People in India are being paid to film themselves at work so AI systems can learn from the footage. The main risk is not the camera. It is that the workers carry most of the privacy and security exposure while having little say over the data.
Key takeaways
- The BBC reports that workers in India are being paid to record themselves doing their jobs, and that the footage is used to train artificial intelligence systems.
- Video recorded at work usually captures much more than the task it was meant to show, including other people, documents, screens and clues to where it was filmed.
- Once footage enters a training dataset it is hard to trace, correct or withdraw, so the people who recorded it largely lose control over how it is used.
- Paid, consensual data collection can be fairer than scraping material without permission, but only if consent, security and the right to withdraw are real in practice.
The argument: the data trade pushes risk onto the people least able to manage it
The BBC reports that workers in India are being paid to film themselves while they work, and that AI developers use the recordings as training material. The report does not say how many people are involved, how much they are paid, which companies buy the footage or which AI systems it goes into. Those details are not known here and should not be assumed.
The general pattern is easier to describe. Systems that interpret the physical world, such as models that recognise objects, follow sequences of actions or guide robots, need large amounts of real-world video. First-person footage of someone doing a manual or routine job is especially useful because it shows hands, tools and steps from the angle a machine might one day work from. Public video on the internet is often badly suited to this, so developers are paying people to produce it on purpose.
This article argues that the arrangement is a privacy and data-security problem as much as a labour story. The worker sells some footage. What is actually handed over is a lasting record of a workplace, the people in it and the worker’s own body and habits, and that record goes into a supply chain the worker cannot see. The buyer gets the long-term value of the data and pays a one-off fee. The worker keeps the long-term risk. This imbalance is the central issue, and it applies whether or not any harm has happened in a given case.
Evidence: workplace video captures far more than the task
A camera recording someone doing a job does not record only the job. Depending on the setting, it may pick up colleagues, customers or family members who never agreed to be filmed. It may show paperwork, labels, computer screens, phone notifications, addresses on parcels, or the inside of a home if the work happens there. Background sound can catch conversations. Even without faces, someone’s hands, voice, clothing and way of moving can help identify them, especially when combined with other information.
The video file itself can also carry metadata, such as device details, timestamps and sometimes location, unless it is removed before upload. It is not known whether the arrangements described by the BBC remove this data or blur faces and screens before the footage leaves the worker’s phone or camera. This is a design choice, not a technical limit. Some collection programmes do it and some do not.
This matters for security as well as privacy. Footage of a shop counter, a warehouse floor or a workshop can show layouts, stock levels, security measures and routines. In aggregate, such recordings could interest people with criminal aims, not just AI developers. A worker filming their own job may not have permission to record the premises at all, and the employer or property owner may not know it is happening. That opens a second gap in consent: the person paid to film is not always the person whose space or information is being captured.
Evidence: training data is hard to recall once it leaves the worker’s hands
Data shared for AI training rarely stays in one place. It is uploaded to a collection platform and may pass through intermediaries such as labelling contractors, annotation tools and cloud storage before reaching the developer that trains a model. Each step is another place it can be copied, stored for too long or exposed through a breach or a badly configured server. Datasets are also sometimes licensed on, merged with others or kept long after the original project ends.
For the person who recorded the footage, this makes withdrawal very difficult. Removing someone’s video from a dataset is possible in principle. Removing its influence from a model already trained on it is much harder, and researchers continue to debate how reliable current methods are. Models trained on large datasets can also, in some conditions, reproduce fragments of their training material, which raises the possibility of identifiable details surfacing again.
India passed the Digital Personal Data Protection Act in 2023, which sets out duties for organisations that process personal data and rights for the people the data describes. How it applies to paid video collection for AI, and how it works when the buyer is outside India, are questions this article cannot answer with confidence. The BBC report does not say whether the companies involved are based in India or abroad. Cross-border data flows generally make rights harder to enforce, because the person whose data it is may have no practical way to reach the organisation holding it.
Evidence: consent given under economic pressure is thin consent
Consent is the main safeguard in any arrangement like this, and its quality depends on the circumstances in which it is given. A worker paid per hour of footage, perhaps through an app with terms in a language or legal style they find hard to follow, is not negotiating on equal terms with a technology buyer. They may not know what the footage will train, who will see it, how long it will be kept or whether it can be resold.
The difficulty grows when the money matters to the household. People who need the income have strong reasons to accept whatever terms are offered, film more than they planned, or overlook the presence of others in the frame. None of this means the workers are being deceived. It means agreement to a contract does not settle the privacy question on its own. Data protection thinking generally treats consent as meaningful only when it is informed, specific and freely given. Each of those conditions is harder to meet when the person giving consent has less information and less bargaining power than the person asking.
The strongest case against: paid, consensual collection is better than the alternatives
There is a serious argument that this framing is too pessimistic. AI developers need real-world data, and much of the material used to train earlier systems was collected with no consent and no payment, by scraping websites, social media and video platforms. Paying people to record footage on purpose, with a contract, is arguably a move in the right direction. Workers are recognised as contributors and paid for something that was previously taken for free.
The footage may also be low in sensitivity. Close-up video of someone folding clothes, sorting goods or using a tool may show little that identifies them, especially if the collection guidelines tell people to keep faces, documents and other people out of shot. Well-run programmes can remove metadata, blur faces and screens automatically, check uploads before accepting them and limit how long data is kept. If those steps are taken, much of the risk described above shrinks.
Finally, the work provides income. For some people, filming tasks they already do may be an easy and flexible way to earn money, and outside critics should be wary of dismissing an opportunity that participants may value. Seeing the workers only as potential victims can understate their own judgement about a trade-off they have chosen to make. On this view, the right response is to set clear standards for the practice, not to treat it as harmful in itself.
That case is a fair one. It does not remove the imbalance, though. It describes how the imbalance could be managed if buyers chose to manage it, and the public information available does not show whether they do.
Evidence that would change the conclusion
Several kinds of information would weaken or strengthen the argument made here. If the companies collecting this footage published clear details of what they gather, how they remove metadata, blur faces and screens, how long they keep the data and who they share it with, much of the concern about hidden supply chains would ease. Independent audits checking those claims would carry even more weight than the companies’ own statements.
Evidence about consent would also matter. If workers were shown to understand what their footage would be used for, to be able to delete it later, and to receive terms in plain language they could read, the claim of thin consent would be much weaker. Surveys or interviews with participants, carried out independently of the buyers, would be the most direct test.
In the other direction, documented breaches of datasets collected this way, footage that identifies bystanders turning up where it should not, or regulators finding that the practice falls outside existing data protection rules would all strengthen the case that the risks fall on the workers. Clear guidance on how India’s data protection law applies to paid AI data collection, including when the buyer is abroad, would make it much easier to judge whether the current arrangements are adequate.
Until that information exists, the reasonable conclusion is a careful one. The practice is not obviously harmful, but its structure lets the costs of any failure fall mainly on the people who were paid least and know least about where their footage ends up.
Sources and further reading
- BBC News technology coverage, which reported that workers in India are being paid to film themselves at work to produce AI training footage.
- Government of India material on the Digital Personal Data Protection Act 2023, for the legal framework covering personal data in India.
- Academic research on the privacy of machine learning training data, including work on memorisation and removing data from trained models.
- Reports from labour and digital rights organisations on data labelling, annotation work and the wider supply chain behind AI datasets.
Surfaced from the rss:bbc_tech signal “workers filming jobs for AI”. AI-assisted draft, editorially reviewed.
