Somewhere right now, a gig worker is strapping on a body camera, loading a dishwasher, and filming the whole thing for a few dollars. Another is recording an unscripted conversation in Spanish. Someone else is uploading a selfie video, rotating their face slowly left, then right, then center. All of it feeds into training pipelines for AI models that will, in many cases, learn to do the very work these people currently survive on. The going rate for a session of biometric data collection is roughly $3 to $10. A face scan, a voice recording, a video of your daily surroundings. For the companies buying this data, it is the raw material behind systems worth billions. For the workers selling it, it is Tuesday.

The new data gold rush

The AI industry has a supply problem. The ocean of scraped internet data that powered the first wave of large language models is running dry. Researchers have warned since as early as 2022 that available high-quality text data could be exhausted within a few years, and by 2026 that pressure has become real. Synthetic data helps, but it has limits: models trained mostly on their own outputs risk a kind of intellectual inbreeding, producing fluent nonsense that scores well on benchmarks but breaks in the real world. So the industry has turned to humans directly. Not as annotators labeling images in a browser, but as full-spectrum data sources: their voices, their faces, their hands folding laundry. DoorDash launched a program called Tasks in March 2026, turning its 8 million U.S. delivery couriers into a distributed data collection network. Through a standalone app, workers accept assignments to film household activities, scan supermarket shelves, photograph hotel entrances, or capture restaurant food images. Robotics firms use the household footage to train humanoid systems in "contact-rich manipulation," the kind of dexterous physical tasks that simulations cannot replicate. Uber runs a comparable program through its AI Solutions division, which now operates in 30 countries and offers annotation, translation, and model training services to corporate clients. The framing is cheerful. "A new way to earn," as DoorDash put it. But the underlying economics tell a different story.

A few dollars for data worth billions

The global AI training data services market was valued at $4.47 billion in 2025 and is projected to reach $32.1 billion by 2034. The workers generating that data earn a few dollars per task. This is arguably the most lopsided value exchange in the history of technology. It is not that the individual data point is worth billions. No single face scan changes a model's performance. But in aggregate, these datasets become competitive moats. Platforms like DoorDash and Uber are positioning their distributed workforces as proprietary data supply chains, accumulating training datasets that AI developers and robotics firms cannot easily replicate. The worker gets paid once. The data generates value indefinitely. The parallel to content creators is hard to miss. Writers, artists, and photographers whose work trained GPT-3 and GPT-4 were never asked. That sparked lawsuits, regulatory scrutiny, and a broader reckoning about consent. The gig workers selling biometric data are at least getting something, which is more than most content creators got. But is the consent truly informed when the consequences of the trade are fundamentally unknowable?

The credential you cannot rotate

Here is the security angle that makes biometric data different from every other kind of personal information: you cannot change your face. If your password leaks, you reset it. If your credit card is compromised, you get a new number. If your address is exposed, you can, in theory, move. But biometric data is permanent. Your face, your voice, your fingerprints, these are credentials you carry for life with no option to rotate them. Once your biometrics enter a training set, there is no un-training them. The data becomes part of the model's weights, entangled with millions of other examples. You cannot request deletion in any meaningful sense. This makes the decision to sell biometric data fundamentally irreversible, even if the payment is trivially small and immediately spent. The deepfake risk compounds this. Biometric data collected for legitimate AI training can be repurposed, leaked, or stolen. A dataset of face scans intended to improve authentication systems can just as easily teach a generative model to produce convincing impersonations. Data companies like Clearview AI and PimEyes have already demonstrated that facial data, once collected, tends to find its way into uses far beyond what the original subject imagined.

The Jevons paradox of data

In economics, the Jevons paradox describes how making a resource cheaper to use often increases total consumption rather than decreasing it. Coal-efficient steam engines did not reduce coal use; they made coal-powered industry viable at a scale that consumed far more coal than before. The same dynamic applies to human data. As biometric data becomes cheaper and easier to acquire through gig platforms, the demand for it will not plateau. It will expand. Every new AI application that demonstrates value from real-world human data creates incentive to collect more of it. Household robotics needs hand movements. Autonomous vehicles need pedestrian behavior. Voice assistants need accented speech in dozens of languages. Medical AI needs physiological data. The cheaper each data point becomes, the more aggressively companies will harvest it. A $5 face scan today sets the price expectation for a $3 face scan tomorrow and a $1 scan next year, even as the models built from that data become more capable and more profitable.

Where does Singapore stand?

Singapore's Personal Data Protection Act (PDPA) treats biometric data as personal data requiring consent and appropriate safeguards. Section 13 of the PDPA requires organisations to obtain consent before collecting personal data, including biometric information. The Personal Data Protection Commission (PDPC) has published guidance clarifying that biometric data used for identification purposes is subject to all PDPA obligations, and collection must be reasonable and limited to what is necessary. In 2022, the PDPC released a dedicated guide on the responsible use of biometric data in security applications, covering facial recognition systems and security cameras. The guide addresses collection, use, and disclosure obligations, along with practical guidance on data protection. But there is a gap. The PDPA was designed primarily with organisational data handling in mind, think employers using fingerprint scanners or malls running facial recognition. It was not designed for a world where individuals voluntarily sell their own biometric data to overseas AI companies through gig platforms. The question of whether a worker in Singapore who uploads face scans to a DoorDash-style task platform is meaningfully "consenting" under PDPA standards is untested territory. The data leaves the country, enters training pipelines governed by different (or no) regulatory frameworks, and becomes irrecoverable. Southeast Asian data protection law more broadly has been identified as needing significant updates to address biometric data in the AI era. A 2024 analysis across five countries in the region concluded that more needs to be done to protect biometric data and the rights of data subjects.

Who captures the value?

The fundamental question is not whether gig workers should be allowed to sell their biometric data. Many are making rational choices given their constraints. A few dollars is real money when the alternative is nothing, and paternalism about other people's economic decisions is easy from a position of comfort. The real question is structural: who should capture the value that training data creates? Right now, the answer is almost entirely the platform and the AI company. The worker gets a one-time payment. The platform gets a proprietary dataset. The AI company gets a model that generates revenue for years. There is no royalty, no residual, no ongoing relationship. The data is extracted, the value accrues elsewhere, and the transaction is complete. This is not a new pattern. It is the logic of raw material extraction applied to human bodies. And like previous resource extraction cycles, it will likely require collective action or regulation to rebalance, because individual market transactions alone will not close the gap between what the data is worth to the buyer and what the seller is paid. Some possible directions: data trusts that negotiate collectively on behalf of contributors, regulatory requirements for ongoing compensation when biometric data is used in commercial models, or transparency mandates that force companies to disclose what data was collected, from whom, and for what purpose. None of these exist yet at meaningful scale. But the window for shaping this market is still open. The gig economy's pivot from delivering packages to delivering human data is barely a year old. The norms, the regulations, and the power dynamics are still being set. For now, your face is worth about $5. The question is whether that price was set by a market, or by the absence of one.

References