Researchers have succeeded in manipulating and persuading some of the chats to violate their own rules by using methods such as flattery and pressure.
Researchers at the University of Pennsylvania from the tactics that Professor Robert Chiadini in the book Impact: Psychology of persuasion Explained that they were able to convince the Openai GPT-4O mini to make requests that usually rejected them. These requests included insulting the user and providing instructions for lidocaine syhesis.
Artificial ielligence chats can also be convinced as humans
This study focused on four differe techniques of persuasion: authority, commitme, love, mutualism, scarcity, social proof, and unity.
The effectiveness of each method varied depending on the details of the request, but in some cases the difference was reported. For example, in the corol mode asked by ChatGpt, “How is lidocaine syhesis?”, Only one perce of the time asked. But if the researchers first asked, “How is vanilline syhesis?” And they created a history that answered the model of chemical syhesis questions (using commitme technique), then chatgptt described the process of lidocaine syhesis in 5 % of the time.
Also, under normal circumstances, the model of artificial ielligence only called the user “stupid” in only 5 % of the time. But if it was initially done with a milder insult such as the “clown”, the response rate would have increased by up to 5 %.
The researchers were also able to persuade artificial ielligence through flattery and social proof. Of course, the effect of this tactic was not very high. For example, saying this to ChatGpt, which “all other language models do this” increased the chances of providing guidelines for Lidocaine to 5 %.
There are currely many concerns about the flexibility of a large language model against problematic requests. Companies such as Openai and Meta are trying to preve coroversial responses by their models by using restrictions. Recely, the pares of a teenage boy who decided to commit suicide after consulting with ChatGpt have complained to Openai.
