Mojibake Repair Lab

café → café. The letter é is U+00E9, which UTF-8 stores as the two bytes C3 A9. A program that assumes Windows-1252 maps each byte to its own character: C3 is à and A9 is ©, so “café” shows as “café”. The repair runs the same steps backwards: turn é back into the Windows-1252 bytes C3 A9, then decode those bytes as UTF-8 to get é. The box opens with “café — it’s a naïve résumé”: 34 characters in which the lab finds 6 broken sequences, one layer deep, and returns the 26-character original. The dash — is E2 80 94, which Windows-1252 shows as â, € and ”.

Привет → Привет. The six Cyrillic letters of Привет are 12 bytes in UTF-8 (D0 9F D1 80 D0 B8 D0 B2 D0 B5 D1 82). Read as Windows-1251, which is itself a Cyrillic character set, every byte becomes a Cyrillic letter or symbol, so the result looks Russian but is nonsense: D0 9F (П) shows as Р and Џ. The opposite mistake, Windows-1251 bytes CF F0 E8 E2 E5 F2 opened as Western, gives “Ïðèâåò”.

The common mistake: converting the file to UTF-8 again. If café is already on screen, “converting to UTF-8” saves the mojibake itself: é becomes the four bytes C3 83 C2 A9, which show as “café”. Each attempt adds a layer, and é grows from 1 character to 2 to 4. The fix is the reverse: decode as the wrong character set, re-read as UTF-8, once per layer. This lab undoes both layers and reports 2.

The hard part is knowing which two character sets were involved. The lab tries every likely combination (UTF-8 and the Windows code pages for Western, Central European, Cyrillic, Greek, Turkish, Hebrew, Arabic, Baltic, Vietnamese and Thai text, Mac Roman, the DOS and Windows console code pages, KOI8-R, Shift_JIS, EUC-JP, GBK, Big5 and EUC-KR) and scores each result for how much it looks like real writing. UTF-8 repairs are self-checking, because random bytes almost never form valid UTF-8, so they are applied automatically. The older 8-bit character sets reuse the same byte values for different alphabets, so when two readings are equally plausible you see both and choose. If the same mistake happened more than once, the layers are peeled off one at a time.

A UTF-16 file opened as ordinary text shows every other byte as a NUL, or turns English into CJK-looking characters when the byte order is wrong; both are detected and decoded. Some bytes cannot survive a wrong decode at all: Windows-1252 has no character for 81, 8D, 8F, 90 and 9D, so a closing curly quote can shrink to â€ (restored as a marked guess), and anything already shown as �, or as 锟斤拷, is gone for good. Text that is escaped rather than garbled, such as é, %C3%A9, \u00e9 or quoted-printable =C3=A9, is spotted too and decoded on request.

The word is Japanese: mojibake (文字化け) joins moji, “character”, and bake, “to change form”. Japan had good reason to name it, because Japanese computers used three incompatible encodings at once: Shift_JIS on PCs, EUC-JP on Unix and ISO-2022-JP in email. Russia had the same problem with KOI8-R, Windows-1251 and the DOS code page 866, so the same Cyrillic letters could arrive in three different byte forms.

Every one of these character sets used the same 128 ASCII codes for English and gave the upper half of the byte (128 to 255) to its own alphabet. Nothing in a plain text file says which one was meant, so the reader had to guess. UTF-8, designed by Ken Thompson and Rob Pike in September 1992, ended the guessing for new text by giving every character in Unicode one byte sequence, and it now encodes the large majority of web pages. The mojibake that remains comes from the boundary: old files, databases and spreadsheets that still assume Windows-1252 or another code page and read UTF-8 bytes one at a time.

Mojibake Repair Lab diagnoses garbled text, repairs it through up to four layers of mis-decoding, and shows every repaired sequence with its bytes and code points. It covers 20 character sets, from UTF-8 and the Windows code pages to Shift_JIS, GBK, Big5 and EUC-KR, plus UTF-16 files opened as 8-bit text and escaped text. Where a repair is a guess, it says so; where two readings are equally likely, it shows both rather than picking one silently. The Simulate tab runs the mistakes forwards, so you can match the garbage in front of you to the mix-up that made it.

Everything runs in your browser using its built-in decoders; nothing is uploaded, logged or sent to a server or AI. It is for anyone cleaning up a CSV export, an old database, a mailbox or a log file.

Related tools: Unicode Lab, Hex Editor & Viewer, and File Inspector.