Thank You for Your Unpaid Labour

Select all squares containing traffic lights.

I click three boxes, hesitate over a fourth where the traffic light occupies perhaps six pixels, and submit my answer. Apparently I am human. The website lets me in.

There is something almost too neat about this exchange. I prove I am not a machine by performing a tiny piece of visual classification—exactly the capability machines have been learning to acquire.

That was once an explicit part of the bargain. Early reCAPTCHA prompts used human answers to digitise books that optical character recognition could not read. After Google acquired reCAPTCHA in 2009, the system moved from distorted words to street numbers and names. Modern reCAPTCHA relies much more heavily on risk analysis, so it would be wrong to suggest that every traffic light we click trains a self-driving car.

As a metaphor, it is difficult to improve upon. Prove you are human. Help the machine understand the world. Continue to the product.

Work that stopped looking like work

Once you notice this arrangement, you see it everywhere. I scan and bag my shopping, print my luggage label, photograph my returns and work through diagnostic scripts before a company will let me speak to a person.

I often prefer the self-checkout: the queue is shorter and I can pack things in the order I want. That convenience makes the shift easy to miss, but work that once sat inside the firm has been divided into pieces, pushed outside it and renamed self-service.

The writer Craig Lambert calls this shadow work: unpaid tasks performed for businesses and institutions without looking like labour. The checkout did not disappear; part of the job moved to my side of the scanner. The language helps: self-service is convenience, CAPTCHA is security and writing reviews is community.

Points, badges and stars

Some systems make the work feel like a game. Google’s Local Guides programme rewards reviews, photographs and corrections with points, levels and badges. Contributors may enjoy the recognition; Google gets a commercial map maintained by people outside its payroll.

Ratings distribute management instead. After an Uber journey, the passenger becomes part customer, part quality inspector and part line manager; Uber states that consistently low ratings may lead to deactivation.

This is not automatically bad. Ratings can surface dangerous behaviour and maps improve when local people correct them. The issue is whether participation creates a shared resource or a privately controlled asset.

Recycling is an interesting edge case because the benefit is shared. The unfairness appears when producers generate packaging while dealing with it is framed as individual responsibility. The UK’s Extended Producer Responsibility scheme will move the cost of household packaging waste from taxpayers to producers. Households still sort it, but financial responsibility travels back up the chain.

The policy does not pay somebody to rinse a yoghurt pot. It reconnects ownership of a decision with its eventual cost. Circularity cannot only mean sending material around the loop while economic responsibility continues in one direction.

Silicon Valley without the jobs

Not all of this work is unpaid. To become an influencer is, in one sense, to apply for a job in Silicon Valley without sending a CV. The public interview can last for years while the applicant buys the equipment, develops the product and finds an audience before the platform promises a penny.

Platforms have opened access to global audiences and incomes, but the ILO describes the resulting work as irregular, insecure and dependent on a few powerful gatekeepers. The creator provides the studio, labour, personality and risk; the platform controls distribution and can change the rules. Streaming makes the arrangement literal: the creator supplies hours of their life, the audience supplies chat, clips and attention, and the platform sells access to both.

YouTube makes this unusually explicit: it may place adverts on videos outside its Partner Programme, even though those creators receive no share. For somebody living thousands of miles from Google, Meta or X, content may look like the most accessible route to earning money from them. Silicon Valley has created a global workforce without needing to employ most of it.

The people inside the machine

AI gives this arrangement industrial scale. Models depend on people classifying, writing, ranking and reviewing information so that machines can learn from human judgement.

OpenAI describes the original ChatGPT training process as reinforcement learning from human feedback, or RLHF. Trainers wrote example conversations and ranked alternative responses, turning their preferences into a signal for what a better answer looked like.

The Kenyan work often mentioned alongside RLHF was related, but not identical. A TIME investigation reported that employees of the outsourcing company Sama labelled descriptions of sexual abuse, hate speech and violence for an OpenAI safety system. Interviewed workers described psychological harm, while take-home pay for junior labelers was reported below $2 an hour. Sama disputed parts of the account.

Calling every human training task RLHF would hide the work as effectively as calling the whole system AI. The broader category is data enrichment: labelling harm, comparing answers, marking objects and writing examples.

A CAPTCHA asks whether I can recognise a traffic light. RLHF may ask which of two answers is better. Safety labelling may ask a worker to classify something nobody should have to read repeatedly. The stakes and conditions are profoundly different, and a few unpaid clicks should not be equated with a traumatic working day. The connection is that human judgement enters the machine while the people supplying it disappear from the finished product.

The Partnership on AI calls these workers essential and argues that AI companies remain responsible throughout their supply chains. That suggests a digital version of Extended Producer Responsibility: accountability should travel towards the organisations specifying the work and capturing the value, not stop at the subcontractor.

The collective second brain has shareholders

In 2024, I described language models as a collective second brain: an interface over humanity’s accumulated text. I still think that holds. I was looking at the collective input rather than ownership of the output.

Language models are possible because people spent centuries writing books, articles, software, documentation, arguments and jokes for other reasons. Nobody writing a programming tutorial in 2008 imagined they were contributing to a commercial coding model in 2026.

Online communities make the transfer visible. Reddit and Stack Overflow grew through people posting, moderating and correcting one another. OpenAI now has structured access to Reddit’s content, while Stack Overflow supplies community knowledge through its Google Cloud partnership.

These arrangements have licensing terms and may promise attribution or new features; they are not simply theft. Yet people created knowledge for one another, a platform organised it, and years later access became an asset negotiated by the platform. They were a community right up until the community became a dataset.

AI labs did not invent unpaid labour. They may represent its most ambitious aggregation: culture produced collectively, refined by a hidden workforce and returned through privately controlled systems.

Prove you are human

This brings the argument back to the CAPTCHA at a larger scale. World, formerly Worldcoin, is building infrastructure through which someone can prove they are a unique human online. The World Foundation stewards the protocol, while Tools for Humanity develops important parts of it. Sam Altman co-founded Tools for Humanity and is also chief executive of OpenAI.

The symmetry is difficult to ignore. One of the most influential figures in making machine-generated activity abundant also co-founded infrastructure designed to prove that a human remains on the other side of the screen. What began as an occasional CAPTCHA could become a persistent economic identity.

World’s founders present it as a way to spread AI’s benefits and a possible route towards AI-funded universal basic income. That may be a sincere attempt to solve the distribution problem. It also creates a remarkable loop: after collectively producing the material from which AI learned, we may need another system from the same technological networks to certify our humanity and distribute some of the prosperity back to us.

Perhaps that is pragmatic. Perhaps it is another layer of dependency. I have not resolved which.

Participation or extraction?

Not all unpaid contribution is exploitation. Football crowds create the atmosphere they came to experience, while volunteers hold Wikipedia together. Communities depend on people contributing more than they are required to.

The difference is whether the contribution remains part of a commons or becomes an input to an asset over which contributors have no meaningful control.

Did I choose to contribute? Could I understand what my effort might become? Who owns the resulting system? If it succeeds, do the benefits flow back towards the people who made it possible?

These questions produce different answers for self-checkout, recycling, Wikipedia and a frontier AI model. That is the point. The problem begins when participation becomes extraction: effort dispersed downwards, ownership concentrated upwards. Terms and conditions are not a sufficient theory of economic participation.

The strangest part is not that machines learned from us. Every tool inherits something from the people who came before it. It is how quickly our contribution disappears once the system becomes valuable: the machine becomes the product, the company becomes the creator and we become merely its users.

Select all squares containing traffic lights.

Thank you for your unpaid labour.