Learn AI /Responsible AI, Human Rights and Safety /Labour behind AI
Labour behind AI
- Intermediate
- Responsible AI, Human Rights and Safety
- Human rights practitioners
- Policy officers
Also called: data workers, ghost work, crowd workers.
Back to the milestone you came from
In five seconds
What it is
Much AI depends on people who collect, label, review, filter and test data and outputs, often in outsourced or hidden forms of work.
In plain words
AI often looks fully automatic, but people are doing important work behind the scenes. Some prepare and label data, some review harmful or incorrect outputs, and some test whether the system fails. Different names highlight different parts of that work.
An everyday example
A town installs sensors on water pumps to warn when one pump behaves unusually. The dashboard looks automatic, but several people keep it useful. One team labels past pump readings as normal, clogged or failing so the software can learn patterns. Another worker checks strange alerts each day, marks false alarms and flags serious cases for an engineer. Residents only see a neat warning light, not the people whose work made that light meaningful.
The pump example suggests a clear technical system with stable labels and obvious ground truth. Many AI tasks are messier. Content moderation, translation review or social data labelling can involve ambiguity, cultural context and exposure to harmful material. Real labour conditions also include contracts, pay, bargaining power and health protections, which the pump story does not show.
Remember thisAI may conceal distributed human work. Responsible use requires scrutiny of who performs it, under what conditions, and who remains accountable for its quality and impacts.
This term is contested
The boundaries among these labels are not fully settled. 'Ghost work' is especially contested because some use it narrowly for hidden platform labour, while others use it more broadly as a critique of invisible human work behind supposedly automated systems.
Understand it
Understand it
Why it matters
If you buy, regulate, deploy or rely on AI, you are also relying on human labour that may be hidden. That matters for three practical reasons. First, system quality depends on this work: weak instructions, rushed labelling or exhausted reviewers can produce poor outputs and unfair errors. Second, rights and safety are at stake: some workers face low pay, insecure contracts, surveillance or repeated exposure to disturbing material. Third, accountability can become blurred across vendors and subcontractors. A non-specialist does not need to know every technical detail, but they should ask who did the work, under what conditions, with what safeguards, and which decisions must stay with accountable human officials.
Human labour across the AI pipeline
Show that human work enters AI at several stages, not only at the start, and that different labels describe different parts of that labour.
01
1. Raw material arrives
Text, images, audio, sensor logs or policy rules are gathered for a task.
Draw a box labelled 'Raw data and rules' with small icons for document, image, audio wave and sensor chart.
02
2. Work is assigned
Tasks are split into jobs such as labelling, rating, moderation, testing or quality checks. Some are done inside an organisation; others through vendors or online platforms.
Draw a second box labelled 'Tasks and instructions' connected from step 1; add small ticket cards named label, review, remove, test.
03
3. People do the work
Workers tag examples, compare outputs, remove harmful material, correct errors or escalate difficult cases. 'Data workers' is the broad label here. 'Crowd workers' is one subset. 'Ghost work' highlights hidden and undervalued labour.
Draw a large central box labelled 'Human labour' with three smaller tags above it: broad: data workers; subset: crowd workers; critical lens: ghost work.
04
4. System behaviour changes
The human work becomes training data, safety rules, benchmark scores or corrected outputs that shape how the AI performs.
Draw a box labelled 'Model training, filtering, evaluation' connected from step 3; add arrows from reviewed data and corrected outputs into the box.
05
5. Users see an 'AI' result
People encounter a chatbot answer, moderation decision, translation, ranking or alert and may not see the workers behind it.
Draw a final box labelled 'User-facing result' with a speech bubble and dashboard icon; add a faded eye-crossed symbol between step 3 and step 5 to show hidden labour.
A left-to-right flow with five boxes: raw material, task allocation, human labour, AI system, and public-facing output. Small tags above the human labour box label broad and narrower terms. Read from left to right to see where people contribute and where their work may become hidden.
On hover or tap, each term tag reveals a short definition: data workers = broad category; crowd workers = platform-based subset; ghost work = hidden labour framing. A toggle can highlight either labour tasks, rights risks, or accountability points.
Describe this without the diagram
The diagram is a five-step flow from left to right. First, raw data and policy rules enter the process. Second, tasks are broken into small jobs such as labelling, reviewing, moderating and testing. Third, people perform those jobs. A note explains that 'data workers' is the broad term for these people, 'crowd workers' is a subset who often work through online task platforms, and 'ghost work' is a critical term for labour that stays hidden from users. Fourth, this human work changes the AI system by creating training data, safety filters and evaluation scores. Fifth, users see an AI output, often without seeing the people whose work helped produce it.
With reduced motion: A static left-to-right flow chart with five boxes and labelled arrows. The term tags and accountability notes appear expanded by default, with no animation.
Many AI systems look automatic from the outside, but people often do important work before, during and after the system runs. They may collect examples, label images or text, review harmful content, check model answers, or test whether safety rules fail. This human labour helps shape what the system can do and where it makes mistakes.
The related terms are not identical. “Data workers” is a broad term for people doing tasks such as data collection, labelling, quality checks or review. “Crowd workers” usually means people doing small digital tasks through online platforms, often paid per task. “Ghost work” is a critical term for labour that keeps digital systems working but is hidden from users and often undervalued. “Labour behind AI” is the widest phrase here: it covers all the human work that supports AI, whether visible or hidden, temporary or permanent, platform-based or not.
What goes in? Raw data, instructions, platform tasks, policy rules and model outputs needing review. What happens? Workers sort, tag, compare, correct, remove, rate or escalate content. They may also help train or evaluate a model. What comes out? Datasets for training, safer filters, quality scores, corrected outputs and systems that appear more reliable than they would be without that labour.
This topic matters because working conditions affect system quality and human rights. Low pay, impossible quotas, exposure to disturbing content, weak appeals, poor safety measures or unclear contracts can harm workers and also harm the public by producing lower-quality data and weaker safeguards. It is also linked to accountability: if an organisation benefits from AI, it cannot treat the people behind that system as irrelevant.
How “A city water pump alarm that still needs people” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Labelled past pump readings | Training or evaluation data prepared by data workers |
| Worker checking strange alerts | Human review of model outputs or risky cases |
| Flagging serious cases to an engineer | Escalation to human oversight for important decisions |
| Warning light that looks automatic | AI output that hides the human labour supporting it |
A worked example
An information management officer in Bogotá is assessing bids for a multilingual tool that classifies public feedback before the procurement unit closes its technical evaluation.
Starting from: Bid documents, synthetic test comments, annotation instructions and each vendor's account of its workforce and subcontractors.
- The officer maps the human tasks behind each service, including data selection, language annotation, quality review, safety testing and referral of uncertain cases.
- The procurement unit asks who performs those tasks, under what contractual arrangements, and whether workers have appropriate language competence, clear instructions, fair review procedures and support when handling disturbing material.
- Evaluators use synthetic comments to check how consistently the service and its human reviewers apply the categories, including how disagreements and appeals are handled.
- The officer records model performance and labour-related findings in the technical evaluation, rather than treating the workforce as an invisible feature of the software.
Result: The evaluation distinguishes genuine automation from work passed to annotators and reviewers. This gives the procurement unit a clearer basis for assessing quality, accountability, working conditions and risks across the vendor chain.
In a United Nations workplace
An OHCHR protection officer in New York is finalising a briefing note for the front office before Human Rights Council informals, while delegations await instructions from capitals. The draft calls a proposed text-classification service automated, although contractor linguists would label examples, review disturbing outputs and resolve cases the model cannot classify.
The briefing should identify who performs annotation, content review, safety testing and exception handling; whether they are employees, subcontractors or platform workers; how pay, workload, language competence, exposure to traumatic material, quality review and appeals are managed; and which supplier remains accountable through the subcontracting chain.
Before you try thisThe officer must not provide survivor, witness or other protection case material to a vendor or unapproved tool for demonstration, annotation or testing. Use synthetic material unless an approved environment and a documented basis permit otherwise. Procurement safeguards should also address workers' exposure to disturbing content and access to support.
Where you may hear it
- A technical evaluation meeting before a procurement recommendation is cleared
- Human Rights Council informals on artificial intelligence and human rights
- An internal audit entrance conference on an automated service
- A vendor demonstration to an ICT governance board
- An inter-agency working group reviewing responsible artificial intelligence safeguards
Avoid confusion
Avoid confusion
Often confused with
Common misconceptions
Common misconception: AI systems work almost entirely without people once the software exists.
Correction: Many systems still rely on people to collect data, label examples, review outputs, remove harmful content, test failures and handle edge cases. Even highly automated services often need ongoing human work to maintain quality and safety.
Common misconception: All labour behind AI means the same thing as annotation.
Correction: Annotation, or data labelling, is one common task, but it is not the whole picture. Labour behind AI can also include data collection, translation checks, quality assurance, red-teaming, content moderation, reinforcement tasks, policy writing and escalation handling.
Common misconception: If workers are contractors or on a platform, the deploying organisation is not responsible.
Correction: Outsourcing can change contracts, but it does not remove responsibility to conduct due diligence, set safeguards and check whether labour practices and system outcomes create harm. Hidden subcontracting can increase accountability risks rather than solve them.
Common misconception: Hidden human labour only raises labour-rights concerns, not technical concerns.
Correction: The labour conditions and the technical result are connected. Poor instructions, unrealistic speed targets, lack of support or weak quality control can directly affect model performance, fairness, safety filtering and error rates.
Quick check
Which statement best captures the relationship among these terms?
The terms overlap, but they are not identical. 'Labour behind AI' covers the full range of human work supporting AI. 'Data workers' is a broad label for people doing such tasks. 'Crowd workers' usually refers to a subset working through online task platforms. 'Ghost work' is a critical term that draws attention to labour kept out of view and often undervalued.
Go deeper
Go deeper
The simple version
AI often looks fully automatic, but people are doing important work behind the scenes. Some prepare and label data, some review harmful or incorrect outputs, and some test whether the system fails. Different names highlight different parts of that work.
The technical version
Labour behind AI refers to the distributed human work embedded in the AI lifecycle: data collection, curation, annotation, content review, quality assurance, model evaluation, safety testing and exception handling. 'Data workers' is a broad occupational label for people performing these tasks. 'Crowd workers' usually refers to workers completing decomposed microtasks through digital labour platforms, often paid per item. 'Ghost work' is an analytical and normative framing for labour that makes automated services function while remaining socially and economically obscured from end users. The mechanism matters because this labour affects dataset composition, label consistency, benchmark validity, moderation outcomes, guardrail performance and downstream error patterns. It also raises questions about labour standards, psychosocial risk, subcontracting chains, transparency and organisational accountability.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives policy-oriented readers an official international framework for ethics, rights, accountability and labour-related governance questions around AI.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
This OHCHR B-Tech Project resource suits readers focused on business responsibility and human rights impacts in digital technology, including due diligence and accountability across value chains.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Labour behind AI (opens in a new tab)Wikipedia
This Wikipedia search results page helps a general reader see the range of related topics and terms, but it is a search page rather than one settled definition.
search results, not one pagelink checked 2026-07-31
VideoVideo explanations: Labour behind AI (opens in a new tab)YouTube search
This YouTube search results page suits readers who prefer audiovisual introductions, but it is a search page with mixed-quality videos rather than one verified explainer.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Labour behind AI (opens in a new tab)Reddit search
This Reddit search results page shows public discussion and debate around the topic for readers who want informal perspectives, but it is a search page and not an authoritative source.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
AI may conceal distributed human work. Responsible use requires scrutiny of who performs it, under what conditions, and who remains accountable for its quality and impacts.