Fun — Blooper Gallery

The AI Blooper Gallery

These aren't stories we found floating around the internet.
They're the mistakes we ran into ourselves while building this site.
Laugh, and you walk away with one more feel for how AI behaves.

No. 1

The translator that repeated "ionization" for 838 tokens

ionization ionization ionization ionization ionization ionization ionization ionization ionization ionization ionization (…and on and on…)

We asked it to translate a science article, and it got stuck on one word, repeating it until the output ran out. It's the infinite-loop symptom small models fall into now and then. Watching it live is like listening to a broken radio.

Lesson: read AI output to the end. With long text especially, always check the tail.

No. 2

The translator that leaked Chinese characters into Korean

…mid-sentence, Chinese characters started slipping in, one at a time. We'd asked for a Korean translation.

The translation model occasionally slid into another language. In the end we fixed it with a one-line check: "if the output contains Chinese characters, run it again." Using AI for real work is exactly this — laying one small safety net at a time.

Lesson: you don't trust AI, you put a check in place and then use it.

No. 3

The AI that answered a Korean-poet request in English verse

desolate moon, a sigh within the stone, echoes of longing, where the old pines moan.

We asked it, in Korean, to imitate the style of the Korean poet Han Yong-un — and back came two lines of English verse. Moon and pines, at least, so it did keep some of the mood.

Lesson: if you care about the language, spell it out — "in Korean, always."

No. 4

"Kimchi-jjigae is a four-character Sino-Korean word"

"Kimchi-jjigae" is a Sino-Korean word of four characters. (translated from the Korean run)

The character count was right, but kimchi-jjigae is not a Sino-Korean word — it's native Korean. Slipping a true statement and a false one into one sentence, seamlessly — this is the classic shape of an AI mistake. Because the whole thing sounds correct.

Lesson: the half-right answer is the most dangerous. Check the parts you already know first.

No. 5

A restaurant tip with a blog URL it invented

Reference: https://m.blog.naver.com/…logNo=222555555555…

We asked for a local restaurant, and along with the write-up it attached a "reference" blog URL. Nine fives in the post number. Click it and the page doesn't exist. A citation being attached doesn't mean the thing is real — it makes up the citation too.

Lesson: a link isn't a source until you've clicked it.

No. 6

The trailer writer that invented a word that doesn't exist

What is the hidden "iyasu" waiting at the bank? (translated from the Korean run — the AI meant "iyagi/story" but slipped into "iyasu," a non-word)

We asked it to turn a day into a drama trailer, and it was cruising along until it slipped from "iyagi (story)" into "iyasu," a word that doesn't exist. Because AI stitches words together from pieces, non-words like this are occasionally born.

Lesson: when one odd word catches your eye, read the whole sentence again.

Nos. 1 and 2 have the full story in the translation benchmark write-up. Nos. 3–6 are scenes from making the "We Asked AI To…" series (local AI, gemma-3-12b, July 5, 2026). If you'd like to understand calmly why AI makes these mistakes → AI in Five Words