Uber Drivers Train AI for Extra Money During Downtime

Your next Uber driver could be using the ride-hailing app to train an artificial intelligence. Now the company is quietly testing a new feature that allows drivers and couriers to complete digital microtasks using their cell phones while they are waiting for rides in their cars, all through its app — think short voice samples, or taking photos of objects or recording snippets of text in multiple languages — and get paid instantly by Uber within 24 hours. It’s a new extension of the gig model: turn downtime into something useful — data that businesses use to build smarter A.I. systems.

How Uber’s microtask training program for drivers works

Participation is opt-in, and assignments appear alongside existing earnings opportunities available with the app in the Driver app experience, Uber said. The tasks themselves are marketed as light, smartphone-friendly prompts: record yourself speaking in your native language, upload photos of things such as receipts or produce, and submit documents that have non-English text. Rates are negotiable and dependent on time commitment and complexity, with pay being credited to a user’s balance within 24 hours.

Most important, Uber says it will route tasks from companies that “need real people to help improve their technology,” though it won’t reveal your employer’s name or goals. The company also suggests that the list of tasks will grow over time. In other words, this is not full-time work — it is opportunistic, ride- or delivery-demand-driven crowd labor that fills in the white space between rides or deliveries.

Why AI Companies Crave Your Data for Training and Safety

Generative AI is thirsty for new, legally clean datasets. Publishers, record labels and news organizations have sued leading AI developers for copyright infringement, making web scraping a legal and reputational liability. By building datasets of explicit consent — photos, voices, handwriting, accents, everyday scenes — one decreases that exposure and makes models better on “long tail” real-world inputs the public internet text or stock imagery might miss.

This model isn’t entirely new. Shutterstock has entered into licensing arrangements to provide images and footage for training, as well as pay contributors; Adobe has paid creators whose Adobe Stock works helped train its Firefly models. Uber’s spin is distribution: a huge, always-on workforce with cameras and microphones, already used to task-based pay and app-based workflows.

What this microtask work pays and what it could cost drivers

Uber has never released a rate card or revealed what its cut is. Crowdwork platforms have been known to demonstrate various characteristics (in the past). In 2018, an analysis published in the Proceedings of the ACM on Human-Computer Interaction found that median pay per hour for jobs posted to Mechanical Turk was abysmally low — a few dollars and frequently below local minimum wages, when there even were local guidelines. A Time magazine investigation into foreign data labeling reported very low wages for moderating and annotating in the Global South. In short, we’ll need to see transparent rate and time tracking if workers are going to make solid money.

For drivers, the attraction is clear: a source of income in down time without burning fuel. But there are hidden costs to consider — battery life, data usage and the danger that microtasks bleed into your on-trip moments if prompts aren’t timed well. For Uber, she added: “What we need is guardrails to avoid overlapping on safety and to ensure that the work remains optional and separate from ride acceptance criteria.”

Privacy and data ownership questions facing Uber’s program

Voice recordings, video images of the face and documents can be sensitive. Uber says it won’t expose users, and that shared material may be kept or moved by the companies asking for it. Which brings us back to a question that is at the heart of the current AI debate: Who does own it, once submitted? Can workers revoke consent? Could their voices or images be used to train biometric systems, and will those applications be made explicit?

Regulators are paying attention. The NIST AI Risk Management Framework recommends organizations document data provenance and consent, while state privacy laws in California, Virginia and Colorado demand clear notice about data collection and secondary use. The Biometric Information Privacy Act of Illinois imposes stringent restrictions on voiceprints and face data. This will be critical for a program that stretches across multiple jurisdictions, where coherent disclosure requirements will be key.

A new link in AI’s data supply chain through rideshare apps

AI developers have long turned to companies such as Appen, Scale AI and Sama for tasks like annotation, red-teaming and preference rating. In a rideshare app, embedding microtasks could even introduce a flexible, localized layer: want audio clips of Midwestern English being spoken, pictures from grocery aisles at night or handwritten Arabic numerals? A distributed driver base can bring fast and geographically diverse data.

But speed isn’t everything. Quality assurance, bias removal and documentation are important too. The Partnership on AI has recommended better labor standards, clear instructions and feedback loops for annotators. If Uber can replicate those best practices — including fair pay floors, quality bonuses and transparent audits — it could change sporadic side gigs into a continuous pipeline with regular payouts that benefit workers and model builders alike.

What to watch next as Uber tests consented data collection

Three signals will indicate whether this experiment has legs: clear pricing and time estimates per task, informed consent in plain language about downstream uses, and provenance labels so that AI builders can follow the threads backwards to ensure permission is traceable back to the source. If Uber nails those, it could unexpectedly emerge as a broker of consented training data — and afford laborers a safer way to monetize their voices, photos and knowledge.

If it doesn’t, the program could turn into another murky corner of the gig economy: sporadic, low-paid and filled with hidden trade-offs. The tech world is hungry for top-quality, ethically sourced data. The unanswered question is whether it can be delivered by a ride-hailing app — without the onus again being put on those who are behind the wheel.