Danish university policies on generative AI: Problems, assumptions and sustainability blind spots

The 'black box' nature of many GenAI systems means there is significant uncertainty about how user data is actually utilized for training or profiling, and whether it is shared with third parties.

“The”

The development and operation of GenAI carry significant environmental costs, including high energy and resource consumption, and raise concerns about labor exploitation and job displacement.

“The development and operation of GenAI carry significant environmental costs, including high energy and resource consumption, and raise concerns about labor exploitation and job displacement.”

GenAI systems inherit and amplify existing biases in their training data, leading to unfair or discriminatory outcomes.

“GenAI systems inherit and amplify existing biases in their training data, leading to unfair or discriminatory outcomes.”

GenAI causes environmental damage, with CO2 emissions and water consumption particularly concerning.

“Finally, genAI causes environmental damage, with CO2 emissions and water consumption particularly concerning (Carrigan, 2024;de Fine Licht, 2024;Driessens & Pischetola, 2024).”

OpenAI created ChatGPT by extracting 300 billion words from online sources without formal consent.

“On the planetary scale, OpenAI created ChatGPT by extracting 300 billion words from online sources without formal consent (Driessens & Pischetola, 2024).”

In the study context, genAI's environmental impact was not discussed in university policy.

“In our study context, as in other international settings, genAI's environmental impact was not discussed in university policy (Dabis & Csáki, 2024;Driessens & Pischetola, 2024;Erhardt et al., 2025).”

AI models, trained on massive datasets reflecting 'long-standing biases existing in society,' can perpetuate undesirable stereotypes.

“AI models, trained on massive datasets reflecting”

Driessens and Pischetola [90] found that the huge planetary costs of GenAI and the exploitative business models and practices are mostly left unproblematized in the policies.

“Driessens and Pischetola [90] also found that the huge planetary costs of GenAI and the exploitative business models and practices are mostly left unproblematized in the policies.”

Driessens and Pischetola [90] and Luo [213] adopted Bacchi's 'What's the problem represented to be?' framework.

“Driessens and Pischetola [90] and Luo [213] adopted Bacchi's 'What's the problem represented to be?' framework [29]”

Driessens and Pischetola [90] identified three key problematisations in the analyzed policies: assessment integrity, legality of data, and veracity.

“Driessens and Pischetola [90] identified three key problematisations in the analyzed policies: assessment integrity, legality of data, and veracity.”

Even seemingly innocuous text input can allow LLMs to infer user characteristics such as age, gender, income, and location, posing re-identification risks.

“Even seemingly innocuous text input can allow Large Language Models (LLMs) to infer user characteristics such as age, gender, income, and location, posing re-identification risks.”