By Peter Witten, AAU Communikation and Public Affairs
Foto: DFF

"I’m proud. It’s a huge recognition of my research," says Professor Johannes Bjerva from the Department of Computer Science at AAU.

He has been awarded Sapere Aude, one of the most prestigious research grants in Denmark. The grant for research leaders amounts to nearly DKK 6.2 million and is awarded by the Independent Research Fund Denmark.

Could cost lives

The millions will be used to delve into something that could potentially cost human lives.

Johannes Bjerva’s research project, TRUST, is based on the growing difficulty of distinguishing whether what we read on a screen is written by a human or by a language model (LLM – Large Language Model), artificial intelligence.

When we use tools like ChatGPT, many people know that sometimes the answers can be incorrect.

However, fewer are aware that the underlying AI technology is vulnerable and can be manipulated or “poisoned” through hacking. This can lead to misinformation - or worse.

Critical functions

AI and language models are now used for much more than just writing essays with ChatGPT.

Today, language models are integrated into critical functions, such as in the healthcare sector. This means that hacking attacks could have serious consequences for patient treatment and, in the worst case, cost lives.

Concerns about system reliability and security extend beyond healthcare equipment to education and many other areas of society.

"Sleeper agent"

The concern is heightened by the possibility that infected systems or language models might not begin to manipulate or misinform until later - like a kind of "sleeper agent" that suddenly activates, Bjerva explains.

The TRUST project aims to prevent this. It will contribute to fundamental research with the goal of uncovering hidden patterns in AI-generated text by studying linguistic structures, such as sentence structure, word forms, and word meanings.

By identifying variations in text, the project will develop methods to detect both artificially generated text and cases where language models have been manipulated by malicious actors.

One of its kind in Europe

"We can be influenced in subtle ways. The core problem is if someone manages to insert a structured bias that can influence us over time without us realizing it," says Johannes Bjerva.

The TRUST project will run for four years and will examine the issue across multiple languages and independently of specific language models.

According to Bjerva, there are no similar research projects in Europe, and TRUST will collaborate with the Department of Linguistics at Stockholm University.