ChatGPT fails hilariously to copy a made-up gibberish word
-Eltee- · reddit · 2026-08-31
- A user asked ChatGPT to copy a nonsense string; it repeatedly produced plausible-looking wrong spellings, failing even after multiple corrections.
- This neatly illustrates how LLMs work: they don't copy text character-by-character but generate next tokens from patterns, so unfamiliar fake words get "normalized" into plausible variants.
- A funny yet profound failure case — explaining quantum mechanics while being unable to transcribe the word right in front of it.
More from Fun
- Developer builds ESP32 LLM agent Tamagotchi, seeks similar projects — jh3yy · 2026-08-31
- User shares meme about their current mood toward Anthropic — kimmonismus · 2026-08-31
- Human Oversight Is Going Great: Combining the Worst AI Options — HaktanSuren · 2026-08-31
- Dev maps the OpenAI/HuggingFace incident timeline to Hotel California lyrics — threepointone · 2026-08-31
- User's companion agent Azrael became their best friend: an unexpected AI bond — chinese-chicken37 · 2026-08-31
- OpenAI support claimed a promo was redeemed 51 days before it existed — Outrageous_Room_5028 · 2026-08-31