How does AI detection work? I designed an EXPERIMENT to show you

AI detection systems don’t judge text sentence-by-sentence; they score documents by looking at full-context patterns. That matters because a workflow that “humanizes” or rewrites sections until a free checker reports 0% can still trigger a high AI score later when a submission tool (such as Turnitin) scans the entire paper at once.

The core mechanism described is context dependence. Detectors analyze the whole document—especially how different passages appear together—rather than treating each paragraph as an isolated unit. This means proximity effects can occur: human-written sections can be flagged as AI-generated when they sit near AI-written sections or when AI-like patterns cluster across boundaries. The practical takeaway is straightforward: feed the detector the full document (or at least all AI-generated text together), not fragmented chunks.

A key warning targets a common workaround: splitting long text into pieces to fit the word limits of free online detectors, then running each chunk separately and pasting the “cleaned” results back into the original document. According to the explanation, this changes what the detector can see. If a free tool only receives half the text, it can’t evaluate cross-document patterns that emerge only when the complete set of passages is present. When the fully assembled document is later scanned by a tool that has access to 100% of the content, the scoring can swing dramatically.

To demonstrate the effect, an experiment was run using a hypothetical “findings” chapter generated by ChatGPT (spelled “Chad GBD” in the transcript). The generated chapter was first placed into Microsoft Word, then split into two parts because it was too long for the free detector being used. Each half was scanned and edited until the free tool reported 0% AI-generated content for that section.

After both halves were pasted back into a single document, a subsequent scan using a detector that considers the full document produced the opposite result: the combined submission was flagged at 100% AI-generated content. The explanation attributes the reversal to pattern detection across the entire text—patterns that the chunk-based workflow prevented the first tool from observing.

Beyond context, the transcript also describes how detectors rely on large language corpora containing millions of human-written and AI-generated documents. Detectors compare expressions against these datasets and compute likelihoods for combinations of phrasing appearing together, even when individual phrases look ordinary. The math is described as too complex for most people to replicate manually.

The bottom line is not that rewriting is pointless, but that evaluation depends on what the detector is allowed to see. For longer assignments, the safest approach is to provide full context to the same type of checker used for submission, because fragmenting text can create a false sense of success—and then a later scan can flag the assembled work as highly AI-generated.