers all content moderators?
No. Work can include labeling images or text, transcription, output evaluation, verification and moderation. The task determines whether exposure to disturbing content is a relevant risk.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Can a single survey tell us how many people do AI data work worldwide?
Not unless it uses a clear definition and a method that reaches the distributed workforce across platforms, outsourcing firms and subcontractors. The cited investigations have defined populations or samples, not a complete global count.
What information would make an AI labor claim easier to verify?
Useful details include the task and project, the country and hiring arrangement, the worker population covered, and the date and method used to gather the evidence. For a pay claim, the unit of pay and treatment of waiting time or rejected tasks also matter.
