For years, customers and partners have asked us whether typing biometrics could help distinguish natural human typing from synthetic or automated input.
That question is becoming more important as bots, browser automation, scripted form filling, and AI agents become increasingly capable of interacting with websites and applications in ways that may appear human.
We decided to bring our experience in typing biometrics, behavioral modeling, anomaly detection, and authentication together into a new capability: TypingDNA Human Likeness.
Detecting whether typing looks human
Most bot-detection systems focus on known automation signatures, devices, traffic patterns, or attack techniques.
Human Likeness starts with a different question:
How closely does this typing sample resemble natural human typing?
The system analyzes the timing and structure of a typing sample and returns a human-likeness score. Higher scores indicate behavior that looks more consistent with natural human typing, while lower scores indicate behavior that appears less human-like.
This can help identify scripted typing, browser automation, replayed or synthetic patterns, paste-based input, and AI-agent-driven form interaction.
How the model works
The model combines several layers of typing biometrics intelligence:
- comparison with multiple human typing references;
- models trained to recognize normal human typing behavior;
- typing-timing and structural features; and
- a gradient-boosted combiner that brings the signals together.
The core of the method is primarily human-centered and is not dependent on seeing every possible bot family.
That matters because automation keeps changing. A system focused only on known AI agents/bot signatures may struggle when a new automation style appears.
What we tested against
During development, we created and evaluated a range of synthetic typing families with increasing levels of realism. They ranged from simple fixed-delay and lightly randomized scripts to more advanced “human-like” generators with varied key-press and key-to-key timing, changes in typing speed, different handling of spaces and punctuation, etc.
Some families were used to calibrate the final model, while others were never shown to it and were used only to test generalization. These controlled families cannot represent every future bot or AI agent, but they help us determine whether the model is merely recognizing known generators or learning broader signals of agent.
The strong performance on the unseen-family test is therefore particularly encouraging and can reduce AI-based typed content fraud. When combined with additional signals, such as typing signatures, session behavior, and activity volume, Human Likeness could also help identify coordinated or heavily engineered automation at scale.
Performance in final testing
The final evaluation included 15,047 samples:
| Test data | Samples |
|---|---|
| Human typing | 4,717 |
| Synthetic or scripted typing | 10,330 |
At the selected operating point:
- 97.5% of all evaluated synthetic or scripted samples were detected;
- 0.7% of human samples were incorrectly classified as not human-like; and
- the model achieved a ROC AUC of 0.9985 across the complete final test set.
The more important generalization test used synthetic families that were not used to train/calibrate the final combiner:
- 96.7% detection of held-out synthetic or scripted samples;
- 0.6% human false-rejection rate; and
- ROC AUC of 0.9982.
At a stricter setting, human false rejection dropped to 0.3%, while held-out synthetic detection remained 93.6%.
These results do not mean every future bot or AI agent will always be detected. However, strong performance on previously unseen synthetic families suggests that both the current solution and the underlying human-likeness method have significant potential for detecting new forms of automated and AI-agent-driven interaction.
Designed as a risk signal
Human Likeness is intended to support broader decision workflows, such as:
- adding risk to a fraud or bot-detection engine;
- triggering step-up verification;
- protecting registration and login flows;
- detecting suspicious form completion;
- detecting non-human content creation,
- reducing automated posting and ranking manipulation in UGC platforms, and
- identifying AI-agent-driven interaction.
The current preview focuses on desktop keyboard input. Mobile implementations may use regular typing patterns alone or optionally combine them with device telemetry, while preserving the same core human-likeness approach.
The underlying method is described in this paper:
Population-Based Typing-Biometric Human-Likeness Scoring for Bot Detection
TypingDNA Human Likeness is currently available for private evaluation. Contact us to schedule a live demo and private API access.