According to philosopher Maren Behrensen, assistant professor of philosophy, debates about ethical AI rarely reach beyond the screen. For Behrensen, ethical AI in research is not just about whether AI should be used to write assignments. It is also about who powers the servers, who trains the models and who takes responsibility when things go wrong.
The hidden costs
One place where this broader perspective becomes unavoidable is in the material and human cost of running AI. The servers that make AI possible do not run on goodwill. They rely on electricity, cooling water and underpaid human labour, which Behrensen describes as “the kind of human cost we don’t often see”. Every click, prompt and generated image depends on infrastructures that most users never think about. Large data centres consume vast amounts of energy to keep AI systems running and cool. As Behrensen explains, the energy costs of these systems are huge, yet they are usually missing from discussions about AI ethics.
There is also a human cost. Thousands of workers, often in countries such as Kenya or the Philippines, are paid to label images, filter harmful content and fine-tune algorithms, frequently for very low wages. In the best case, Behrensen says, this work is dull and repetitive. In the worst case, people are exposed to extremely disturbing content, and some leave the job with PTSD because of what they have had to look at. For Behrensen, this labour and energy use form part of the ethical footprint of every AI-generated sentence, whether it appears in a student essay or a scientific paper. Yet these costs rarely play a role in debates about AI use at universities, where the focus often remains on surface-level questions such as whether AI use is allowed or whether it counts as plagiarism. Important questions, Behrensen agrees, but not the full picture.
Ethical AI is more than plagiarism
Authorship and accountability
The systems behind AI also shape who is held responsible when things go wrong. Most academic journals and universities currently agree that AI cannot be listed as an author and that a short statement explaining how and why AI was used usually suffices. But this leaves a larger question unanswered: who is responsible when AI produces errors or causes harm? At the moment, accountability is mostly pushed onto the user, Behrensen says. Researchers are expected to disclose AI use, check for mistakes and take responsibility, while the people who built or trained these models remain largely out of sight.
This creates an uneven situation. Students and academics are held to strict standards, while developers can release flawed or biased systems without facing the same level of oversight. If a tool hallucinates references or produces harmful outputs, Behrensen argues, the makers should also be held responsible, especially when real people are affected. They point to recent lawsuits involving AI chatbots that allegedly encouraged self harm, as well as deepfake campaigns that distort political discourse. Poor research has always existed, Behrensen notes. What is new is how easily responsibility can now be shifted. Errors that once would have been traced back to an author can now be blamed on “the AI”, leaving no one clearly accountable.
The myth of neutral data
Even when AI works as intended, its outputs depend entirely on the data it learns from. That raises a deeper ethical question: can a system trained on human history ever be neutral? As Behrensen puts it, this is “the classic garbage in, garbage out problem”. Amazon encountered this issue when it experimented with AI-based recruitment software. The system was trained on historical hiring data that favoured men. As a result, the algorithm downgraded women’s CVs. The project was eventually abandoned, but the lesson remained: a system designed to be objective can still repeat existing inequalities.
Similar patterns have appeared in predictive policing tools. These systems use historical crime data to identify so-called risk areas. Because those records already reflect social and racial inequalities, the algorithms often reinforce them. Reducing bias requires human effort and investment. To make AI less biased, Behrensen says, many people need to do expert work on the data. The question is whether companies that want to deploy AI everywhere are willing to pay for that work. Even then, Behrensen doubts full neutrality is achievable. Data, they explain, is always just a snapshot of reality. It can be made more balanced, but there are limits to how far that can go. For researchers, this means that using AI without understanding its training data risks quietly importing old biases into new scientific work.
Seeing the bigger picture
In the end, ethical AI use in research is not about drawing a single clear line. It is about recognising how many ethical choices are already made before a laptop is even opened, from energy use and labour conditions to data selection and accountability. AI will not replace science, but it is changing what responsible scientific practice looks like. Ethical AI use does not mean rejecting these tools altogether, nor embracing them without question. As Behrensen reminds us, AI systems are never neutral. They run on energy, on human labour and on data shaped by real histories.
Using AI responsibly means being aware of this and consciously deciding when AI truly adds value, and when it risks distancing us from the thinking research is meant to encourage. Responsibility, at least for now, remains human.
