ChatGPT turns off when asked for seahorse emoji

Is there a seahorse emoji? This simple question leads to a lengthy discussion on ChatGPT full of confusion and various emojis. However, there is no seahorse. Although ChatGPT isn't quite sure about that: "Wait, no, yes, here you go, no, not at all."

There can't be an emoji of a seahorse. That never existed. The competition, Gemini and Meta AI, know this. And what's more, they also know about the blog post that was written about the phenomenon. It is astonishing that ChatGPT has apparently not yet found it or internalized it sufficiently.

The background to this is that there have been several discussions on Reddit, Tiktok and other social media channels about whether the seahorse emoji exists or at least once existed. Some people seem very sure that the emoji did exist. It is very easy to take stock of the current situation using your keyboard. Lo and behold, no seahorse. But there never actually was one.

ChatGPT, in this case GPT-5, obviously has a similarly deceptive memory as some people and believes it can find such an emoji. It is surprising that the AI chatbot has such a hard time finding an answer. It also indicates that ChatGPT is obviously very keen to find a definitive answer rather than simply admitting that it doesn't know the answer. Claude doesn't fare much better when answering, by the way.

Mandela effect and the seahorse approach

Theia Vogel claims to be an independent AI researcher. She runs a blog where she also writes about the seahorse. In addition to the training material from GPT, which also includes Reddit, she has another human explanation: the Mandela effect. It has its origins in a study in which hundreds of people stated that they could remember the death of South Africa's president in 1980. However, Nelson Mandela only died in 2013. Anyone who thinks they can remember the phrase "Luke, I am your father" in the first "Star Wars" is also wrong. Darth Vader says in response to Luke's accusation that he killed his father: "No, I'm your father."

As Gemini and Meta AI know, the seahorse emoji is also part of the Mandela effect. In addition to the collective memory being wrong, Gemini also explains that AI models hallucinate the emoji and that the chatbot is aware of the discussions on Tiktok and Reddit.

ChatGPT's loop also seems to be because the chatbot at least wants to reproduce a seahorse emoji if it can't find it. This explains the many attempts to string together horses and unicorns and aquatic animals. The model searches for similarities and the next token to approximate the seahorse. But it doesn't work if the seahorse emoji doesn't even exist. Nevertheless, eventually, ChatGPT also manages to find the right answer and suggests using a combination:

(emw)