Under the Automation Carpet: The Invisible People Who Make AI Work

Imagine a moderation dashboard. A photograph appears; a worker has a few moments to decide which label fits. Violence? Sexual content? Hate? Context changes everything, but the interface prefers a box. The next item arrives before the fir st has emotionally left the room.

Far away, a sales presentation describes the resulting system as ‘automated safety at scale’. There is a tasteful blue cloud on the slide. No human beings are visible inside it.

This is not simply a branding problem. The cleaner the story of autonomy becomes, the easier it is for buyers to ignore who prepared the data, evaluated the model, corrected failures and absorbed disturbing material. Artificial intelligence often looks like a magic carpet because somebody has been swept underneath it.

The machine that arrives with a hidden payroll

Every AI system sits inside a chain of human decisions and tasks. Researchers select a problem. Engineers design an architecture. People collect, clean and annotate data. Others compare responses, flag unsafe outputs, test edge cases, translate instructions or moderate what users upload. Staff in client organisations repair the system’s mistakes after

deployment. Users themselves may supply feedback that improves a commercial product.

Some of this labour is highly paid and publicly credited. Some is fragmented into tasks, outsourced through several firms and made difficult to see. The term ghost work draws attention to human activity that appears, from the customer’s side, to have been performed by software (Gray & Suri, 2019). The ghost is not imaginary. Its visibility has been engineered away.

Platform labour makes that engineering easier. A complex job can be split into thousands of apparently simple judgements and distributed across workers who may never know the final product, the client or the value created. ‘Click the correct label’ sounds like a mechanical instruction. Yet the difficult cases demand language, culture, inference and emotional judgement. The machine receives a clean category; the person keeps the ambiguity.

Recent analyses challenge the myth of AI autonomy by tracing these human inputs across a system’s life cycle (Rani & Williams, 2026). Work on ‘ghostcrafting’ similarly focuses on how labour is concealed while the platform presents intelligence as a property of the model alone (Rahman & Sultana, 2025). The crucial point is not that AI is fake. It is that AI is socio-technical: computational capacity and organised human labour produce the service together.

Why invisibility is part of the product

If hidden labour were merely forgotten, a better annual report might solve the problem. But invisibility performs economic and rhetorical work.

First, it supports the autonomy premium. A tool described as self-operating appears more advanced and scalable than one described as dependent on continuing human judgement. Investors, managers and journalists are more likely to treat it as a replacement for labour rather than a reorganisation of labour.

Second, it protects price opacity. The buyer sees a subscription or an API fee, not the distribution of money along the supply chain. Several layers of

contracting can separate the brand from the worker. Each layer can plausibly say that working conditions belong to somebody else.

Third, fragmentation weakens epistemic status. When thousands of decisions are called microtasks, the people making them cease to look like contributors with knowledge. Their disagreements become ‘noise’; their rejected work becomes a quality metric; their cultural expertise becomes a column in a spreadsheet. The product learns from human judgement while the product story denies that judgement authorship.

Fourth, opacity narrows responsibility. A client may say the supplier handles the data. The supplier may say a contractor handles annotation. The contractor may say the platform merely matches independent workers with tasks. Responsibility descends the staircase until it reaches the person with the least bargaining power — and then disappears through a trapdoor.

This is why a broader observation about digital inequality matters here: ‘Exclusion begins at the level of system construction’ (Kaniecki, 2026, p. 13, author’s translation). Who is represented in the data matters; so does who is permitted to shape the categories, challenge an instruction and share in the value.

The human cost is not an unfortunate footnote

The cost can be financial. Workers may face uncertain task supply, opaque assessment, rejected outputs or unpaid time spent qualifying and searching for work. It can be psychological: repeated exposure to abuse, violence or sexual material may produce harm, while support varies and may be distant from the people setting performance targets. It can also be civic. A person whose labour trains a moderation system may have no channel to contest a culturally ignorant policy that the system will later enforce at scale.

Geography matters, but the familiar phrase ‘workers in the Global South’ can itself flatten people into scenery. These workers are not a single passive group. They bring expertise, organise, negotiate, refuse, improvise and build careers under different legal and economic conditions. Ethical reporting should not replace the myth of autonomous AI with a sentimental photograph

of an anonymous victim. It should make agency and constraints visible at the same time.

The harm also reaches users. A poorly supported annotator may be asked to resolve a linguistic or cultural distinction that an instruction manual has ignored. That decision can later influence whose speech is blocked, whose dialect is misclassified or whose distress is missed. What looks like a model error may be an organisational error fossilised in data.

Finally, there is a cost to the buyer. A university, newsroom or public body that cannot reconstruct its AI labour chain cannot honestly assess operational risk. Labour abuse, low-quality annotation, sudden contractor turnover and culturally narrow evaluation are not separate from product quality. They are among its causes.

Dignity at work means more than keeping a human ‘in the loop’

The phrase human in the loop can make any process sound reassuring. But a worker is not protected merely by being present. Dignified work requires the ability to understand the task, exercise judgement, receive fair treatment, question an evaluation and avoid preventable harm.

There is also a difference between having a person perform a step and recognising that person as a participant in governance. If workers encounter the same ambiguous category hundreds of times, they possess evidence about a system’s design. A responsible organisation needs a route for that knowledge to travel upwards. Otherwise the feedback loop is human only in the narrowest anatomical sense.

The purchaser has moral agency too. ‘We did not employ them’ is not a complete answer when an organisation deliberately benefits from a supply chain. Buyers routinely ask suppliers about cybersecurity, data location and service continuity. Labour conditions belong in the same serious conversation. International work on the AI divide frames technology and employment as a global distribution question, not simply a race to adoption

(United Nations et al., 2024). The relevant question is therefore not only Can we afford this system? but What has made this price possible?

Responsibility should follow three things: control, knowledge and benefit. A brand that controls specifications, can demand information and captures substantial value has duties even if payroll is handled elsewhere. A public institution has an additional obligation because procurement spends collective money and can set norms for an entire market.

The AI Human Labour Label

A useful labour label should not pretend that six boxes can certify justice. Its job is to make missing information visible and purchasing decisions contestable. It should travel with a system from tender to renewal.

1. Categories of work. Name the human tasks used across the life cycle: data collection, cleaning, annotation, translation, red-teaming, response ranking, safety evaluation, content moderation, customer support and post-deployment correction. Distinguish one-off preparation from continuing labour. ‘Human review’ is too vague; it is the nutritional equivalent of listing ‘ingredients’ as the only ingredient.

2. Place, employer and contracting chain. State where the work is performed, which entities employ or contract the workers and how many layers separate them from the principal supplier. Report significant changes during the contract. Location is not a moral score, but it reveals which law, language, remedy and cost structure apply.

3. Pay and evaluation. Provide pay bands or a defensible equivalent, the unit of payment, expected unpaid time, performance metrics, grounds for rejection and whether workers can see how an assessment was reached. The buyer should ask whether targets reward speed at the expense of careful judgement. A model trained on rushed certainty may later sell that certainty back as intelligence.

4. Exposure to harm and support. Identify foreseeable disturbing content, maximum exposure patterns, opt-out rules, rotation, breaks, psychological support and emergency escalation. Do not accept ‘wellbeing resources available’ without asking whether they are

accessible in the worker’s language, confidential and usable without losing income.

5. Voice, appeal and remedy. Workers need a functioning path to challenge rejected work, harmful instructions, discrimination or retaliation. Record response times and outcomes, not merely the existence of an email address. Where collective representation exists, disclose how it is included in changes to tasks or metrics.

6. Credit and distribution of value. Explain whether substantial human contribution is acknowledged in technical documentation, publications or product descriptions. Where appropriate, show how productivity gains, royalties, bonuses, training or longer-term contracts reach contributors. Not every label earns co-authorship, but no product should advertise autonomous judgement while systematically denying that judgement’s human origin.

Turning the label into a procurement decision

The label becomes useful only when absence has consequences. A newsroom or university can apply it in five steps.

Before the tender, map which planned uses are high risk for workers and users. Safety moderation, emotionally disturbing datasets, children’s data and culturally sensitive classification deserve deeper scrutiny than an office autocomplete tool.

During supplier selection, request evidence for every field: sample contracts, audit summaries, grievance statistics, support protocols and the names of responsible entities. Commercial confidentiality may justify redacting individual or competitive details; it does not justify a blank page where human rights should be.

Score uncertainty as risk. ‘Unknown’ is not neutral. If a supplier cannot identify who performs essential evaluation, the buyer should lower the score, require remediation or reject the bid. A cheap system with an unauditable labour chain carries a deferred invoice.

Write enforceable clauses. Contracts should require notification of subcontractor changes, minimum support standards, access to independent

audit and a remedy plan. They should prohibit retaliation against workers who report harms through protected channels. Price reviews should not automatically squeeze the least powerful tier.

Review after deployment. Compare promised and actual reliance on human correction. Log the cases in which client staff, freelancers or users repair outputs without recognition. The hidden workforce may move after purchase; sometimes it moves directly into the newsroom, where an editor spends every afternoon cleaning ‘automated’ copy.

Publish a concise version of the label. Readers do not need every commercial detail, but they should know whether a product depends on ongoing human moderation, where major uncertainties remain and who is accountable for improvement.

Where a label stops

Disclosure can become ethics theatre. A company may publish an elegant label while pay remains inadequate or workers lack meaningful power. Information also carries risks: revealing a precise location or subcontractor could expose workers, trade unions or security-sensitive operations. The standard should be independently governed and designed with workers, not only buyers and suppliers.

Comparisons across countries require care. A single global wage number can ignore living costs, social protection, unpaid time and bargaining power. Nor should every task be forced into a permanent job if workers value genuine flexibility. The test is whether flexibility is reciprocal: can the worker refuse without punishment, understand the rules and predict enough income to act freely?

The label cannot by itself redistribute ownership or dismantle platform concentration. It can, however, break the rhetorical spell that turns organised labour into atmospheric computing. Once a buyer knows that a function depends on people, choosing not to ask about their conditions becomes a decision rather than an oversight.

Return to the blue cloud in the sales presentation. It need not be replaced by a crowded organisational chart. One honest sentence may be enough to begin: This service is produced by software and by people whose judgements remain essential.

Then ask the adult question hidden beneath the magic trick: who are they, what do they risk, what can they contest, and how much of the value reaches them?

Automation should remove drudgery. It should not remove the worker from the moral field of vision.