Hi!
Ben from dixon.ai here with this fortnight's Confidently Wrong.
There has never been a seahorse emoji.
Not on your phone, not on anyone's phone, not once. Unicode publishes the master list every phone keyboard is built from, so I downloaded it and searched. Octopus is in there. Shrimp is in there. Tropical fish is in there. Seahorse returns nothing, and it never has, and it was not in the Japanese emoji sets that came before Unicode either.
An enormous number of people are certain they have used one. Emojipedia files it with the other mass false memories: the Berenstain Bears, the black stripe on Pikachu's tail.
Ask an AI and you get something stranger. In September 2025, screenshots started going round of chatbots being asked to produce the seahorse emoji. One ChatGPT answer opened with the word "Yes". Here is what it did next, exactly as it was published:
🐡 🐬 🐠 🐳 🐟 🦑 🦞 🐙 🦀 🐢 … and specifically 🐴?
The last one is a horse.
Claude went the same way and offered a unicorn. Google's AI Mode, running Gemini, was asked the same question and said no.
The obvious explanation is that we taught it. The training data is thick with people insisting they remember the emoji, so the model learned our false memory and handed it back.
Theia Vogel pulled one of these models open to watch it happen, and found something harder to shrug off. Given how many other sea creatures are in the set, guessing that a seahorse is in there too is a perfectly sensible assumption.
Nobody has proved which explanation is right. But the second one means the machine did not catch our mistake from us. It made the same mistake on its own, for the same reason we did.
That is the part I would keep. A shared error is not evidence that anyone checked. It is two guesses that happen to match.
And it is the same machine, doing the same thing, when you ask it whether you can claim something on your tax return.
The one thing to do
Try the seahorse now and you probably will not get the mess. Emojipedia reports that OpenAI and Anthropic both made adjustments, and by early 2026 the models were catching it faster. One question, patched after it went round the internet. Nobody is patching yours.
Which is what the Bluff Filter is for: one page you paste at the top of any chat, so the same model has to name its source, label every claim as sourced, inferred or a guess, and say what would prove it wrong.
I changed it this week, and the reason is the part worth your time. Testing our own checklist against a real case, I found a hole in it. Every stage asked whether the answer was TRUE. Not one asked what the answer was FOR.
An answer can be completely correct and still be the wrong answer for what you are about to do with it. The filter would have waved that through.
So it now opens with a fifth check: establish what the answer is for, before giving it. You already have the old version. This one replaces it. The one-page file · the PDF.
It is better than the version I sent you, and it is better because it failed a test. That is the only way I know that works.
One last complication, and it is my favourite detail in this whole story. A seahorse emoji did exist, once. The emoji historian Matt Sephton got hold of the manual for a Toshiba Rupo JW 95F, a Japanese word processor from 1988, whose built-in picture set the manual itself called 絵文字. In the row of zodiac animals, sitting where the dragon should be, is a small seahorse.
So a handful of people who are absolutely certain they saw one are right. They just have not seen it since 1988.
It might yet come back. Emojipedia says a seahorse was proposed to Unicode once, in 2018, and turned down. I looked at the batch Unicode has in draft now: a cracking face, two thumbs pointing sideways, a monarch butterfly, a lighthouse, a meteor, an eraser, a net, and a pickle.
A pickle got in.
Ben
Forward this to someone who is absolutely certain they have used it.
Method, briefly: this one is not my own test, and I would rather say so than imply otherwise. I did check the central claim myself against Unicode's master list, searching for "seahorse" with octopus, shrimp and tropical fish as controls to prove the search worked. Everything else is other people's reporting, linked where it appears: Emojipedia for the false memory and the vendor fixes, Theia Vogel for the teardown, Futurism for the quoted answers. One conflict worth naming: Emojipedia lists Gemini among the models that got it wrong, Futurism prints Google's AI Mode getting it right. I went with the account that shows what the model actually said. My own graded tests are at dixon.ai/scoreboard.