human-slop-detector

we talk a lot about ai slop these days. but you know what nobody talks about? humans. sloppy, keyboard-mashing, form-spamming humans.

i still get humans slopping all over my signup forms. hammering my apis with asdfghjkl. registering with [email protected].

this library aims to detect the most egregious human slop so you can protect your expensive ai powered services from the chaos we humans like to bring to them.

install

npm install human-slop-detector

quick try

npx human-slop-detector [email protected]
# => { "isSlop": true, "isLikelyHumanSlop": true, "confidence": 1 }

usage

import { detect, detectSentence } from "human-slop-detector";

// detect single strings
detect("asdfghjkl");
// => { isSlop: true, isKeysmash: true, isLikelyHumanSlop: true, confidence: 0.98 }

detect("hello world");
// => { isSlop: false, isKeysmash: false, isLikelyHumanSlop: false, confidence: 0.12 }

// detect whole sentences of slop
detectSentence("asdf jkl qwerty hjkl");
// => { isSlop: true, slopPercentage: 0.8, isLikelyHumanSlop: true }

// even catches sneaky email slop
detect("[email protected]");
// => { isSlop: true, isKeysmash: true, isLikelyHumanSlop: true }

what it detects

| type | example | human? | description | | ------------- | ---------------- | -------------- | ---------------------------- | | keysmash | asdfghjkl | ✓ yes | home row keyboard mashing | | keysmash | qwertyuiop | ✓ yes | enthusiastic jazz pianist | | keysmash | asjkdfhaskjdfh | ✓ yes | very anrgy human | | gibberish | xkjqzpmw | ✗ probably bot | random unpronounceable chars | | repeated | wsadwsadwsad | ✓ yes | i'm playing counter-strike |

real-world test results

we tested against like 100 real-world examples from twitter, reddit, github and signup forms then tried a simple openai prompt vs this libary:

| metric | gpt-5.1 | human-slop-detector | | --------- | ------- | ------------------- | | precision | 88.89% | 96.97% | | recall | 24.24% | 96.97% | | accuracy | 74.00% | 98.00% |

ground truth: what a human (me) thinks is human slop. run npx tsx scripts/benchmark.ts to regenerate.

want to help? we need more keysmashes from the wild to improve coverage, open keysmashes.txt and smash your keyboard - one per line, no questions asked.

api

detect(input, options?)

const result = detect("asdfghjkl");

result.isSlop; // true - it's slop (keysmash OR gibberish)
result.isKeysmash; // true - keyboard-based pattern
result.isGibberish; // false - not random gibberish
result.isLikelyHumanSlop; // true - probably a frustrated human
result.rageScore; // 0.29 - frustrated but not furious
result.confidence; // 0.98 - very confident
result.gibberishConfidence; // 0.76 - also looks gibberish-y

rage score

how angry is the human? we analyze keyboard patterns to estimate frustration levels:

| input | rage | interpretation | |-------|------|----------------| | qwerty | 0.11 | mild annoyance | | asdfghjkl | 0.29 | frustrated | | ASDFGHJKL | 0.54 | angry | | asjkdfhaskjdfhaskjdfh | 0.44 | angry | | ASDFGHJKLASDFGHJKL | 0.67 | very angry |

rage factors:

• length - longer keysmashes = more sustained frustration
• CAPS LOCK - uppercase = SCREAMING
• repetition - rhythmic aggressive patterns
• chaos - controlled keyboard walks vs frantic mashing
• hand clustering - one-handed rage (other hand facepalming)

detectSentence(input, options?)

const result = detectSentence("asdf jkl qwerty hjkl");

result.isSlop; // true
result.slopPercentage; // 0.8 - 80% of words are slop
result.slopWordCount; // 4
result.totalWords; // 5
result.words; // per-word breakdown

options

detect("qwerty", {
  layout: "dvorak", // keyboard layout for analysis
  threshold: 0.6, // confidence threshold (default: 0.5)
});

supported keyboard layouts

because apparently not everyone uses qwerty (looking at you, dvorak enthusiasts):

• qwerty - the og, the classic, the one true layout (default)
• azerty - bonjour mon amis
• qwertz - german efficiency
• dvorak - for the enlightened few
• colemak - the new hotness

supported languages

gibberish detection uses bigrams from 20 languages so we don't accidentally flag legitimate words:

english, chinese (pinyin), spanish, arabic, portuguese, indonesian, french, japanese (romaji), russian, german, korean, italian, turkish, vietnamese, polish, dutch, thai, hindi, persian, swedish

how it works

keysmash detection looks for keyboard-based patterns:

• key proximity (adjacent keys = sus)
• home row concentration (asdfghjkl gang)
• keyboard walk patterns (qwerty, zxcvbn)
• hand clustering (one hand mashing)
• limited vowel diversity (only 'a' because it's on the home row)

gibberish detection looks for unpronounceable chaos:

• uncommon letter combinations across 20 languages
• excessive consonant clusters (xkjqzp is not a word)
• abnormal vowel patterns

human detection distinguishes keyboard mashing from random generation:

• humans hit adjacent keys
• humans favor the home row (lazy fingers)
• bots generate uniformly random garbage

contributing

i've added a few test cases, keyboards and languagues, please add more.

why "human slop"?

because "keysmash-detector" was boring and i used to work somwhere where the all function names in our codebase were keysmashes, long before claude started coding.

license

mit - do whatever you want with it