When artificial intelligence is given a choice between texts written by humans and those generated by another AI system, it tends to favour its own kind. This is the finding of a recent study by Charles University in Prague (Large language models favor communications generated by large language models).
More specifically, Large Language Models were asked to make purchasing decisions based on different types of text, including product descriptions from e-commerce websites, academic texts and film summaries. The models were not told whether the texts had been written by humans or generated by AI. The results showed a clear preference for AI-generated content. In the case of product descriptions, GPT-4 selected the AI-generated version in 89% of cases. The researchers were unable to provide a definitive explanation for this effect.
What does this mean for the use of AI in professional research?
The findings provide further evidence that human oversight remains essential when using the outputs of AI chatbots. It cannot automatically be assumed that AI-generated texts are more error-prone than human-written ones. However, a cycle in which texts are first generated by AI without review and are then used by other AI tools to derive insights or support decisions can increase the risk that errors remain undetected or are amplified.
By Markus Hoffmann
