Corrupted Text Generator

Real corrupted text is not random noise. Each kind of damage has a cause, and the cause decides what the reader sees: a lost byte gives a �, a wrong encoding gives à and â€, a flipped bit swaps one letter for its neighbour in the ASCII table. The styles here copy those patterns, so a “recovered file” in a game or story looks like the real thing rather than a random mix of symbols.

Only the Corrupted style mixes faults. If you want the damage to tell a story, pick the one that fits it: mojibake for an old email archive or a badly exported spreadsheet, broken bytes for a half-downloaded file, bit flips for failing hardware or a transmission from deep space, escaped bytes for a programmer’s log.

Mojibake (Japanese for “character transformation”) is what happens when text saved in one encoding is read in another. The common case today is UTF-8 read as Windows-1252. In UTF-8, é is two bytes, C3 A9. Windows-1252 has one character per byte, so it shows C3 as à and A9 as ©, and café becomes café. The curly apostrophe ’ is three bytes, E2 80 99, which Windows-1252 reads as â, € and ™.

Plain ASCII letters are the same in both encodings, so mojibake only bites on accented letters, curly quotes, dashes, symbols and emoji. That is why the Mojibake row has no damage setting, and why the generator first turns straight quotes into curly ones as Word does when you type. Without that step an English sentence would usually come through untouched.

Worked example. “Record 0417: Don’t open the door – it’s still inside.” has three characters outside ASCII: two apostrophes and an en dash. Read as Windows-1252 it becomes “Record 0417: Don’t open the door – it’s still inside.” The other 41 non-space characters are unchanged, and the line grows from 53 to 59 characters because each of the three became three.

The right double quote ” leaves only “— on screen because its last byte, 9D, is not a printable character in Windows-1252; browsers decode it as an invisible control character, and it is often lost when the text is copied. The last row shows mojibake applied twice: each pass makes it longer, and à followed by Æ, ƒ or  is the sign that it happened more than once.

“Latin-1” and “ISO-8859-1” behave the same way in a browser. The WHATWG Encoding Standard, which all current browsers follow, treats those labels, and even “ascii”, as names for Windows-1252.

� is U+FFFD REPLACEMENT CHARACTER, the symbol a decoder writes when it meets bytes that are not valid in the encoding it was told to use. Unlike mojibake, the original is gone: the program has already thrown the bad bytes away and kept only the marker. In UTF-8, � itself is stored as EF BF BD, so a � that is then garbled as mojibake shows up as �.

Unicode’s recommended practice, followed by browsers, is one � for each maximal subpart: the longest run of bytes that could have been the start of a valid character. So a file cut off in the middle of € (E2 82 AC) ends in a single �, not two, and a lone é saved in Windows-1252 (byte E9) and read as UTF-8 gives “caf�”.

The Broken bytes row replaces each hit character with one �, whatever its length. That matches what you see when a single byte inside a character is damaged; real files with long damaged stretches can show many � in a row.

Every ASCII character is a 7-bit number, and flipping any one bit gives another character a fixed distance away in the table. That makes bit-flip damage look strangely regular: some letters swap case, some turn into digits or punctuation, and some become control codes that have no visible form. The Bit flips row shows control codes as the Unicode Control Pictures ␀ to ␟ and ␡, so they can be seen and copied.

Worked example. e is 0x65, binary 1100101. Flipping each of its seven bits in turn gives d, g, a, m, u, E and %. Bit 5 is worth 32, the exact gap between capitals and small letters, which is why a flipped e so often becomes E.

Bit flips happen outside fiction too. In 2011 Artem Dinaburg registered domains one bit away from popular ones, such as mic2osoft.com (r is 0x72, 2 is 0x32, so bit 6 differs), and logged 52,317 requests from 12,949 IP addresses over about seven months from machines whose memory had flipped a bit in a domain name. He called it bitsquatting; most consumer computers and phones have no error-correcting memory to catch it.

The Redacted row replaces whole words with one █ per letter, so the length of each hidden word still shows, as it does on a real blacked-out page. Punctuation and spaces stay where they were. It is meant for case files, puzzle props and fictional dossiers.

Do not use it to hide anything real. You choose the damage level, not which words go, so a name can survive, and the letter count of a hidden word can be enough to guess it. To redact a real document, use the redaction feature of a PDF editor, which removes the text underneath; a black box drawn on top leaves the words in the file, where they can be copied out.

Mojibake can usually be undone, because no information was lost: the bytes are still there, just shown with the wrong character set. Paste text like “it’s “fresh†and café costs £3” into the box and a Repaired text row appears with “it’s “fresh” and café costs £3”. It reverses UTF-8-as-Windows-1252 up to three times, so double mojibake (é) comes back too, and it restores a bare †as ”, the case where the invisible last byte was lost. Text that is already correct, such as café or naïve, is left alone.

Better still is to stop it at the source. A CSV that shows é in Excel was saved as UTF-8 and opened as Windows-1252: import it through Data, From Text/CSV and set File origin to 65001: Unicode (UTF-8), or save the file as “CSV UTF-8” in the first place. On a web page, declare <meta charset="utf-8"> in the first 1,024 bytes. For large amounts of text in Python, the ftfy library detects and fixes mojibake across many encodings, not just this one.

� and bit flips cannot be fixed from the text alone. The original bytes were replaced or changed, so the only real fix is an older copy of the file, or a format with a checksum or error correction that can rebuild it.

Damage is the chance that each character is hit, so at 30% roughly three characters in ten change, though a short line can land above or below that. Spaces and line breaks are never touched, which keeps word shapes visible even when most letters have gone. Each position has its own fixed roll, so raising the slider only adds damage and typing more text does not reshuffle the start. New pattern picks a different set of rolls.

Worked example. The sample sentence “Record 0417: Don't open the door - it's still inside.” has 44 characters that can be hit. In the Corrupted row on the starting pattern, 10% hits 5 of them and leaves it easy to read; 30% hits 16 and gives “R�od 04�7: Don't op�n ▒ ▒oor - it's ���l �n▒id�.”; 60% hits 30 and only fragments remain. With “Keep the first and last letter of every word” ticked, only 26 characters can be hit, and 30% gives “R�od 04�7: Don't op�n te door - it's s��ll in▒ide.”, which most readers can still work out.

For marks stacked above and below letters, the Zalgo look, use the glitch text generator or the cursed text generator. Those keep every letter and pile combining marks on top; this page removes and replaces the characters themselves.