people assume ai detectors are like forensic tools. paste in text and they'll tell you exactly which model generated it, when, and with what prompt.
that's not remotely how they work.
what detectors actually tell you
they give you a probability: "this text has a 73% chance of being ai-generated." that's it.
they can't tell you which model wrote it. chatgpt, claude, gemini, llama — they all get lumped together as "ai." some detectors claim to distinguish between models but the accuracy on that is terrible.
they can't tell you which parts are ai and which are human. some tools highlight individual sentences but the confidence on per-sentence detection is much lower than document-level. a highlighted sentence might just be formal writing.
they definitely can't tell you who the human author is. there's no fingerprinting. no identification. it's a binary classification at best.
the technical reason
detectors are trained on two categories: "text that looks like ai" and "text that looks like human." they learn statistical differences between the two categories. that's the full extent of their capability.
distinguishing between chatgpt-4 and claude would require a much more sophisticated model with access to the specific output distributions of each model. nobody has that at scale.
identifying individual human authors would require stylometric analysis — a completely different technology. some research exists but it's not in commercial detectors.
why this matters
the limitation means detectors are blunt instruments. they answer a yes/no question with varying accuracy. they can't provide context, attribution, or nuance.
when a professor sees "67% ai-generated," that number tells them almost nothing useful. which 67%? how confident is the tool? could it be a false positive? the detector doesn't say.
the implication for ai-assisted writing
if detectors can't tell who wrote what or which tool was used, the distinction between "ai-generated" and "ai-assisted" is invisible to them. text that was 90% ai and lightly edited looks the same as text that was 90% human and slightly assisted.
this is why the binary ai/human classification is broken. real-world writing exists on a spectrum. detectors can't see the spectrum. they see a coin flip.
what actually works
instead of trying to classify text after the fact, focus on producing text that's genuinely good. natural rhythm, real variation, actual substance.
paraai's fine-tuned models produce text with human writing patterns because they learned those patterns from real writing. detectors classify the output as human-like because the statistical properties match. not because we're tricking them — because the text genuinely has those properties.
untraceable ai writing doesn't need to fool a forensic analysis. it just needs to be good enough that a probabilistic classifier says "looks human." and good writing clears that bar easily.