How Detection Tools Handle AI-Assisted vs. Fully AI-Generated Writing

Detection tools can’t genuinely separate AI-assisted writing from fully AI-generated writing — they score how machine-like the words read, not how the document was made, so a polished human draft and a touched-up AI draft can land on the same number. That blind spot sits at the center of nearly every modern authorship dispute, because almost nobody’s workflow is purely one or the other anymore. Understanding how tools behave on the messy middle — the real subject of any honest AI assisted vs AI generated detection comparison — is what makes their scores readable.

Key takeaways

• Detectors observe text, not process. “Assisted” and “generated” are process categories, invisible in the artifact.
• Mixed documents produce blended, unstable scores — the ambiguous middle range where tools are least reliable and stakes are highest.
• Sentence-level highlighting is more informative than one document score, but it inherits every error mode at higher resolution.
• AI *editing* of human text moves scores toward “AI” — meaning tools punish some permitted assistance and miss some prohibited generation.
• The distinction that matters is set by policy and proven by process (drafts, history), never by a percentage.

AI-assisted vs AI-generated detection: two categories the tool can’t see

Start with what the words mean in practice. AI-assisted writing is human-driven: you drafted the ideas and structure, and a model helped — tightening sentences, fixing grammar, suggesting a transition. AI-generated writing is model-driven: the model produced the substance, and you prompted, maybe edited. Institutions draw their lines in very different places along that spectrum, which we’ve mapped in detail in where the line sits between AI-assisted and AI-generated work.

Now the crucial point: a detector receives only the final text. It computes how predictable the word choices are, how even the rhythm is, how closely the passage resembles machine output in its training data. Nothing in that computation touches *who drove*. The categories everyone argues about are simply not features the classifier has.

What actually happens to a mixed document

Feed tools a half-human, half-AI draft and the behavior splits by architecture.

Document-level scorers output something like “46% AI.” That number is not a measurement of proportion — it’s a probability blended from competing signals, and it lands in exactly the range where every vendor’s own documentation gets cagey. Turnitin, notably, treats low scores as unreliable enough to display with an asterisk and warns against acting on them. The middle of the scale is where false positives and false negatives crowd together.

Sentence-level tools — GPTZero and similar — classify passage by passage and highlight the machine-like stretches. On paper this is the right shape for the hybrid problem, and it *is* more useful: you can see whether flags cluster in one suspicious section or sprinkle randomly. But resolution multiplies error surface. Individual sentences are short, and short text is where statistical classification is weakest — a distinctive human sentence gets flagged, a lightly reworded AI sentence passes, and the highlight map acquires an authority its per-sentence accuracy can’t back.

Either way, the tool’s output on hybrid text is a fog with a number on it.

The asymmetry nobody designed

Here’s the perverse part. Because detectors key on statistical smoothness, AI *assistance applied to human text* — the widely permitted use — actively pushes scores toward “AI.” Accept enough phrasing suggestions and rewrites and your genuinely self-authored argument starts reading machine-polished; the Grammarly-then-GPTZero pipeline has burned plenty of students whose only sin was accepting suggestions. Meanwhile, *human effort applied to AI text* — paraphrasing tools, or a determined manual rework of generated content — pulls scores toward “human,” which is exactly the evasion result Krishna and colleagues demonstrated at NeurIPS 2023.

So the technology systematically risks punishing the sanctioned behavior and missing the prohibited one. It’s not that tools fail randomly on the assisted-vs-generated distinction; on the margins, they can get it *backwards*. We dug into the editing side specifically in whether detectors can tell AI editing from AI writing — the short answer is no, with consequences in both directions. And all of this stacks on top of the baseline unfairness that detectors flag non-native English writers’ unassisted prose at wildly elevated rates, as the Stanford Patterns study documented.

Reading scores sanely on hybrid work

Given all that, here’s the posture that survives contact with reality.

If you’re the writer: know your policy first — assisted and generated are defined by your instructor or employer, not by any tool. Keep process evidence as a habit: version history, notes, prompts used. Check your own draft before submission, because you inherit the score’s errors either way; a sentence-level view shows you *which* passages read machine-flat so you can revise them into your own voice. You can try it on your own text for free.

If you’re the evaluator: treat a mid-range score on a suspected hybrid as a prompt for process questions, never as a finding. “Walk me through this draft” separates authors from prompt-passers far more accurately than any classifier — a genuine author explains choices effortlessly. And weight the known failure modes: polished-by-tools human work scoring high, reworked AI scoring low, second-language writers scoring high for reasons that have nothing to do with AI at all.

If you’re setting policy: write the assisted/generated line explicitly and say how it will be checked. A policy that silently outsources the distinction to a detector is outsourcing it to a tool that cannot make it. For checking at scale, see what each plan handles — and whatever tool you use, pair it with process evidence rather than substituting for it.

For more on how these systems work and where they break, you can browse the rest of our writing on detection.

Frequently asked questions

Can AI detectors tell the difference between AI-assisted and fully AI-generated writing? Not reliably, and most don’t really try. Detectors score how statistically machine-like text reads; they have no way to observe your process. A human draft heavily polished by AI and an AI draft lightly edited by a human can produce similar scores, even though one is assisted writing and the other is generated writing with cosmetics.

Will using Grammarly or an AI editor get me flagged? It can move your score, yes. Grammar and phrasing suggestions push word choice toward statistically safe patterns, and full-sentence rewrites do so strongly. Accepting scattered fixes rarely flips a verdict on its own, but accepting wholesale rewrites across a document makes human writing read measurably more machine-like. Keep drafts and version history so you can show your process if asked.

How do detectors score a document that’s half human, half AI? Roughly as a blend, with the details depending on the tool. Document-level scores average competing signals and often land in an ambiguous middle range. Sentence-level tools highlight which passages read as machine-like, which is more informative but still error-prone — human sentences get highlighted and edited AI sentences slip through.

Is AI-assisted writing treated as cheating? It depends entirely on the rules you’re under, and they vary wildly. Some courses ban all AI contact; many allow editing help but not generation; workplaces often encourage both. The detector doesn’t know which policy applies — it outputs a number either way. What matters is disclosure where required and matching your use to the stated policy, not the score itself.

What’s the fairest way for a teacher to read a mid-range score on a mixed draft? As a question, not a finding. A 40-60% score on a hybrid document is consistent with permitted editing help, unpermitted generation, or a false positive on distinctive human writing. The reliable next step is process evidence: version history, earlier drafts, and a short conversation about the work’s choices — things a genuine author handles easily.

The bottom line

“Assisted or generated?” is the question everyone wants answered, and it’s precisely the question detection tools cannot answer — they grade the texture of the final words, while the distinction lives in the process that produced them. Scores on hybrid documents are blends of competing signals, least trustworthy exactly where the modern cases are. Use detectors to find machine-flat passages, use policy to define the line, and use process — drafts, history, conversation — to establish which side of it a piece of writing actually sits on.