Elisabeth Asser often gives talks on artificial intelligence, and is constantly asked the same question: Is it possible to use language models in an ethical way? So far, her answer has been no.

"It has to do with the way these language models are designed," she says.

by Walter Wehus

Asser compares using language models such as ChatGPT and Claude to buying blood diamonds. She has drawn this comparison from the American researcher Joy Buolamwini.

“These are diamonds mined at the cost of great human suffering. Brutal exploitation of resources and cynical exploitation of labour,” says Asser.

Morten Goodwin does not immediately agree with the comparison.

“That’s a strong claim,” he says.

People are involved in the training process

Goodwin explains that a language model is trained in two main phases.

"The first phase is basic training. Essentially, this means that the model is trained using everything available on the internet", he says.

This initial phase is followed by fine-tuning.

“That’s when people assess the answers the model provides and rate the content it generates. Perhaps it’s as simple as giving a thumbs-up or thumbs-down to what the model delivers,” says Goodwin.

He believes it is difficult to envisage a good language model without people being involved. And it is precisely in these two phases that Asser believes the problems lie.

Used without permission

“The text data that the large models have used for training has, after all, been stolen,” says Asser.

Books, images and music have been taken from the internet and are being used to train the language models without the original creators having any say in the matter. Asser believes the way we talk about language models obscures what this is really about.

"When a million books are referred to as a ‘dataset’, it’s easy to forget that these are a million different forms of human artistic expression, each of which has its own intrinsic value", she says.

Goodwin is more ambivalent. He points out that researchers can also benefit from being discovered by language models, in the same way that they benefit from being discovered by Google.

“So not everything is pirated, even though this applies to a great deal of the material the models are trained on,” he says.

The work nobody sees

During fine-tuning, the models must learn to recognise what can be described as hateful, racist, violent or related to abuse. And for that to happen, someone has to flag that content first. This is currently done by people in countries such as Kenya and Venezuela. These people work in short-term employment arrangements for third-party firms with loose ties to those developing the language models.

“Being exposed to so much horrific content – very graphic and over a long period of time – takes its toll on a person,” says Asser.

She refers to the journalist Karen Hao and her book *Empire of AI*. Hao describes how the workers have little bargaining power. They are not made aware of the risks they face in advance, and receive neither protection nor support during or after the work.

“Police officers who deal with abuse material are offered psychological support. I doubt these people are offered the same support, says Goodwin.

Time Magazine has documented that Kenyan workers who carried out this work for OpenAI earned less than two dollars an hour.

"Perhaps we are heading towards a new underclass here, a bit like the factories during the Industrial Revolution", says Goodwin.

Both Asser and Goodwin point out in the podcast that there is little the individual can do about it. Language models are already embedded, for example, in Word, in Outlook and in Google’s search results.

"I can choose not to buy a blood diamond. But in practice, it is difficult to choose not to use artificial intelligence," says Goodwin.

Asser believes that demands must be directed at those with the most power. That is to say, developers, procurers, public authorities and large companies. If a Norwegian local authority chooses the cheapest model without setting any requirements, it will likely be at the expense of someone else.

"We cannot individualise a structural problem. A responsible stakeholder should ask where the training data comes from, which subcontractors are used, and how working conditions are managed", says Asser.

The two researchers believe that much can be solved with smaller and cheaper models.

"You don’t need a gigantic language model to summarise a text", says Goodwin.

Small language models are faster, and they are easier to keep under control. The downside is that you have to think a bit more for yourself.

"Artificial intelligence is a technology that has a huge impact on education and working life, the public sector and democracy. So perhaps the question shouldn’t be whether you or I are acting ethically when we use ChatGPT. Perhaps the question should rather be why we, as a society, accept that this technology has a value chain about which we have so little insight,” says Asser.

The value chain behind a language model:

Basic training: The model is fed with vast amounts of text, code and images sourced from the internet.

Fine-tuning: People assess responses and flag content. This includes, amongst other things, violence, abuse and hate speech.

Wages: Time Magazine has documented hourly wages ranging from $1.32 to $2 for workers in Kenya, depending on seniority.

Transparency: Stanford University’s Foundation Model Transparency Index shows that AI companies are becoming less transparent, falling from an average of 58 out of 100 in 2024 to 40 in 2025.