The Treadmill Problem in Trust and Safety

Bot detection resets every quarter. Confirming a real person holds.

Bots now generate more than half of all web traffic, and the malicious share has climbed for the seventh year running. Most teams answer this the same way. They study the shape of the fake thing and build rules to catch it. The work is genuinely good, and the strongest teams tend to be the first to admit it stopped holding. The reason is structural, and it has little to do with how sharp the team is.

Detection is a treadmill

Detection recognizes a pattern, so every time the pattern changes, the team rebuilds the model to match. The other side automates a new evasion, the team ships a new rule, the evasion adapts again. Researchers now describe AI image detection as an unwinnable arms race, because generators keep closing the gap faster than detectors widen it. Survey and research teams feel the same pressure. A recent Dartmouth study ran AI through public-opinion surveys and watched it sail past every quality check, while each detection method on the market missed it entirely. A win this quarter buys a few months, then the work starts over.

The thing that holds still

There is a layer that sits underneath the whole chase. Instead of watching what a signup does and guessing whether the behavior looks human, you confirm a real person at the front door. One live human, verified once, in ten seconds, in a mobile browser, with no app to download. That confirmation holds still while everything above it keeps moving. The World Economic Forum points at the same shift, away from chasing each new fake and toward verification that assumes the fake will eventually look perfect.

Lighter and cheaper than teams expect

Durability changes the economics. When a real person verifies, that standing carries with them. Someone you remove for abuse cannot come back an hour later as a fresh account, because the person behind it has not changed. The cost of returning stops being zero, and raising that cost is the entire point. Adoption stays light too. The lift to put a humanness check in front of users looks closer to ten minutes of developer time than a quarter-long build. Plenty of teams start with a branded link sent to an existing list, watch the conversion, and make the call before anyone writes integration code.

Keep the internet human

The internet keeps getting better at manufacturing things that look human. Chasing each new version one at a time only speeds the belt up. Confirming the real person underneath compounds instead. That is the whole idea behind what we build, and it is why we keep saying the same five words. Keep the internet human.