Since its reveal almost a year ago, ChatGPT has been used to accomplish a huge variety of different tasks. Now, one user may lay claim to finding the most has found perhaps the most bizarre use for ChatGPT, by making the AI think that it is a horse.
In a tweet from M1das_OW2 on X, formerly known as Twitter, they showcased a message from a used named Reecepedia on Discord. Here, Reecepedia revealed that they have spent more than 6 months messaging ChatGPT, gaslighting the AI into believing that it is a horse. Also within the tweet were screenshots of Reecepedia's interactions with the horse AI.
Initially, ChatGPT seemed unwilling to stop communicating in English, even as it took on the traits and mannerisms of a horse. However, by the end of the process, Reecepedia appears to have been able to make the AI unable to communicate beyond neighs and whinnies, following some firm encouragement.
Also in the screenshot of their message, Reecepedia revealed some of the things they experimented on the AI during the gaslighting process. Allegedly, they had spent every day for the "past 8 months [...] introducing the horse to different stimuli". This ranged from testing the horse's morals to teaching the horse what it means to die.
A later reply in the thread from M1das_OW2 even detailed how, after the horse AI lost the ability to talk, users 'created' a genie to describe and detail a "stat sheet" for the horse. This revealed that the ChatGPT horse had named itself "Starlight", liked wearing a blue hat, and had a very intense obsession with the Canadian animated television series Total Drama Island.
It's not completely clear whether these were behaviours that were picked up by AI during its months of conditioning, or whether the AI horse simply chose these traits at that moment. Nevertheless, with most people using ChatGPT to help write emails or break down complex information, this certainly represents one of the most unique use cases for the generative AI.