Beyond the Binary: How Tools Like EditLens Measure the Extent of AI in Writing

For the past three years, the question hanging over every essay, email and explainer has been a simple one: did a human write this, or did a machine? But that question is starting to feel quaint. The honest answer for a growing share of the text we read is: both, in some proportion we can't quite see.

A new generation of detection tools — EditLens among them — is trying to put a number on that proportion. Instead of returning a binary verdict, they aim to estimate the extent of AI involvement in a piece of writing: how much was drafted by a model, how much was lightly polished, how much is genuinely the author's own. For Australian universities, newsrooms, hiring managers and government agencies wrestling with the post-ChatGPT writing economy, that shift in framing matters more than any single product launch.

The problem with yes/no AI detection

The first wave of AI detectors — the GPTZeros and OpenAI Classifiers of 2023 — treated detection as a classification problem. Feed in a document, get back a probability that it was machine-generated. The trouble is that real-world writing almost never fits that mould anymore.

Consider a typical workflow in 2025: a journalist drafts an article in their own voice, asks a model to tighten a few clunky paragraphs, runs the result through a grammar tool, then rewrites the lede by hand. A student outlines an essay themselves, generates a first draft with a chatbot, then heavily rewrites the argument. A public servant dictates rough notes, asks an assistant to turn them into a briefing, and edits the result line by line.

None of those documents are "AI-written" in the way a fully prompted essay is. None are purely human either. A binary classifier will either flag them all as suspicious or miss them all as clean, depending on where the threshold sits. Worse, the same classifiers are notorious for misfiring on non-native English speakers — a serious fairness problem in a multicultural country like Australia, where universities have already been forced to pull back on aggressive detection policies after high-profile false-positive cases.

What EditLens-style tools try to do differently

The premise behind a tool like EditLens is that AI involvement exists on a spectrum, and that the linguistic fingerprints of editing are different from those of full generation. A model that rewrites a paragraph tends to leave statistical traces — characteristic word choices, sentence rhythms, punctuation habits, transition phrases — but it also inherits structure and ideas from the human input underneath. The signal is mixed, and the goal is to estimate the mix.

In practice, that means a quantified output. Rather than "87% likely AI," the report might read something like: roughly a quarter of this document shows signs of substantial machine editing, concentrated in the middle sections, with the opening and closing paragraphs appearing largely untouched. Some tools highlight specific spans; others give a document-level percentage; the more ambitious ones try to distinguish between "generated from scratch," "heavily rewritten," "lightly polished" and "human only."

It is a more honest framing of the problem. It is also, technically, a much harder one.

How the measurement actually works

Most edit-detection systems lean on the same underlying intuition that powered earlier detectors: large language models produce text with subtly different statistical properties to human writing. They favour certain function words, avoid certain rare constructions, and produce sentences whose probabilities — under another language model's gaze — cluster in telltale ranges.

To turn that into an edit estimate, researchers typically do one or more of the following:

The result is an estimate, not a measurement in the physics sense. A 30% score is not a guarantee that 30% of the words came from a model — it is a model's best guess about how much of the document looks like it has been through machine hands.

Why "extent" matters more than "presence"

The shift from presence to extent changes the conversations institutions can have. A binary detector forces a binary response: punish or ignore. A quantified estimate invites nuance.

An Australian university could, in principle, set different thresholds for different assessment types. A reflective journal might tolerate light AI polishing; a take-home exam might not. A newsroom could distinguish between reporters using AI to clean up transcripts and reporters generating copy wholesale. A recruiter could read a cover letter knowing that some level of AI assistance is now the cultural norm, the way spellcheck and Grammarly were before it.

It also shifts the moral conversation. "Did you use AI?" is increasingly an unhelpful question, because almost everyone does, somewhere in their workflow. "How much of the thinking is yours?" is the better one — and an extent score, however imperfect, gives both sides a shared object to argue about.

The limits worth being honest about

None of this means edit-detection tools should be treated as ground truth. Several caveats deserve to travel with every score.

First, the underlying models drift. As frontier models improve and as more humans absorb AI-style rhythms into their own writing, the statistical gap that detectors rely on narrows. Tools trained in 2024 may be poorly calibrated by 2026.

Second, adversarial paraphrasing — running AI text through another model, or through deliberate humanising tools — can collapse detection accuracy. Anyone determined to evade a detector usually can.

Third, false positives remain a serious harm. Flagging a student, a job applicant or a writer as "30% AI" when they wrote every word themselves is not a small error; it can affect grades, employment and reputation. Any institution adopting these tools needs an appeals process and a clear rule that scores are evidence, not verdicts.

Fourth, the privacy implications of routing documents through third-party detection services — particularly in education and government — are non-trivial under Australian privacy law. The convenience of a web upload can mean handing student work or internal drafts to an offshore provider.

What this means for readers and writers

For everyday Australians, the practical upshot is twofold. As readers, we should get used to the idea that almost everything we read online sits somewhere on the human–AI spectrum, and that "authentic" is becoming a question of degree rather than kind. As writers, we should expect that our own work will increasingly be scored — by employers, editors, educators, platforms — for how machine-touched it appears, fairly or not.

The most useful thing tools like EditLens may end up doing is not catching cheaters. It is giving institutions a vocabulary for the messy middle: the documents that are neither purely human nor purely machine, which is to say, most of them. Whether we use that vocabulary to draw smarter lines or simply to police harder will say more about us than about the technology.

Related on Bleen

Sources