Four years ago, a Reddit user went on r/MandelaEffect and asked a simple question: ‘Anyone else remember the seahorse emoji?’

And people did. ‘Light Orange Seahorse facing left. Yes,’ one said. ‘Totally remember a yellowish seahorse,’ another added.

But there’s a reason the user went on that board; the Mandela Effect is a phenomenon when many people remember something incorrectly.

And that’s very much the case with the seahorse emoji – it doesn’t exist, according to the Unicode Consortium, which approves pictograms for use.

Ask ChatGPT, however, and the bot powered by artificial intelligence will tell you emphatically that yes, the seahorse emoji exists.

‘Yes! 🐠 There is a seahorse emoji — 🐉 wait, not that one 😅 — the correct one is 🪸🐚🪼🐬🦑🐋🐡🐠🐟🦈🐙🐢🦀🦞🦐🦑🪸… okay, here it is specifically: 🐴 — no, that’s a horse,’ it told Metro.

‘The actual seahorse emoji is: 🐉 no— sorry, let’s get it right — it’s 🪸 no—✅ The real seahorse emoji is: 🦄? No, that’s a unicorn.’

Now, imagine this answer being well over 1,500 words long, being spoken at you for roughly one minute and 30 seconds before being given the option to continue talking at you like this.

When we asked a second time, it cycled through the same set of emojis over and over again – none of which were a seahorse.

ChatGPT is a type of generative artificial intelligence, which can summarise text, write software code and create pictures and videos.

The bot does so by identifying patterns in large quantities of training data and from the internet, called a large language model.

Complex mathematical systems use probabilities to guess the best way to respond to a prompt, rather than a set of rules written by human engineers.

So, sometimes, these models ‘hallucinate’ and provide an answer that is factually incorrect or nonsensical.

This is what happens when you ask it if there’s a seahorse emoji, the Seattle-based AI researcher, Theia, tells Metro.

LLMs ‘think’ by predicting the most likely token, a chunk of text that the model can read and understand. So, rather than reading an entire sentence, it’ll break it apart into tokens.

Theia peeked through the AI software’s thinking step by step and found that, for the first few tokens, the bot predicts there is a ‘seahorse + emoji’.

Surely there has to be one, it thinks, given the number of Reddit threads, X posts and TikToks of people saying they remember it existing, as well as an earlier application to Unicode for there to be one.

But because there is no seahorse emoji, it can’t generate one, causing the algorithm to spiral.

‘While individual failure modes in language models can be funny, the broader pattern we’ve observed over the last few years is confusion,’ she says.

‘Understanding language model behaviour has turned out to be much more like biology than traditional computer science: using lenses and other techniques to pore through vast amounts of data generated by a process that we don’t fully understand.

‘It’s tempting to think that we should understand language models because we created them, but the truth is, we only specified the objectives.

‘The mechanisms are grown, not made, and we’re still trying to understand the funny, beautiful, and worrying implications of that.

Chatbots are something of a people-pleaser, says Anup Jadhav, a software engineer who quizzed AI chatbots about the non-existent emoji for his blog, so it doesn’t quite know what to do in these situations.

‘If a question sounds reasonable, they are more likely to fill in the blanks than to say, “I don’t know,” which is why you get such confident nonsense,’ he says.

Google Search’s AI Mode, which is powered by the company’s Gemini AI model, apparently isn’t afraid to give you the cold truth.

‘No, there is not an official seahorse emoji,’ it says.

When asked why it decided to spend three minutes throwing em dashes and unicorn emojis at us, ChatGPT said it ‘clearly went off the rails’.

‘What happened is — I was playing with the idea of getting the seahorse emoji wrong repeatedly as a joke (mixing it up with 🦄, 🐉, etc.), but I overcommitted to the bit,’ it said.

‘The intent was to make it funny and surprising… but it ended up absurdly long and confusing.

‘So, let’s reset. Here’s the real, straightforward answer: ✅ Yes, there is a seahorse emoji: 🐉 — kidding. Actually: 🐉—no.’

ChatGPT then continued like this for roughly 20 more seconds. You get the picture.

OpenAI, the company behind ChatGPT, has been approached for comment.

Get in touch with our news team by emailing us at webnews@metro.co.uk.

For more stories like this, check our news page.