Every system that needs to tell a human apart from a machine is built on the same brittle assumption: that the machine moves differently, and you can spot it from the statistics. Mean flick speed, snap-to-target acceleration, reaction time, distribution of inter-event timings. Behavioral anti-cheat, bot detection, fraud scoring, CAPTCHA — all of it.
That assumption is now false. Modern AI is trained to reproduce exactly those statistics. The surface a detector analyzes is precisely the surface a humanizer spoofs. And the pressure is only increasing: as agents get good enough to drive a mouse, fill a form, or play a ranked match, "prove a human did this" breaks across games, fraud, identity, and RPA simultaneously.
The generative reframe
The detectors that survive scrutiny share one move: model the causal generative process of human motor control, not its output histogram.
Human cursor and aim movement isn't just noisy — it's the output of a stochastic optimal-control loop with well-characterized structure: signal-dependent noise (variance that scales with command magnitude), a stretch-reflex feedback delay, and a 1/fᵝ power spectrum of correction sub-movements that emerges from limb-plus-neural-feedback dynamics. Fit a hidden-state stochastic differential equation of that plant per session, and you can compute a likelihood ratio: how probable is this trajectory under the human-motor model versus a best-fit smoothed-AI model?
This is fundamentally different from classifying features. To fool a likelihood-ratio test on the causal covariance kernel, an AI can't match aggregate speed and acceleration — it has to instantiate a real-time closed-loop controller with correct signal-dependent noise and feedback delay, i.e. solve a meaningful slice of the motor-control problem inside its injection latency budget. That's dramatically more expensive than the Bezier-curve smoothing that defeats today's classifiers.
Honest about the arms race
The defense isn't permanent — once the kernel is public, attackers can train a generator to sample from it. But a generator that matches the output distribution still tends to miss the conditional structure (noise scaling on instantaneous command, feedback-delay phase), and closing that gap pushes the attacker toward actually simulating the plant. So the right way to talk about it is a measured cost gradient, not a wall: ship a red-team harness that trains humanizers against the detector and quantifies the cost-to-evade. That number — how much neuromuscular machinery an attacker must add to pass — is the defensible result, and it compounds with a corpus of human-vs-synthetic signatures that no static product can replicate.
Why gaming is the wedge, not the market
Competitive games are the hardest proving ground: an adversary actively optimizing against you, in real time, at scale. Win there and the same engine — was this input stream produced by a human neuromuscular system? — sells into the far larger markets that need the same answer: bot and RPA detection, payment fraud, synthetic-input detection, and continuous CAPTCHA-replacement authentication. In an era where AI increasingly drives real UIs, "human or machine control?" becomes a universal need. Gaming just gets you there first.