When you mix your own writing with AI's, how often do AI-text detectors wrongly call it machine-made, and has anyone measured that?

This explores how often AI-text detectors wrongly accuse writers when a draft mixes human and AI work. The short answer is that the collection has no study measuring detector false-positive rates on blended drafts. Treat any specific percentage you've seen elsewhere as something to check against its source. What the collection does have is the surrounding evidence, and it explains why that number is so hard to pin down and why it matters more than headline accuracy suggests.

Start with the human baseline. A review of 30 studies found that people tell AI-made content from human content at roughly coin-flip accuracy, across text, images, and voice. Their accuracy hasn't improved as AI output has become more realistic (Can people reliably spot content made by AI?). That's why institutions turn to automated detectors in the first place. It also means there's no reliable human check on what a detector decides. A mixed draft makes this worse, because the 'correct' label is unclear even before any tool is involved. If a person outlines, a model drafts, and the person rewrites, there is no clean ground truth to measure false positives against.

The most useful idea for thinking about detector error rates comes from a note that isn't about detection at all. It argues that a system with 95% accuracy, used at scale, still wrongly convicts thousands of people, and that impressive accuracy figures can hide basic statistical mistakes (Can AI models be truly free from human bias?). The same arithmetic applies to detectors. When most submissions are honest, even a small false-positive rate means many of the people flagged are innocent. Mixed drafts fall exactly where a detector's confidence means least. A related note explains why: AI output is inherently changeable, shifting with sampling, prompt wording, and context (Why does AI output change with every prompt and context?). That makes it hard to find a stable 'AI signature' that survives human editing.

Two neighboring threads show what's at stake. In hiring, applicants use AI and prompt tricks to get past filters while recruiters spend large parts of their week screening out what they see as spam. That's an escalating loop in which filtering errors fall on real people (Are job applicants and employers locked in an escalating AI arms race?). And AI gatekeeping can be uneven across groups. One study found guardrails refusing requests at different rates depending on the user's apparent age, gender, or ethnicity (Do AI guardrails refuse differently based on who is asking?). That study is about refusals, not detectors, but it raises a fair question: are a detector's false positives spread evenly across writers, or do they cluster on certain styles of writing?

The collection also points to a different design. Instead of a detector that hands down a verdict, a system can highlight what a human reviewer should look at and leave the judgment with that person. In the decision-making studies behind this idea, that approach reduced anchoring bias (Can AI guidance reduce anchoring bias better than AI decisions?). Another note finds that we still lack ways to measure whether an AI system's errors stay visible and can be challenged (How can we measure whether AI errors stay visible and recoverable?). For a writer falsely flagged on a draft they co-wrote with AI, that is the real gap: the error rate is unknown, and there is often no way to see or contest the error.