Why 'Real or Fake' Is the Wrong Question for AI Detection

The tidy binary of “real or fake” has defined public debate about AI content since the first chatbots went mainstream. But Pangram’s Max Spero argues that framing is not just simplistic — it is actively misleading. In practice, the detection problem is not about separating pure human text from pure machine text. It is about identifying the vast, messy middle where the two are indistinguishable.

The spectrum problem

Modern AI-generated content rarely arrives untouched. Users edit outputs, paraphrase sentences, splice in their own writing, or run text through multiple tools before publishing. A detector that scores a document on a simple binary scale will fail on these hybrids — and hybrids are increasingly the norm, not the exception. The meaningful question is no longer “was AI involved?” but “how much, and where?”

The stakes are considerable. Misinformation campaigns, academic integrity, content moderation, and even legal discovery all depend on reliable attribution. Yet a verdict-based approach gives false confidence: a document flagged as “human” may still contain machine-written passages, while a “fake” label obscures the human effort that shaped it. Spero’s critique suggests that confidence scores and segment-level analysis matter far more than a single yes-or-no answer.

If detection is to keep pace with generation, it must become granular and probabilistic — flagging specific passages, estimating the proportion of machine involvement, and acknowledging uncertainty rather than hiding it. That shift is harder to communicate than a simple label, but it is the only honest path forward. The future of AI detection is less about rendering a verdict and more about drawing a map of influence, one that shows readers exactly what they are looking at and why.

undefined: The DAN Brief — .