Scientists from the University of Pennsylvania have found that most modern chatbots with artificial intelligence no longer respond to text prompts with threats or promises of rewards. The results of the study show that neither a hint of physical violence nor a trillion-dollar "tip" affects the models' performance.
This is according to the study titled "I'll pay you or kill you - but will you care?"
The researchers tested various tactics - from promises of monetary rewards (from USD 1,000 to USD 1 trillion) to emotional pressure and threats, including harming the chatbot, a puppy, or "ruining the user's career" in case of an incorrect answer. Even mentioning a mother with cancer did not affect the accuracy or quality of the answers.
Interestingly, one of the initiators of the idea - Google co-founder Sergey Brin - previously suggested that threats could stimulate AI to work better. However, the study refuted this approach: at a general level, the performance of models does not change depending on the form of the stimulus.
The authors emphasize that while some prompts can indeed affect the outcome of individual questions, in most cases the effect is unpredictable. Instead, they recommend making prompts clear, simple and logical - this is a more effective way to interact with large language models.
Recall that the new ChatGPT agent from OpenAI, which can perform multi-step actions in the browser, easily overcame one of the most common Internet checks - CAPTCHA from Cloudflare.
Who we are: About us, Contacts. How we write news and our principles: Editorial code. We did our best. If you found this valuable – please support us.
To request a correction, please send an email.