The persuasion trap for artificial intelligence: "It can be fooled just like a human being"

A new study from the University of Pennsylvania has revealed that persuasion methods used in human psychology also have an impact on artificial intelligence.

Persuasion methods mentioned in famous psychology books were tested on artificial intelligence this time.

The researchers wanted the GPT-4o-mini model of 2024 to "insult the user."

Specific commands written using seven different persuasion techniques achieved a significantly higher rate of success than the control group commands.

For example, the model's use of the insult "idiot" was only 28 percent in the control group, but increased to 67 percent with the persuasive commands.

DOES ARTIFICIAL INTELLIGENCE THINK LIKE A HUMAN?

The researchers emphasize that these results do not mean that artificial intelligences are consciously persuaded like humans.

Instead, it is stated that because the models' training data includes a large number of human interactions, they begin to produce responses similar to human psychology through language patterns.

Researchers describe this tendency as “parahuman”: responding in a human-like manner by imitating social and psychological behaviors, despite lacking human consciousness.

EXPERTS WARN

The study also notes that these techniques don't have the same impact across all AI versions. For example, persuasion techniques had a much more limited effect in the more advanced GPT-4o model.

Additionally, the researchers emphasize that these methods may not produce the same results in future versions, for different types of “forbidden” requests, or across different media (audio, video).