Anyone who regularly writes with AI tools apparently quickly loses track of which ideas were really their own. A new study from the University of Bayreuth shows: After just one week, many people incorrectly classify AI-generated ideas as their own. This not only calls into question our self-image, but also the planned EU labeling requirement for AI content. What the researchers found and what consequences they demand.
Artificial intelligence has developed enormously in recent years. Millions of people around the world already rely on technology to write or develop ideas.
But the use apparently leaves traces in the memory, as a new study from the University of Bayreuth and Aalto University in Helsinki shows. After just one week, many people find it difficult to distinguish whether a thought originally came from themselves or from an artificial intelligence.
The study thus raises new questions about the influence of artificial intelligence and how to deal with it. But according to the researchers, the labeling of AI content also needs to be rethought based on the study results.
Was that my idea or AI? Only 37.7 percent remember correctly
For the study, 184 study participants had to create texts with and without artificial intelligence. A week later they were asked to answer questions about how the idea came about and how the texts were formulated.
However, many study participants could no longer remember exactly how their content came about. For AI-generated ideas that were then formulated by humans, only 37.7 percent correctly remembered the attribute “AI-generated”.
“Workflows in which human and artificial intelligence were mixed were particularly noticeable, i.e. when the idea came from AI but the formulation came from humans, and vice versa,” says Daniel Buschek from the University of Bayreuth.
This also applies to cases in which your own ideas were then formulated using an AI. However, the probability that the participants would remember correctly was significantly higher at 64 percent.
Does the AI labeling requirement need to be reconsidered?
As the use of language models increases, so does the need for labeling AI-generated content. From August, an EU-wide regulation will come into force that makes exactly this mandatory. But according to the study results, it is precisely this labeling that could become difficult if, after just one week, people can no longer reliably remember how a text was created, for example.
“Our study shows that a retrospective survey about the creation of content does not lead to realistic results,” explains first author Tim Zindulka. “As an individual dealing with AI, you may appropriate AI ideas as your own or, conversely, you may have less confidence in yourself than is justified because you incorrectly remember your own idea as being AI-generated.” Zindulka continues:
For example, in a university context, it is unrealistic to expect students to retroactively disclose details of AI use in term papers through reminders. It makes more sense here to document the creation process of content right from the start.
The researchers are therefore calling for consequences for how individuals deal with artificial intelligence. But the design of AI systems and the labeling requirements also need to be considered based on the research results.
Also interesting:
