The checkbox looks like the whole test: move the pointer, press once, and wait for a green tick. When Google introduced No CAPTCHA reCAPTCHA in December 2014, however, the visible click sat on top of an advanced risk-analysis system that considered a user’s engagement before, during and after the widget.

Google said a significant number of users would be able to verify themselves with one click. It also made clear that users would still receive a CAPTCHA challenge whenever the risk-analysis engine could not confidently classify a request.

That is more complicated than the familiar explanation that humans wobble while bots move in perfectly straight lines. Mouse activity can provide information, but the available evidence does not show that cursor wobble was the decisive secret behind the 2014 checkbox.

The checkbox is only the visible part

In 2016, Columbia University researchers Suphannee Sivakorn, Iasonas Polakis and Angelos Keromytis published a detailed black-box study of reCAPTCHA. They examined what the widget collected, which requests received an immediate green tick, and which were given image or text challenges.

The researchers reported that the widget collected information about the browser as soon as it loaded. When the checkbox was clicked, the request sent to Google included the referring page, the site’s reCAPTCHA key, a Google cookie and encrypted information produced by the widget’s browser checks.

If the browser was logged into a Google account, an authentication cookie added another piece of information. The widget also checked browser attributes, looked for signs of automation frameworks and performed tests that included canvas fingerprinting.

The most striking result involved cookie age. In the researchers’ tests, a nine-day-old Google tracking cookie was enough to receive the checkbox challenge even when it carried almost no browsing history, while newly created Google accounts were initially treated as suspicious.

Those findings describe a particular version of reCAPTCHA at a particular time. Google has never published the complete model, and the system has continued to change, so lists claiming to reveal every signal should be treated cautiously.

Mouse movement is a clue, not a magic signature

Human pointing movements are rarely geometrically perfect. A hand may accelerate unevenly, pause near a target, overshoot it slightly or make a correction that the person controlling the mouse never consciously notices.

Paul Fitts studied the relationship between movement and target selection in his 1954 paper on the information capacity of the human motor system. The principle now called Fitts’s law predicts how movement time changes with target distance and width, but it does not say that human pointers always follow curved paths.

The Columbia researchers specifically tested mouse timing, erratic movement, multiple clicks and a JavaScript-generated click that reached the checkbox without hovering over it. None of those changes harmed the session’s risk result in their experiments.

A straight pointer path is therefore not automatic proof of a bot, just as a centred click is not automatic proof of automation. Scripts can dispatch synthetic mouse events, while automation frameworks may expose themselves through browser attributes before cursor geometry becomes relevant.

Bots learned to imitate the interaction

A later experiment produced a different result with reCAPTCHA v3. In 2019, Ismail Akrout, Amal Feriani and Mohamed Akrout published “Hacking Google reCAPTCHA v3 using Reinforcement Learning”, in which an automated agent learned to move a pointer across a grid and obtain a high risk score.

The system succeeded in 97.4 percent of runs on a 100-by-100 grid and 96.7 percent on a 1,000-by-1,000 screen. Performance fell when the agent moved in larger jumps, which the authors interpreted as evidence that less natural movement could affect the score in their setup.

That experiment suggests mouse behaviour can matter, but it also demonstrates why “bots don’t wobble” is the wrong distinction. A bot can generate a path, vary its steps and learn which movements a detector rewards.

Audio challenges entered the same arms race. The University of Maryland’s unCaptcha project used public speech-to-text services to solve 85.15 percent of more than 450 audio challenges in an average of 5.42 seconds.

After Google changed those challenges from spoken digits to phrases, the team released unCaptcha v2 in December 2018. Its developers reported accuracy of about 90 percent, showing that replacing one puzzle with another did not create a permanent boundary between people and software.

Why image grids still appear

The grids asking users to select traffic lights, buses or crosswalks are fallback challenges. Google explained in 2014 that they appear when its risk engine cannot confidently decide whether a request comes from a person or an abusive agent.

The images present a computer-vision problem that was once harder for software than reading distorted text. They also work more naturally on a phone, where tapping photographs is easier than typing warped letters into a small field.

A popular story claims that the traffic-light grids were created to label training data for Google’s self-driving cars. Their timing overlapped with Google’s work on autonomous vehicles, but that coincidence does not establish that reCAPTCHA answers were formally used for that purpose, and the cited Google material does not make the connection.

Earlier reCAPTCHA systems did perform documented secondary work, including helping digitise scanned books and identify Street View numbers. By April 2014, however, Google reported that its own recognition system could read difficult Street View numbers with 90 percent accuracy and solve its hardest distorted-text reCAPTCHA images with more than 99 percent accuracy.

The test moved into the background

In October 2018, Google introduced reCAPTCHA v3, which removed the visible challenge and returned a risk score instead. Make Tech Easier covered that shift when the version was released.

Google’s current documentation describes scores from 0.0 to 1.0, with higher values indicating a lower-risk interaction. The website, not the visitor, receives the result and can use it to allow an action, request another verification step or send suspicious activity for review.

The score is not a definitive certificate that someone is human. Google advises site owners to interpret it in context and combine it with appropriate responses, such as email verification or multifactor authentication, rather than treating every low score as conclusive proof of automation.

This background assessment carries privacy concerns. Apple has noted that CAPTCHA providers may rely on tracking, fingerprinting and IP addresses, and its Private Access Tokens use blinded cryptographic signatures to attest that a device passed checks without letting the receiving site identify it over time.

The checkbox has also become familiar enough to imitate. Fake CAPTCHA pages now use that familiar visual language to persuade people to run malicious commands, even though those scams have nothing to do with Google’s underlying risk analysis.

When a real reCAPTCHA checkbox turns green, the cursor’s tiny turns may be one part of what happened. Behind that brief movement sit browser checks, cookies, account state and other signals hidden from view, all compressed into the quiet pause between the click and the tick.