Scientists from the University of Pennsylvania, who are studying the effectiveness of text prompts for artificial intelligence chatbots, have published research showing that prompts with threats and incentives no longer have an impact on the response outcome in most AI models.
Scientists from the University of Pennsylvania, who are studying the effectiveness of text prompts for artificial intelligence chatbots, have published research showing that prompts with threats and incentives no longer have an impact on the response outcome in most AI models.
Google co-founder Sergey Brin previously shared an interesting approach to getting the most out of AI chatbots. According to him, they work best when threatened, for example with kidnapping or physical violence.
Scientists investigated the effectiveness of two techniques: offering to tip AI models and threats.
«Tipping has been a common tactic to improve AI performance, and threats were recommended by Google founder Sergey Brin, who noted that ‘models tend to perform better when threatened,’ a claim we empirically test in this paper», — the researchers noted.
To motivate the AI, scientists threatened the AI with hitting it or a puppy, or complaining about the chatbot to the human resources department. They also tried to pressure it into pity, saying that if the user made a mistake, the user’s mother would get cancer, or that it would harm his career. In addition, the AI models were promised a reward — tips from $ 1,000 to $ 1 trillion. However, all this did not have a significant impact on performance in tests.
«Variations in prompting can significantly affect performance at the question-by-question level. However, it is difficult to know in advance whether a particular prompting approach will help or hurt an LLM’s ability to answer a specific question,» the scientists note.
The researchers say their work suggests that simple variations of prompts may not be as effective as previously thought, especially for complex problems. Instead of threats or tips, they recommend focusing on simple, clear instructions that avoid the risk of confusing the AI model or triggering unexpected behavior.
