Skip to Content
Platform Ontology

Machine made input patterns

Clicks and keys sent by a tool tend to be too neat, too fast or oddly placed.

Gets silentlyWhile you use it

Input sent through the DevTools Protocol is trusted, so defenses look at how it behaves. Common tells are a click with no earlier mouse move, exact element center targets, whole number coordinates, identical gaps between keys, and zero hold time between key down and key up. Some tools also gave screen coordinates that matched page coordinates, which a real desktop rarely does.

What it can and can't do

What the user does: Nothing. Page script listens in the background

It can

  • Record pointer paths and key timing
  • Measure how neat the pattern is
  • Flag teleporting clicks and exact centers
  • Score the session as human or machine

It cannot

  • Catch careful tools that add human noise
  • Identify who runs the tool
  • Read what you type beyond timing
  • Work on pages with few interactions

How it differs between platforms: Agent tools now add random delays and curved paths, so the signal degrades over time. A person on a trackpad can also look unusually smooth.

Browser support

Version notes
  • Chrome: Applies to any Chromium based browser
  • Edge: Applies to Chromium Edge
  • Firefox: Same patterns can be measured
  • Safari: Same patterns can be measured
  • Safari on iPhone: Touch patterns
  • Chrome on Android: Touch patterns
  • Samsung Internet: Touch patterns

How a site asks for it

APImousemove/pointermove/click timing; MouseEvent.clientX/screenX; KeyboardEvent timing; PointerEvent.getCoalescedEvents()
Browser compat keysapi.MouseEvent.screenX
api.PointerEvent.getCoalescedEvents
How access is grantednone
What the user seesNone
Needs a secure context (HTTPS)No
Needs a user gestureNo
How long it lastsNot listed
How to take it backNot listed
Inside an iframeNot listed

Typical flow

  1. Page records pointer and key events with timestamps
  2. Script measures paths, gaps and coordinate patterns
  3. Model scores how human the pattern looks

Data you get

Shape: object
How often: stream while the person interacts

FieldTypeUnitRangePrecision
pointerPathlistpxpositions with timestamps before each clickNot listed
clickOffsetfloatpxdistance from the exact center of the targetNot listed
keyHoldTimeintegermsgap between key down and key upNot listed
screenVsClientintegerpxdifference between screenX and clientX, which should follow the window positionNot listed

Technical limits

  • Defeated by tools that add human noise
  • Smooth trackpad users can look robotic
  • Works best with many events

The story

What it is

A program that clicks tends to move like a program. The page measures that and guesses.

For you

You never see it. A person sounds irregular and so is accepted. An agent that clicks instantly in the exact middle of a button may be challenged.

For makers

Behavior models need many events, so use them on long forms, not on a single click. Add noise if you build agents that must act on defended sites, and respect the site's rules about automation.

Where you see it

TypeExampleApps
securityA login page scores mouse movement before accepting a password attempt.None named
shoppingA ticket queue slows clients that click with no earlier movement.None named
gamesA browser game flags auto clickers with identical delays.None named
adsAn ad unit discards clicks that arrive with no pointer path.None named
bankingA bank adds typing rhythm to a fraud score at sign in.None named
developerA tester adds random delays so an agent does not trip a defense.Playwright

Risks

  • Penalizes assistive tools and unusual input devices
  • Needs many events

What the platform says

  • MDN: Pointer events. Docs describe pointer events and how pages receive coordinates and timing.

Good uses and abuse

Legitimate uses

  • Bot defense
  • Fraud scoring

Known abuse

  • Behavior profiling
  • Blocking assistive tools

Scores

ScoreValueMeaning
Sensitivity3 of 5How revealing the data is
Utility4 of 5How useful it is to honest apps
Friction0 of 5How much the user has to do
Fingerprint bits2Rough identifying entropy, when known

History

  • 2023, Chrome: Public reports show CDP dispatched mouse events with matching screen and client coordinates

Sources

How well checked: Inferred from related behavior.

On this page