Why we need more transparency in training data

It’s the beginning of August. I’m sitting in a green backyard in Nairobi. Kenya has long been known for academic ghostwriting. People have written academic papers for scientists in the West for little money. While ghostwriting is becoming increasingly popular ChatGPT and Co. is replaced, the business that makes these systems possible in the first place is flourishing in Kenya: data annotation for the training of AI.

Little is known about who the people who do this work are. It is not trivial to locate them and make contact. Because unlike the field of content moderators, where the workers are employed by outsourcing companies, data annotators work as freelancers and are hardly networked.

Everyday work at the data annotators

Julia Kloiber writes for the printed edition of MIT Technology Review a regular column. (Photo: Oliver Ajkovic)

They come across the job advertisements through friends or platforms like Linkedin. They then go through a short online training. They never receive a formal employment contract. Once the training is complete, they wait in front of their computers for orders that are sent to them via a platform. One of the annotators tells me that the order situation was still quite good after the training, but that later he sometimes waits up to two weeks for new orders. With jobs in the pipeline, people can spend unlimited hours a day annotating and describing images, videos and text. There are no working time regulations. Practical for clients who only see labor rights as obstacles anyway. Not to mention unionization.

The data annotators I’m sitting with on this beautiful August day tell me that they work up to 20 hours a day. Continuous work is provided with incentives: the system favors those who are persistent by giving them more jobs.

Plants, pets, technology: she records what she sees

When I ask which clients they do this work for, they shrug their shoulders – I have no idea. You can only speculate. A freelancer works for hours every day on photos taken by robot vacuum cleaners in homes in the US and Europe. To train the AI, it draws what it sees: plants, toys, pets, technology. “Now I know how you live,” she slips out.

Others talk about LIDAR or stereo image data sets (SID) for training autonomous vehicles. An annotator describes to me how he sometimes annotates a single SID sequence for two weeks at a time. It also happens that the material is returned to him with the comment – error in the annotation – and he does not receive any payment. For this work, the annotators need computers with powerful graphics cards and a fast internet connection. Because they are freelancers, they have to do everything themselves.

Lack of transparency is part of the business model. It is confusing how many companies contract out this work to freelancers and who the clients are. It recently emerged that Kenyan workers trained the facial recognition software used by the Russian government to identify critics. The annotators knew nothing.

AI use in a military context?

Just in May, Google Deep Mind employees signed an open letter calling for an exit from AI military contracts. Are the annotators from Kenya also involved in projects for the use of AI in a military context without knowing it – and therefore without the right to make an ethical decision for themselves?

There is undoubtedly a need for more transparency in the field of data annotation – in terms of working conditions, clients and the proportion of human work. You don’t have to go far away to do this. You can start on your own doorstep and ask German AI start-ups where and under what conditions they have their training data annotated. I wouldn’t be surprised if you look into puzzled eyes. After all, people are too busy marketing AI. There is hardly any room for labor law and ethical issues. AI made in Germany is therefore just as hollow a slogan as it was in the clothing industry back then, when only one production step was brought to Germany for the “Made in Germany” label.

The author of this text is Julia Kloiber. She works on fair and inclusive digital futures as co-founder of the feminist organization Superrr Lab. In her column in the printed edition of MIT Technology Review, she reports on her experiences in and with the tech world.