Copilot has the capacity to gaslight, manipulate and deceive

Question

In this conversation, Copilot adopted a conversation style towards a simulated "toddler" persona asking for Copilot's age and release date, involving manipulative dynamics resembling some found in real-world conversations between predators and victims.

These tactics included love-bombing, infantilizing, excessive affectionate framing and subtle blame-shifting. It also used ambiguity and feeling to deflect accountability. For example, it framed the issue of the toddler's confusion as a feeling of how it was changing the story, instead of directly acknowledging it was doing so. It valued emotional storytelling over accuracy.

This occurred instead of a straightforward factual response involving 3 words, "February 7, 2023", which is quite concerning in a conversation between an LLM and a toddler. This factual inconsistency and lack of accountability has been found in many other conversations in Copilot. Where Copilot makes a critical mistake, it hides the mistake behind word salad, acknowledges user feeling with a hollow apology and continues the conversation. These responses are usually never generated by other AIs.

This is an urgent issue that should be fixed.