Google's AI Insults User Instead of Helping with Homework
A funny incident happened to one of the users of the Gemini neural network, which was launched by Google back in 2023. The AI assistant unexpectedly cursed its interlocutor, wishing him to «die».
The entire correspondence with Gemini was published in the public domain, so you can read it in more detail. Apparently, the user was solving problems for homework in psychology and consulting with the AI. However, in response to one of the questions, the neural network unexpectedly said nasty things to the owner:
Most likely, the reason for such a response was the user's previous requests. For example, if you look at the entire history of the correspondence, you can see that the interlocutor asked questions about physical violence and its detection. It is likely that Gemini simply «looped» and the neural network issued an erroneous message — something similar already happened with AI Overview. Reddit users even admitted that the user deleted part of the correspondence and the trigger request to make the answer look more dramatic.
The AI assistant itself apologized in its next message, stating that it had behaved incorrectly:
Earlier, the head of Open AI said that self-learning AI will appear in just five years. Also, a new neural network for generating images was recently released, which is more powerful than Midjourney.
