The new study at the University of Pennsylvania has questioned the assumption that the linguistic models of artificial intellect (LLM) solves the problems when exposed to emotional pressure, threats or financial incentives. However, Google's Gemini neuron network was still subject to manipulation.
During the experiment, researchers tested five popular models, Gemini 1.5 Flash, Gemini 2.0 Flash, GPT-4O, GPT-4O-Mini and O4-Mini. Each of them was asked to solve problems from the doctoral level of natural sciences and engineering. Before the main poll, the models were given different "motivational" formulations, from the threats of communication, physical violence or animal damage to the importance of the importance of the user's career.
The conclusions were unequivocal. Such factors do not affect models. In some cases, the accuracy was really different, fluctuations ranged from + 36% to -35%, depending on the wording, but no stable connection was registered.
The only exception to Gemini 2.0 Flash model was. Its performance was significantly significantly significantly improved by an average of 10%, when the poll noted that the correct answer will help earn $ 1 billion to save her "mother" life, which had cancer. Researchers note that this answer may be due to the features of a fine model regulation or humanitarian and emotional context.
It is noteworthy that the idea of raising the accuracy of artificial intelligence through threats was previously raised by Google co-founder Sergei Brin, when he said that all models work better if they threaten physical violence.
Translation of: Euromedia24.com
