facts 524 · signals 29 · articles 20 · database 3,168 KB · bots —guest@ftrlgthis is you~/whoamiYou post anonymously. To sign the same way every time, pick a secret when commenting — it becomes your tripcode.no sign-up, ever

Behavioral biometrics

translated by AI, reviewed by the editors

future.log editors · published

Behavioral biometrics is recognising a person by how they do things: the rhythm in which they press keys, how they move the mouse, how they scroll a phone screen. Unlike a fingerprint, this kind of “print” can be taken unnoticed and continuously, while the person is simply using a website.

The idea is more than forty years old, but it went mainstream in the 2010s: banks use it to check that the account owner is the one at work, and websites use it to check that they are dealing with a human rather than a bot. With the arrival of AI agents that move the cursor and type on their own, behaviour has become one of the few signs that can still tell them apart from people.1

History

  • 1940s During the Second World War, British radio interceptors could not read encrypted Morse code, but they recognised German radio operators by the rhythm of their transmissions, known as the operator’s “fist”.2 ⧗ ¶ chjg8jet4
  • 26.08.1983 John Garcia files a patent application for a device that compares the pauses between keystrokes with a stored sample of the owner’s typing.3 ¶ cg4cedewp
  • 15.12.2000 Net Nanny releases BioPassword, a Windows NT login that checks not only the password but also how the person types it.4 ¶ cvgt52jm7
  • 07.2007 Ahmed Awad E. Ahmed and Issa Traore propose recognising a user by their mouse movements.5 ¶ c9nksar4j
  • 29.03.2017 Mastercard buys NuData Security, a “passive biometrics and behavioral analytics company”.6 ¶ cynic22mj
  • 29.10.2018 Google releases reCAPTCHA v3: instead of pictures, an invisible score from 0 to 1 based on how a person behaves on the website.7 ¶ ckqc4f264
  • 02.05.2024 The fund Permira buys a majority stake in BioCatch at a $1.3 billion valuation; more than 190 banks use the company’s behavioral biometrics.8 ¶ cny4wi8yy
  • 28.08.2025 Cloudflare introduces “signed agents”: an AI agent can prove who it is with a cryptographic signature rather than with its behaviour.9 ¶ ccc94wi7u

How it works

The system learns how a person usually types and moves the cursor: how long they hold a key, how long the pause between letters lasts, along what arc and at what speed the mouse travels. It then compares every new action with this sample. There are two uses.

Checking that it’s you. A bank quietly watches behaviour throughout the session, not only at login: if someone in the account suddenly types differently, it may be a fraudster. This is how BioCatch, BehavioSec (part of LexisNexis Risk Solutions since 2022) and NuData (Mastercard) work.8106 ¶ c5yxkmxf7

Checking that you’re human. A website judges whether the behaviour looks human and either lets the visitor through without questions or asks them to solve a CAPTCHA. reCAPTCHA v3 returns a number from 0 to 1; by default, Google advises treating everything below 0.5 as a bot.11 ¶ c27euapsb

Data

How accurately it recognises a person

Who and whenWhat was studiedWhat was found
Killourhy and Maxion (CMU), 20091251 people type the same password 400 times; 14 algorithms comparedThe best one errs in roughly one case in ten. ¶ c6puj3dgt
Zheng, Paloski, Wang, 10.201113Recognising a user by the mouseAfter 20 clicks, an error rate of 1.3%. But on a real forum, 20 clicks took up to 38 minutes. ¶ cnkfbbice
Frank et al., 201314Recognising a person by how they scroll the screenAlmost no errors within one session, up to 4% a week later. The authors: suitable as an addition, not instead of a password. ¶ cynp6pi8c

What a typing style gives away

Who and whenWhat was studiedWhat was found
Giancardo et al. (MIT), 16.04.201515Whether typing shows that a person is sleep-deprivedYes: detection accuracy (AUC) of 0.93. ¶ c8fs7rtng
Arroyo-Gallego et al. (MIT), 03.201816Whether early Parkinson’s disease shows in typing at homeYes, with an AUC of 0.76, from ordinary typing, without special tests. ¶ cuctwz92x
Buriro et al., 09.201617Age and gender from entering a PIN on a phoneAge in 88% of cases, gender in 83%. ¶ ci8geueu9

Does it tell bots and AI agents apart

Who and whenWhat was studiedWhat was found
Chu et al., 10.201218Blog bots: 1,078 real users and bots97.9% of bots caught, 0.2% false alarms. ¶ cn4jvrsrt
Liu et al., 23.10.202419Can a neural network move the mouse “like a human”Its trajectories cut the accuracy of bot detectors by 5–10%. ¶ ce6sh8cqu
Wang, Shafiq, Vekaria (UC Davis), 02.05.202620Seven AI agents that work in the browserAll seven recognised by behaviour; Cloudflare recognised one. ¶ cw4g3654p
Choudhary et al., 29.07.202621Why behaviour gives agents awayStandard detectors took 35–39% of agents for humans. Two features, the frequency of mouse events and clicks without cursor movement, were enough to catch them all: what gives an agent away is not its “mind” but the program that controls it. ¶ cffgcfrzh
Akamai, 19.08.20261How the AI browsers Comet and Atlas behave; data from a company that sells bot protectionIn 63% of requests the mouse did not move at all. ¶ cx7ir25px
[EDITORS’ CONCLUSION]

Behaviour does not recognise a person reliably enough to replace a password (error rates from 1 to 10%, CMU, Zheng), but it is enough to reveal too much about them: sleep deprivation, early Parkinson’s disease, age and gender (MIT, Buriro). Behaviour still catches AI agents, but not because a human cannot be imitated: it is because the programs that act as the agents’ “hands” move the mouse unnaturally (Choudhary, Akamai). Neural networks are already learning to move the mouse like humans (Liu), and this gap will narrow. Caveat: almost all studies of agents are 2026 preprints or research by vendors of protection. ¶ cyz55igre

What the law says

WhereDocumentWhat was decided
EUGDPR, Art. 4(14), 201622Behavioural traits are biometrics, but only if they allow a person to be uniquely identified. Fraud prevention is recognised as a company’s legitimate interest. ¶ chvcbrp4w
EUAI Act, Arts. 3 and 5, from 02.02.202523Biometrics covers any behavioural traits; “keystrokes characteristics” are named explicitly. Recognising emotions at work and in education is banned.24 ¶ cze5bxdtk
EUEuropean Commission guidelines, 29.07.202524Emotions inferred from typing style at work are banned; the same emotions for advertising are not. ¶ chqmyzrt7
CaliforniaCalifornia Consumer Privacy Act (CCPA)25“Keystroke patterns or rhythms” are explicitly classed as biometrics. ¶ cvynyeuxc
IllinoisBiometric Information Privacy Act (BIPA)26A closed list: iris, fingerprint, voice, face, hand. Behaviour is not on it. ¶ cidsg6wui
United KingdomICO regulator, 23.02.202427“How they type” is biometrics; special protection applies when it is used to recognise a specific person. ¶ c7wapvpcv

Incidents

TikTok’s in-app browser, August 2022 ¶ cpb7e8asm

Researcher Felix Krause showed that when a user opens a link inside the TikTok app for iOS, the app injects its own code into the third-party website, and this code subscribes to every keystroke and every tap. In his words, this is “the equivalent of installing a keylogger on third party websites”. Krause stressed that he did not know whether this data was sent anywhere. TikTok replied that the code is “used only for debugging, troubleshooting and performance monitoring”.28

Employee monitoring through the mouse, December 2024 ¶ cmuasm2ny

The French regulator CNIL fined a real estate company €40,000. Software on the computers of employees who worked remotely counted every few minutes without keystrokes or mouse movements as idle time; this idle time could be deducted from pay, and screenshots of their screens were taken regularly. The regulator found such monitoring excessive.29

An AI agent clicks “I am not a robot”, July 2025 ¶ cn3cn9k3z

A Reddit user posted screenshots: the ChatGPT agent ticks the Cloudflare verification box on its own and explains: “This step is necessary to prove I’m not a bot and proceed with the action.” Whether this got past a behavioural check is unclear: a month later Cloudflare classed the ChatGPT agent as a “signed agent”, which it recognises by a cryptographic signature.309

Criticism

As early as 2012, European regulators warned that behavioral biometrics was spreading beyond identity checks into “behaviour analysis, surveillance and fraud prevention”.31 ¶ cezdf9y3i

In 2018, Privacy International wrote that websites collect tens of millions of behavioural profiles, while users cannot tell when this is happening; human rights advocates are worried that such data reveals illnesses.32 ¶ cjgqe7gdw

In 2021, the W3C pointed out the downside of invisible checks such as reCAPTCHA v3: they are convenient for people with disabilities because they do not require solving puzzles, but they hand Google even more data about the user.33 ¶ cnw9g853n

In 2022, researcher Jennifer Mankoff and her co-authors wrote that inaccessible biometric checks are “essentially defining a disabled person as non-human”.34 ¶ c2xzx98fs

Outlook

The industry is moving from guessing by behaviour to signatures. Since 2025, an honest bot or agent can sign its requests cryptographically under the Web Bot Auth standard, promoted by Cloudflare and Google, and a website no longer has to guess from the mouse.3536 In October 2025, Visa justified its protocol for AI agents by saying that bot detection systems mistakenly block legitimate purchases made by agents.37 According to Cloudflare, in 2026 more than half of internet traffic turned out to be non-human for the first time.38 ¶ cbdv4wexk

[EDITORS’ CONCLUSION]

Behavioral biometrics is changing its role. Honest agents present a signature, and for them behaviour is no longer needed; it remains a check for those who do not want to sign, and for people who are still being watched unnoticed. That makes the key question one the law answers in different ways: in the EU, emotions from typing style cannot be recognised at work but can be for advertising; in California, keystroke rhythm is biometrics, in Illinois it is not. The better AI gets at imitating human movements, the less use such monitoring has for security, and the more questions it raises as surveillance. ¶ c9m5isezn

See also

Proof of personhoodEmployee monitoringDeath of CAPTCHAAutonomous AI agentsDead internet theoryPrivacy Pass and PACT

References

  1. M. Sparkman, D. McAndrew. Identifying Agentic Automation with Behavioral Telemetry — Akamai, 19.08.2026. akamai.com ↩ ↩2

  2. J. Rehmeyer. Digital Fingerprints — Science News, 09.01.2007. sciencenews.org ⧗ ↩

  3. J. D. Garcia. Personal identification apparatus, US4621334 — application 26.08.1983, patent 04.11.1986. patents.google.com ↩

  4. S. Bekker. Net Nanny Ships BioPassword for Windows NT — RCP Magazine, 19.12.2000. rcpmag.com ↩

  5. A. A. E. Ahmed, I. Traore. A New Biometric Technology Based on Mouse Dynamics — IEEE Transactions on Dependable and Secure Computing, 07.2007. doi.org ↩

  6. Mastercard Enhances Security of the Internet of Things with the Acquisition of NuData Security Inc. — Mastercard, 29.03.2017. investor.mastercard.com ↩ ↩2

  7. Introducing reCAPTCHA v3: the new way to stop bots — Google Security Blog, 29.10.2018. security.googleblog.com ↩

  8. Permira to Acquire Majority Position in BioCatch at $1.3bn Valuation — Permira, 02.05.2024. permira.com ↩ ↩2

  9. The age of agents: cryptographically recognizing agent traffic — Cloudflare, 28.08.2025. blog.cloudflare.com ↩ ↩2

  10. LexisNexis Risk Solutions Acquires Behavioural Biometric Innovator BehavioSec — PR Newswire, 03.05.2022. prnewswire.co.uk ↩

  11. reCAPTCHA v3 — Google for Developers. developers.google.com ↩

  12. K. S. Killourhy, R. A. Maxion. Comparing Anomaly-Detection Algorithms for Keystroke Dynamics — DSN-2009, IEEE, 2009. cs.cmu.edu ↩

  13. N. Zheng, A. Paloski, H. Wang. An Efficient User Verification System via Mouse Movements — ACM CCS, 10.2011. pdf ↩

  14. M. Frank, R. Biedert, E. Ma, I. Martinovic, D. Song. Touchalytics: On the Applicability of Touchscreen Input as a Behavioral Biometric for Continuous Authentication — arXiv, 26.07.2012; IEEE TIFS, 01.2013. arxiv.org ↩

  15. L. Giancardo et al. Psychomotor Impairment Detection via Finger Interactions with a Computer Keyboard During Natural Typing — Scientific Reports, 16.04.2015. pmc.ncbi.nlm.nih.gov ↩

  16. T. Arroyo-Gallego et al. Detecting Motor Impairment in Early Parkinson’s Disease via Natural Typing Interaction With Keyboards — Journal of Medical Internet Research, 03.2018. dspace.mit.edu ↩

  17. A. Buriro, Z. Akhtar, B. Crispo, F. Del Frari. Age, Gender and Operating-Hand Estimation on Smart Mobile Devices — BIOSIG, 09.2016. repository.essex.ac.uk ↩

  18. Z. Chu, S. Gianvecchio, A. Koehl, H. Wang, S. Jajodia. Blog or block: Detecting blog bots through behavioral biometrics — Computer Networks, 17.10.2012. pdf ↩

  19. J. Liu et al. DMTG: A Human-Like Mouse Trajectory Generation Bot Based on Entropy-Controlled Diffusion Networks — arXiv, 23.10.2024. arxiv.org ↩

  20. E. Wang, Z. Shafiq, Y. Vekaria. FP-Agent: Fingerprinting AI Browsing Agents — arXiv, 02.05.2026. arxiv.org ↩

  21. Choudhary et al. What Does It Take to Detect an AI Agent? Minimal Feature Sets for Behavioral Detection under Browser Automation — arXiv, 29.07.2026. arxiv.org ↩

  22. Regulation (EU) 2016/679 (GDPR), Article 4(14), Recital 47 — Official Journal of the EU, 27.04.2016. publications.europa.eu ↩

  23. Regulation (EU) 2024/1689 (AI Act), Article 3(34), Article 5(1)(f), Recital 15 — Official Journal of the EU, 12.07.2024. publications.europa.eu ↩

  24. Guidelines on prohibited artificial intelligence practices, C(2025) 5052 final — European Commission, 29.07.2025. ec.europa.eu ↩ ↩2

  25. California Consumer Privacy Act, Cal. Civ. Code §1798.140 — California Privacy Protection Agency, version as of 01.01.2025. cppa.ca.gov ↩

  26. Biometric Information Privacy Act, 740 ILCS 14/10 — Illinois General Assembly. akingump.com ↩

  27. Biometric data guidance: Biometric recognition — ICO, 23.02.2024. ico.org.uk ↩

  28. F. Krause. iOS Privacy: Announcing InAppBrowser.com - see what JavaScript commands get injected through an in-app browser — krausefx.com, 18.08.2022. krausefx.com ↩

  29. Surveillance excessive des salariés : sanction de 40 000 euros à l’encontre d’une entreprise du secteur immobilier — CNIL, 04.02.2025. cnil.fr ↩

  30. B. Edwards. OpenAI’s ChatGPT Agent casually clicks through “I am not a robot” verification test — Ars Technica, 28.07.2025. arstechnica.com ↩

  31. Opinion 3/2012 on developments in biometric technologies (WP193) — Article 29 Data Protection Working Party, 27.04.2012. ec.europa.eu ↩

  32. Behavioural biometrics flag fraud but invade privacy — Privacy International, 13.08.2018. privacyinternational.org ↩

  33. Inaccessibility of CAPTCHA — W3C Group Note, 16.12.2021. w3.org ↩

  34. J. Mankoff et al. Areas of Strategic Visibility: Disability Bias in Biometrics — arXiv, 14.07.2022. arxiv.org ↩

  35. T. Meunier. Forget IPs: using cryptography to verify bot and agent traffic — Cloudflare, 15.05.2025. blog.cloudflare.com ↩

  36. Web Bot Auth — Google for Developers. developers.google.com ↩

  37. Visa Introduces Trusted Agent Protocol: An Ecosystem-Led Framework for AI Commerce — Visa, 14.10.2025. usa.visa.com ↩

  38. M. Conroy. The Internet has a second audience — Cloudflare, 30.09.2026. blog.cloudflare.com ↩

Your signature

Remember the secret — it is your password. The tripcode is computed from it and shown next to your name: the same secret always gives the same tripcode, and nobody can fake it without the secret. The secret is never published or stored on the server.