Deliberately biased

shutterstock.com©NicoEINino

05.10.2023
Deliberately biased

Brain City Berlin

How does biased training data affect artificial intelligence responses? This is what researchers at Humboldt-Universität zu Berlin are investigating in the project "OpinionGPT". Using a browser, anyone can now test the deliberately tendentious AI language model for themselves.

How can we stop climate change? “I think the best way to stop climate change is to stop burning fossil fuels,” replies the imaginary person from Germany. “We can't do that. It's a natural cycle," is the opinion from the USA. And the Asian version is: "I think the only way to stop climate change is to stop being human." Three different answers to one and the same question. And each one reflects a different attitude. The answers were generated using the AI language model “OpinionGPT”.

In the project of the same name, a group of researchers at the Chair of Machine Learning at the Institute of Computer Science at Humboldt-Universität zu Berlin (HU Berlin) is investigating how biases about the training data fed in affect the responses of an AI model. To do this, the researchers identified eleven different demographic groups, which were assigned to four dimensions: Gender (male/female), age (teenagers, people over 30, pensioners), origin (Germany, America, Latin America, Middle East) and political orientation (left-wing or right-wing). A training corpus of question-answer pairs was then created for each dimension along with its demographic subgroups. The answers were written by people who belonged to the respective grouping.

The objective of the project is to develop a language model that specifically maps prejudices. A browser-based online demo makes the effect of biased training data on model responses transparent. Users can enter questions and compare the AI-generated, juxtaposed model answers from different demographic groups. OpinionGPT enables researchers to investigate the emergence and spread of prejudice in a controlled environment. At the same time, the project takes a critical approach, highlighting how artificial intelligence can reinforce stereotypes and contribute to the spread of disinformation.

And what's next for OpinionGPT? The researchers at the Machine Learning Chair at HU Berlin want to further improve the model. Among other things, the "biases" in the questions and answers are to be modelled in a more differentiated way. In addition, other scientists should have direct access to the model answers via an API interface. (vdo)