The problem

The preponderance of LLM generated content presents the problem: how can I prove that someone has done work to complete a task, or how can I prove that a document I'm reading was written or reviewed by a human?

A common approach is to run some algorithm on the text itself to see: is this text something it seems an LLM would produce? This is problematic in two ways: it's a moving target as LLMs continue to change their own internal word choices, and more importantly small edits like removing all em-dashes after the text is generated can bypass these detectors.

What we do differently

This website proposes an alternative: obfuscated keylogging. If we transform the actions a user takes (pauses, pastes, other special keys) into a transferrable chunk of data, we can pair that data with the content itself and perform some static analysis on them together to provide confidence that the content was produced via the associated human session.

The internal mechanisms of how the singature and analysis work are omitted here, per the clear objection: someone will figure out how to spoof these signatures. And when they do we can increment the version of the signature or analysis algorithms; this isn't an undefeatable approach, but just as its simple to break a window or pick a front door lock, the presence of the barrier itself denies opportunists.