Types of CAPTCHA, and what each one costs real people

A CAPTCHA can take many forms: letters to read, pictures to pick, a sound to transcribe, a slider to drag, or nothing visible at all. Each type was invented because automation learned to pass the one before it, and each one shifts the cost to a different group of real people. Choosing a type is choosing whom you are prepared to inconvenience.

Section Bots Updated Covers six types

Quick facts

Visible types
Distorted text, image selection, audio, puzzles and sliders, simple logic questions
Invisible types
Behavioural scoring and honeypot fields
Accessibility
Every visual or audio test excludes some people; see W3C, Inaccessibility of CAPTCHA
Accessible requirement
Offer an alternative in a different mode, not only a different puzzle
What none of them prove
That the visitor is a person; they only make automation more expensive

What a CAPTCHA is and why bots still pass is on CAPTCHA.

The main types of CAPTCHA

TypeWhat the visitor doesWho it locks out
Text CAPTCHAReads distorted letters and types themPeople with low vision, dyslexia; many people on small screens
Image CAPTCHAPicks every square showing an objectBlind and low-vision users; slow connections
Audio CAPTCHAListens to spoken characters and types themDeaf and hard-of-hearing users; noisy places; non-native speakers
Puzzle or sliderDrags a piece into placePeople with motor impairments; keyboard-only users
Logic questionAnswers a simple question or sumPeople with cognitive disabilities; non-native speakers
Invisible scoringNothing, unless the score is lowPeople with privacy tools, unusual setups, shared networks

The last row is the hardest to notice: when invisible scoring fails, the visitor gets a harder puzzle or a refusal with no explanation.

Image based CAPTCHA and sample CAPTCHA images

An image based CAPTCHA replaces typed letters with a picture task: the visitor has to recognise what a picture shows. Researchers proposed image recognition CAPTCHAs that ask people to identify simple objects in the images presented. The argument was that object recognition is more complex for software than text recognition, so these tests should resist machine learning attacks better.

Sample CAPTCHA images are easiest to tell apart by the task they set:

Sample CAPTCHA images by task
SampleWhat the picture showsWhat the visitor does
Object tilesSeveral photos, some with crosswalks or street lightsPicks every tile that matches the prompt
Still-image comparisonOne image, for example a human-shaped figure or an avatar-shaped oneSays which kind it is

CAPTCHA image verification and who it excludes

CAPTCHA image verification asks one thing of every visitor: can you see the picture, identify it and select the right answer? The W3C finds that relying on visual comparison alone makes these approaches difficult, if not impossible, for people with vision-related disabilities, and very difficult for people with visual processing disorders. A site that uses image verification needs a second route that does not depend on sight.

Google image CAPTCHA

The Google image CAPTCHA is the picture challenge of Google's reCAPTCHA service. In 2012 reCAPTCHA began using photographs taken from the Google Street View project, asking the visitor to identify crosswalks, street lights and other objects. Google also introduced a challenge, designed to be more accessible to mobile users, where the visitor selects the images that match a prompt from a grid.

Audio CAPTCHA

Audio CAPTCHAs exist so that people who cannot see an image test can still pass. They play spoken letters or numbers over background noise. The W3C notes that they are often hard to understand even for people with good hearing, and that they replace one excluded group with another rather than removing the barrier.

Speech recognition has improved to the point where many audio tests are easier for programs than for people. Offering audio as the only alternative to images is no longer enough.

Audio based CAPTCHA: what the test sounds like

An audio based CAPTCHA reads out characters, words or phrases for the visitor to enter in a form. Some implementations let people opt for the audio version, reCAPTCHA among them. Because recognising speech is no longer hard for software, the spoken characters are distorted and played over a sonic environment of obfuscating sounds.

That design creates its own failures. The W3C describes confusion over whether a number should be entered as a numeral or as a word, for example “7” or “seven”. It also notes that a listener can hear sounds that seem to be words or numbers and turn out to be background noise. A test meant to help people who cannot see the image can therefore become a second barrier.

Audio CAPTCHA problems: replay, language and controls

The W3C lists practical failures that a listener meets before any puzzle is solved. Few audio CAPTCHAs provide an easily invoked and reliable replay feature, let alone independent volume control or pause, rewind and fast-forward. When part of the test is not understood, the visitor may be played an entirely new one. Non-English audio CAPTCHAs appear to be very rare, and the W3C says visitors should not be presumed capable of understanding and transcribing aural English.

Some people are shut out entirely: users who are deaf-blind, who have no sound card or do not use one, or who are in a noisy environment. For screen reader users, the W3C says accuracy improves when the controls for the audio do not require moving focus away from the text response field, so that the screen reader and the CAPTCHA audio are not mixed.

Software does not share these problems. The W3C cites a University of Maryland study that reached a 90% success rate cracking Google's audio reCAPTCHA with Google's own speech recognition service.

Photo and animated CAPTCHA

Photo CAPTCHAs ask the visitor to pick images that match a description, such as every square with a crossing, or to rotate a photo until it is upright. Animated ones show moving letters or objects and ask the visitor to identify them, on the theory that motion is harder for software to follow.

Both have the same weakness as distorted text: image and video recognition has improved faster than puzzles have become harder, while the puzzles stay difficult for people with low vision or on small screens. They also depend on images that load and play reliably, which slow connections and data-saving settings do not guarantee.

Invisible CAPTCHA and behavioural scoring

An invisible CAPTCHA asks nothing at first. It watches how the page is used, how the browser behaves and where the request comes from, and gives the visit a score. Only low scores see a challenge.

Most people never notice it, which is its advantage. Its cost falls on people whose setups look unusual and who cannot see or contest the score; why they keep failing is on “Are you a robot?”.

COST AND BENEFIT BY TYPE VISIBLE TO EVERYONE EXCLUDES SOME USERS STOPS SIMPLE BOTS TEXT IMAGE AUDIO INVISIBLE SCORING HONEYPOT FIELD // a honeypot asks nothing of anyone, and stops only the simplest bots
// five types, and who pays for each

Honeypot CAPTCHA

A honeypot is a form field hidden from people but present in the page's code. People leave it empty; simple bots that fill every field do not, and their submissions are discarded.

It costs real users nothing when built correctly, which means hidden from screen readers as well as from sight, and not filled in by the browser's autofill. It stops only bots that fill fields blindly, so it is a first filter rather than a complete answer.

Why text CAPTCHAs faded

Distorted text was the first common CAPTCHA, because reading warped letters was once hard for software. Each time the text became harder to read for machines, it also became harder for people, and character recognition kept catching up.

Most sites have moved to image selection, invisible scoring or checks that ask nothing, covered on CAPTCHA alternatives. Text puzzles survive mainly on older systems and simple self-hosted forms, where they still stop the least capable bots.

Choosing a CAPTCHA type

  1. Start with what asks nothing. Honeypot fields, rate limits on the form, and checks on what was submitted.
  2. Add a challenge only where abuse is real, and only for visits that already look suspicious.
  3. Offer at least two modes, for example visual and a non-puzzle route such as an email link.
  4. Measure abandonment. Compare how many people leave at the check with how much abuse it stops.

Responses to automation that ask nothing of people are on blocking bots.

Questions

What are the types of CAPTCHA?

Distorted text, image selection, audio, puzzles and sliders, logic questions, invisible behavioural scoring, and honeypot fields.

What is an audio CAPTCHA?

An audio CAPTCHA plays spoken letters or numbers over background noise for the visitor to type, as an alternative for people who cannot use image tests.

Why are audio CAPTCHAs hard to use?

Few offer reliable replay or independent volume, pause and rewind controls, non-English versions are very rare, and users who are deaf-blind, have no sound card or are in a noisy place are prevented from proceeding, according to the W3C.

What is an invisible CAPTCHA?

An invisible CAPTCHA scores a visit from behaviour and browser signals without asking anything, and shows a challenge only when the score is low.

What is a honeypot CAPTCHA?

A honeypot is a form field hidden from people; bots that fill every field reveal themselves by filling it, and their submissions are discarded.

What is an image based CAPTCHA?

An image based CAPTCHA asks the visitor to recognise what a picture shows, for example by picking the tiles that contain a crosswalk, instead of typing distorted letters. It excludes people who cannot see the picture.

What does CAPTCHA image verification check?

It checks whether the visitor can identify the objects in the images shown. The W3C notes that relying on visual comparison alone is difficult, if not impossible, for people with vision-related disabilities, so a second route is needed.

What is the Google image CAPTCHA?

It is the picture challenge in Google's reCAPTCHA. Since 2012 it has used Street View photographs and asked visitors to identify objects such as crosswalks and street lights; a grid version asks them to select the images that match a prompt.