README

A small, dependency-free PHP library that detects and removes invisible watermarks — and other tell-tale characters — that AI providers leave in generated text.

Large language models routinely emit characters that render as nothing but survive copy/paste: zero-width spaces and joiners, Unicode tag characters (U+E0000–E007F) and variation selectors used for steganographic watermarking, bidirectional controls, and unusual spaces such as the narrow no-break space (U+202F). This library strips them, and can optionally normalize the visible typography (smart quotes, dashes, ellipses) that also signals machine authorship.

Zero dependencies. Pure PHP 8.1+, PCRE with the /u modifier, and mbstring.
Framework-agnostic. Drop it into any PSR-4 project and adjust the namespace.
Detect and clean. Use it as a filter, or as a "was this AI-watermarked?" check.

Install

composer require iserter/php-ai-text-sanitizer

Usage

One-liner

use iSerter\AiTextSanitizer\AITextSanitizer;

$clean = AITextSanitizer::clean_text($aiOutput);

Configured instance

$sanitizer = new AITextSanitizer([
    'normalize_smart_quotes' => true,
    'normalize_dashes'       => true,
]);

$clean = $sanitizer->sanitize($aiOutput);

Clean with a report of what changed

$result = $sanitizer->clean($aiOutput);

if ($result->wasModified()) {
    error_log($result->getReport()->summary());
}

echo $result->getText();

Detect only (don't modify)

$report = $sanitizer->detect($aiOutput);

if ($report->hasWatermarks()) {
    foreach ($report->getFindings() as $f) {
        printf("%s %s [%s] x%d\n", $f->notation(), $f->name, $f->category, $f->count);
    }
}

Options

Option	Default	Effect
`remove_zero_width`	`true`	Remove ZWSP, ZWNJ, ZWJ, word joiner, BOM (`U+200B–200D`, `U+2060`, `U+FEFF`).
`remove_bidi`	`true`	Remove bidirectional marks/controls (`U+200E/200F`, `U+202A–202E`, `U+2066–2069`, `U+061C`).
`remove_invisible_math`	`true`	Remove invisible math operators (`U+2061–2064`).
`remove_variation_selectors`	`true`	Remove variation selectors (`U+FE00–FE0F`, `U+E0100–E01EF`).
`remove_tag_chars`	`true`	Remove the Unicode Tags block (`U+E0000–E007F`).
`remove_format_chars`	`true`	Remove soft hyphen, CGJ, Hangul fillers, interlinear annotations, etc.
`remove_braille_blank`	`true`	Remove `U+2800` BRAILLE PATTERN BLANK.
`strip_control`	`true`	Remove stray C0/C1 control characters (keeps TAB, LF, CR).
`keep_emoji`	`true`	Preserve ZWJ and variation selectors when they are part of a valid emoji sequence.
`remove_citations`	`false`	Remove AI-generated citations like `(oaicite:5){index=5}` or `【13†source】`.
`normalize_homoglyphs`	`false`	Apply NFKC normalization to neutralize visually identical homoglyphs (requires `ext-intl`).
`normalize_spaces`	`true`	Fold unusual spaces (NBSP, thin space, ideographic space, …) to `U+0020`.
`normalize_line_separators`	`true`	`U+2028` → `\n`, `U+2029` → `\n\n`.
`normalize_smart_quotes`	`false`	Curly/angle quotes → straight `'` and `"`.
`normalize_dashes`	`false`	Dashes and minus (`U+2010–2015`, `U+2212`) → `-`.
`normalize_ellipsis`	`false`	`U+2026` → `...`.
`collapse_whitespace`	`false`	Collapse runs of horizontal whitespace, trim line-end spaces, cap blank lines.
`trim`	`false`	Trim leading/trailing whitespace of the whole string.

Docker

docker build -t php-ai-text-sanitizer .
docker run --rm php-ai-text-sanitizer

Demo

php src/examples/demo.php

iserter/php-ai-text-sanitizer

包简介

README 文档