CAPTCHAs were designed as a test that humans pass and computers fail. Twenty years later, that description is hard to defend. Published research shows automated systems matching or beating people on most CAPTCHA types, while people still spend seconds, sometimes most of a minute, on every challenge.
So do CAPTCHAs still stop bots? The research points to a more precise answer: they rarely stop a determined operator, but they add cost and friction, to attackers and legitimate users alike. This study brings together the measurements behind that answer: how long CAPTCHAs take humans, how often machines solve them, what AI agents manage today, and how the defence has shifted from puzzles to risk scoring.
Every figure below comes from the paper’s own text or tables. Several recent results are preprints, which we mark. This is a summary of published research, not a guide to solving CAPTCHAs.
Key takeaways
- Humans pay in seconds. In the largest recent lab study, average solving times ranged from 3.6 to 42.7 seconds per CAPTCHA, depending on the type. In a live deployment, image challenges took people 10.3 seconds on average against 1.85 seconds for a checkbox.
- Machines match or beat humans on accuracy. The same lab study compared its human results with bot results from earlier research and concluded that “bots can outperform humans, both in terms of solving time and accuracy, across all these CAPTCHA types.”
- General-purpose AI agents still fall short, purpose-built solvers do not. In one 2025 benchmark the best general agent solved 40% of interactive CAPTCHAs against 93.3% for humans. Systems built for specific CAPTCHA families report results close to or above human level.
- The real mechanism is cost. As early as 2010, researchers concluded that a CAPTCHA “reduces an attacker’s expected profit by the cost of solving the CAPTCHA” rather than preventing automated access.
- Defences have moved behind the scenes. Major providers now score requests without showing a puzzle, and show a challenge only when the score is doubtful.
What CAPTCHAs cost humans
The most detailed measurement of modern CAPTCHAs is a 2023 USENIX Security study by Searles and colleagues. In it, 1,400 participants solved 14,000 CAPTCHAs across ten types.
| CAPTCHA type | Human solve time (seconds) |
|---|---|
| Checkbox (click) | 3.1 to 4.9 |
| Distorted text | 9 to 15.3 |
| Image selection (reCAPTCHA) | 15 to 26 |
| Image selection (hCaptcha) | 18 to 32 |
| Game-based | 18 to 42 |
| Slider puzzle | 28 to 30 |
Source: Searles et al. (2023), Table 3. Ranges cover the study’s different settings.
Human accuracy was far from perfect: around 81% on image challenges, and between 50% and 85% across the types where accuracy could be measured. Framing the CAPTCHA as part of an account sign-up, rather than as a standalone task, made people slower, by 26.7% on average across types.
People also give up. In a follow-up study within the same paper, 30% of participants abandoned before finishing, and those who saw the CAPTCHAs as part of an account sign-up were 120% more likely to abandon. These were paid study participants, so the figures describe task abandonment rather than lost sales, but the direction is clear.
A 2025 study by Searles, Prapty and Tsudik measured the same thing in production. Over 13 months, more than 3,600 real users of a university account service passed through reCAPTCHA v2 without knowing they were being studied:
| Measure | Result |
|---|---|
| Sessions resolved by the checkbox alone | 79.98% |
| Mean solve time, checkbox | 1.85 seconds |
| Mean solve time, image challenge | 10.3 seconds |
| Image challenges passed | 92.96% |
| Survey respondents who found image challenges annoying | 40% (under 10% for the checkbox) |
The authors also estimated the total human time spent on reCAPTCHA across its lifetime at 819 million hours, worth at least 6.1 billion dollars at an hourly wage of 7.50 dollars. They describe the estimate as a rough lower bound built on assumptions, including a 2010 figure of 100 million CAPTCHAs a day held constant, so it is best read as an order of magnitude. For comparison, reCAPTCHA’s creators reported in 2008 that people solved more than 100 million CAPTCHAs a day, “hundreds of thousands of human hours per day.”
What machines achieve
Automated solving has tracked every generation of CAPTCHA. The 2023 study compared its human results with bot results reported in earlier research:
| CAPTCHA type | Human accuracy (2023 study) | Bot accuracy (earlier studies) | Bot solve time |
|---|---|---|---|
| Checkbox (click) | not directly measured | 100% | 1.4 s |
| Slider puzzle | not directly measured | 96% | 5.3 s |
| Distorted text | 50% to 85% | 99.8% | under 1 s |
| Image selection (reCAPTCHA) | 81% | 85% | 17.5 s |
| Image selection (hCaptcha) | 71% to 81% | 98% | 14.9 s |
Source: Searles et al. (2023), Table 3 and section 5.5, citing earlier studies for the bot results.
The bot figures come from different papers under different conditions, so this is not a controlled head-to-head. But the authors’ conclusion is direct: bots “can outperform humans, both in terms of solving time and accuracy.” Distorted text fell first; a Google research team reported 99.8% accuracy on the hardest category of reCAPTCHA text puzzles in 2014.
A 2024 study from ETH Zurich reported that an automated system passed every reCAPTCHA v2 session it attempted, sometimes after several image rounds, and needed a similar number of challenges to a human tester (one person, in their comparison). Its most telling finding concerns what the system is actually measuring: the authors concluded that reCAPTCHA v2 “is heavily based on cookie and browser history data when evaluating whether a user is human or not.” In their tests, a browser without history needed far more challenges than one with an established history. The image task, in other words, matters less than what the browser already looks like.
What AI agents achieve today
Since 2024, researchers have tested general-purpose AI models and agents on CAPTCHAs, with mixed results.
| Study | Status | Result |
|---|---|---|
| Open CaptchaWorld (2025) | Preprint | Best general web agent solved 40.0% of 225 interactive CAPTCHAs; humans 93.3% |
| MCA-Bench (2026) | AAAI 2026 | A fine-tuned vision model scored 0.025 on arithmetic selection against 0.78 for humans, but 0.985 on distorted words against 0.84 for humans |
| Oedipus (2024) | Preprint | General models solved 12% to 21% of reasoning CAPTCHAs unaided, and 51% to 64% with a task-decomposition framework |
| VIPERStrike (2026) | Accepted, USENIX Security 2026 | Up to 93.2% on visual-reasoning CAPTCHAs from six providers, against 87.6% to 95.4% for a small group of human volunteers |
The pattern is consistent. Off-the-shelf agents struggle with multi-step, interactive challenges. Systems built for a specific family of CAPTCHAs close most of the gap with humans, and on simple visual tasks exceed them. One encouraging result for defenders: in the VIPERStrike paper, randomising the wording of challenge prompts reduced every solver’s success rate in a preliminary test, by between 4.5 and 20.5 percentage points.
Deterrence is a cost, not a wall
If machines can solve CAPTCHAs, why do they still exist? A 2010 USENIX Security study by Motoyama and colleagues answered that question before the AI era. Measuring the market for human CAPTCHA-solving labour, they found solving available at retail prices “as low as $1 per thousand”, with a median response time of 14 seconds and 86% to 89% accuracy at most services they tested. Their conclusion still frames the debate:
“In an absolute sense, CAPTCHAs do not prevent large-scale automated site access.” A CAPTCHA “reduces an attacker’s expected profit by the cost of solving the CAPTCHA.”
That makes a CAPTCHA an economic control. It works where the value of each automated action is small compared with the cost of passing the challenge, such as mass account creation or spam, and fails where each action is valuable. It also costs every legitimate user the same seconds and the same frustration.
Recent preprint research puts AI-based solving in the same cost range: the Oedipus authors estimated a few dollars to around twenty dollars per hundred solved reasoning CAPTCHAs, depending on the model and the type, which they described as within the same order of magnitude as human labour.
From puzzles to risk scores
The industry has largely accepted this, and moved the decision out of sight.
- Google reCAPTCHA v3 “returns a score for each request without user friction”, from 0.0 (very likely a bot) to 1.0 (very likely a good interaction), with a default threshold of 0.5. Google’s documentation suggests using the score to decide what to do, such as requiring two-factor authentication on a risky login or sending a risky comment to moderation, rather than to block outright.
- Cloudflare Turnstile “works without showing visitors a CAPTCHA”, running small non-interactive browser challenges in the background and escalating to a checkbox only when needed.
The puzzle has become the fallback, shown when the background score is unsure. That matches the ETH Zurich finding about browser history, and the wider shift in bot defence described in our article on how anti-bot defence evolved past CAPTCHAs.
What this means
For site owners, the research suggests treating CAPTCHAs as one cost-raising layer, not a gate:
- Show challenges only on actions that are worth protecting, such as sign-ups, logins and checkouts, and only when a risk score is doubtful.
- Measure what challenges cost you. Abandonment in the studies above ran well into double digits.
- Expect a determined operator to pass, and combine challenges with rate limits, account-level controls and server-side checks on the actions themselves.
For data teams, a CAPTCHA is a clear signal: the site has judged your traffic as automated and does not want it on that path at that rate. The right response is to slow down, check whether an official API or data licence exists, and stay within what the site allows, not to look for a way past the challenge. Our web scraping best practices and our guide to robots.txt and AI opt-outs cover how to read those signals.
FAQ
Do CAPTCHAs still work?
As a strict barrier, not reliably: research shows automated systems matching or beating human accuracy on most CAPTCHA types. As a cost that slows down low-value automated abuse, such as mass sign-ups, they still have a role, especially combined with background risk scoring.
How long do CAPTCHAs take humans to solve?
In a 2023 study of 1,400 people, average solving times ranged from 3.6 to 42.7 seconds depending on the type. In a live deployment of reCAPTCHA v2, a checkbox took 1.85 seconds on average and an image challenge 10.3 seconds.
Can AI solve CAPTCHAs?
General-purpose AI agents still trail humans on interactive CAPTCHAs; in one 2025 benchmark the best agent solved 40% against 93.3% for humans. Systems built for specific CAPTCHA types report results close to human level.
Why are CAPTCHAs being replaced?
Because they cost real users time and conversions while determined bots get through. Providers now score requests in the background and show a challenge only when the score is uncertain.
The bottom line
CAPTCHAs no longer separate humans from machines reliably; on most types, machines are now at least as accurate. What they still do is add cost: seconds and frustration for people, money and effort for automated traffic. That makes them useful where each automated action is worth little, and weak where it is worth a lot. The industry has responded by moving the real decision to invisible risk scoring and keeping the puzzle as a fallback.
Sources and references
- Searles, A. et al. (2023). An Empirical Study & Evaluation of Modern CAPTCHAs. USENIX Security 2023.
- Searles, A., Prapty, R. T. and Tsudik, G. (2025). Understanding reCAPTCHAv2 via a Large-Scale Live User Study. NDSS Symposium on Usable Security and Privacy (USEC) 2025.
- Plesner, A., Vontobel, T. and Wattenhofer, R. (2024). Breaking reCAPTCHAv2. COMPSAC 2024.
- Motoyama, M. et al. (2010). Re: CAPTCHAs: Understanding CAPTCHA-Solving Services in an Economic Context. USENIX Security 2010.
- von Ahn, L. et al. (2008). reCAPTCHA: Human-Based Character Recognition via Web Security Measures. Science 321.
- Goodfellow, I. et al. (2014). Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks.
- Luo et al. (2025). Open CaptchaWorld. Preprint.
- Wu et al. (2026). MCA-Bench. AAAI 2026.
- Deng et al. (2024). Oedipus. Preprint.
- Qi et al. (2026). VIPERStrike. Accepted at USENIX Security 2026.
- Google, reCAPTCHA v3 documentation; Cloudflare, Turnstile documentation.