typingnorm

Tell whether keystrokes came from a human. Three rules, each with a false-positive rate measured on 2,245 real typists. Zero dependencies, works in the browser.

npm install typingnorm

import { check } from "typingnorm";

// events: [milliseconds, "D" | "U", physical key code, optional event.key]
const report = check(events);

report.verdict              // "human-consistent" | "synthetic-signals" | "insufficient-data"
report.flags                // ["zero_rollover"]
report.measures             // { wpm, rollover_pct, dwell_cv, ... }
report.falsePositiveBudget  // 0.0031

Collect events straight from the DOM. Use event.code, not event.key: the rules depend on the physical key, and event.key changes with the input method.

const events = [];
input.addEventListener("keydown", (e) => {
  if (!e.repeat) events.push([e.timeStamp, "D", e.code, e.key]);
});
input.addEventListener("keyup", (e) => events.push([e.timeStamp, "U", e.code, e.key]));

The rules

| Rule | What it catches | False positives on 2,245 humans | |---|---|---| | same_key_overlap | A physical key pressed again before release | 0.00% | | constant_dwell | Every key held for the same time, or too short to be real | 0.00% | | zero_rollover | No key overlap at all (only applied above 35 WPM) | 0.31% |

Every rule is speed-invariant on purpose. Fast and slow typists differ enormously in rollover, interval, and hold time, so any "how far from the human average" rule systematically misjudges people at the extremes. Those are exactly the people you should not misjudge: password managers, dictation, switch access, and on-screen keyboards all look robotic in the time domain.

Hand alternation looked like a strong feature and was cut: the effect reverses for slow typists, so using it as a hard rule misjudges 15.46% of real humans.

What this does and does not do

It catches careless automation. It does not catch a program that deliberately imitates human timing, including this project's own simulator.

check() returns signals, not a verdict on a person, and deliberately offers no isHuman boolean, because that invites callers to use it as a gate. Treat it as one input to a risk score, never as identity verification.

Below 150 keystrokes it returns insufficient-data rather than guessing. A login form gives you twenty or thirty keystrokes, which is not enough for any conclusion.

Parity with the Python package

This is a port of typingnorm on PyPI. Both implementations run the same test vectors from vectors.json in CI, so they cannot drift apart.

Method, data sources, and the measured false-positive rates: METHOD.md.

MIT.