AI Detection vs Authorship Tracking: Two Very Different Ideas

The short answer to AI detection vs authorship tracking: a detector looks at finished text and estimates, from its style, how likely it is that AI wrote it. Authorship tracking records how each piece of text entered the document while it was being written. The first produces a probability you have to interpret. The second produces a record you can read.

That difference decides what happens to real people. A student accused on the strength of a detector score has little to argue with. A writer who kept a record of which paragraphs came from ChatGPT has something concrete to show. Below: what each approach can tell you, where each breaks, and who controls the answer.

What is the difference in AI detection vs authorship tracking?

Here is the analogy I'll use throughout. Imagine two ways of deciding whether a goal in a football match was scored fairly.

The first way: someone hands you a single photo taken after the final whistle. Players standing around, the ball in the net. From that photo alone, you have to say whether there was a foul in the build-up. You can make an educated guess from where people are standing and who looks guilty. Sometimes you'll be right. But you were not there.

The second way: a referee watched the whole match and wrote down what happened as it happened. You don't need to guess. You read the notes.

AI detection is the photo. It only ever sees the finished text, and it infers the past from the present. Authorship tracking is the referee. It records events as they occur: this paragraph was typed, this one was pasted in from an AI tool, this one was pasted and then rewritten.

The two approaches answer different questions:

• AI detection asks: "Does this text look like something a language model would produce?"

• Authorship tracking asks: "How did this text get into this document?"

Those questions overlap, but they are not the same. A human can write in a flat, predictable style that looks machine-made, and an AI draft can be edited until it doesn't. The photo can mislead in both directions. The referee's notes can be wrong too, but for different reasons.

How does AI detection work, and what can it actually tell you?

Most AI detectors are classifiers. They are trained on large sets of text known to be human-written and text known to be AI-generated, and they learn statistical patterns that separate the two. Common signals include how predictable each next word is and how much sentence length and structure vary. Text that is very smooth, very even and very predictable tends to score as "AI".

The output is a likelihood, sometimes shown as a percentage of the document, sometimes as highlighted sentences. It is never a record of what happened. A detector has no idea whether you had ChatGPT open in another tab. It only knows what your words look like.

Where detectors do reasonably well

It would be unfair to say detectors never work. Jisc, the UK's digital body for further and higher education, reviewed the research in June 2025 and concluded that mainstream paid detectors perform reasonably well on text that is entirely AI-generated, as long as nobody has tried to disguise it. The same review noted that the best detection tools are better than most people at spotting AI text and have a far lower false positive rate than human markers.

That second point matters. A 2024 study by Peter Scarfe and colleagues at the University of Reading, published in PLOS One, slipped AI-written exam answers into real marking batches. According to the authors, 94% of the AI submissions went undetected by the markers. If the alternative to a detector is a tired teacher's hunch, the detector is not obviously worse.

In the analogy: if the photo shows a player holding the ball in his hands, you can be fairly sure something was off. Blatant cases show up in photos.

Where detectors break down

The trouble starts with everything that isn't blatant.

Accuracy has been poor even from the people who build the models. OpenAI released its own AI text classifier in January 2023. On its own challenge set, it correctly identified only 26% of AI-written text and wrongly labelled human text as AI-written 9% of the time. OpenAI withdrew it on 20 July 2023, citing its "low rate of accuracy". The company that makes ChatGPT could not reliably tell when text came from ChatGPT.

Some writers are flagged far more than others. A 2023 Stanford study by Weixin Liang, James Zou and colleagues, published in the journal Patterns, ran seven widely used detectors over essays written by non-native English speakers for the TOEFL exam. The detectors misclassified more than half of them as AI-generated, with an average false positive rate of 61.22%. On essays by native English-speaking US students, the same detectors made almost no mistakes. Writers with a smaller vocabulary and simpler sentence structures produce exactly the "predictable" text that detectors associate with machines.

Light editing defeats them. The Jisc review found that paraphrasing or manual changes such as adjusting sentence structure or vocabulary can significantly reduce detection accuracy, and that a market of "humaniser" tools now exists purely to do this. Someone determined to cheat edits the photo before handing it in.

Small error rates become large numbers at scale. In April 2023 Turnitin launched its AI writing indicator. Its chief product officer wrote in May 2023 that for documents with more than 20% AI writing, the document-level false positive rate is under 1%. That sounds excellent until you multiply it. Vanderbilt University did the sum in August 2023: it had submitted 75,000 papers to Turnitin in 2022, so a 1% rate would have meant around 750 student papers wrongly labelled as partly AI-written. Vanderbilt disabled the AI detector, also citing the lack of detail on how it works, privacy concerns and the bias against non-native speakers.

A worked example of the base-rate problem

Take a hypothetical teacher with 150 students across five classes, each handing in six essays a year. That's 900 essays. Suppose every one of those students writes honestly. A detector with a 1% false positive rate would still be expected to flag around nine essays as AI-written. Nine conversations with honest students, each of whom now has to prove a negative.

If some essays really did involve undisclosed AI, the detector will catch many of them, but the nine false flags are mixed in, and the score alone can't tell the teacher which is which. A probability about a population is not a finding about a person.

Why do detectors struggle most with mixed human and AI writing?

This is the gap many comparisons skip, and it is the most important one for anyone writing in 2026.

Very few people now either "use AI" or "don't use AI". A typical document might have an outline suggested by Claude, two paragraphs drafted by Gemini and then rewritten, a summary pasted from ChatGPT and left mostly as it was, and the rest typed from scratch. Many school and workplace policies allow some of that and forbid the rest.

Detectors are weakest precisely here. In the same May 2023 update, Turnitin said its sentence-level false positive rate is around 4%, and that these errors are more common in documents mixing human and AI writing, particularly at the transitions between the two. It reported that 54% of its false positive sentences sat right next to actual AI writing. In other words, when a document is mixed, the detector smudges the boundary and starts colouring your own sentences.

Go back to the photo. If two players were both involved in the build-up, a still image can't show who touched the ball and in what order. A referee who watched it happen can.

There's a second problem with mixed writing. Even a perfect detector would only tell you that a sentence looks AI-generated. It cannot tell you whether that sentence was a permitted use (say, AI-suggested phrasing under a policy that allows it) or a forbidden one. The Jisc review raised a related question about AI assessment scales and "traffic light" systems: if a policy allows some AI use, how is anyone supposed to enforce the line? A score that says "AI somewhere" doesn't answer it.

How does authorship tracking work?

Authorship tracking, sometimes called writing provenance, takes the opposite approach. Instead of judging the result, it records the process. When text enters the document, the tool notes how: typed, pasted, pasted from an AI tool, generated inside the editor. When text changes, the record changes with it.

There are a few flavours, and they differ in who does the recording and how much they record.

Version history

The simplest form is the version history built into Google Docs and Microsoft Word. It doesn't say where pasted text came from, but a document that grew over three evenings tells a different story from one that went from empty to 2,000 words in a single paste.

Process-monitoring tools

Grammarly Authorship is an opt-in feature that tracks writing activity in Google Docs, Microsoft Word and the Grammarly editor. Grammarly says it records what was typed, pasted or produced with AI help, and can replay the writing process. Grammarly's support page is clear that it does not track anything unless the user turns it on.

Turnitin Clarity, which became generally available on 15 July 2025, is a composition space that institutions license. Students draft inside it, and educators see indicators such as revision timelines and pasted versus typed text, alongside an educator-guided AI assistant. When Turnitin announced it in March 2025, the integrity blog Plagiarism Today noted that the approach gives Turnitin more control over the writing environment and could draw pushback over privacy.

Self-declared provenance

The third flavour puts the writer in charge of the labels. Assist Writer is an example: a browser-based editor where the writer marks AI text as it comes in. Text pasted with Ctrl+Shift+V (Cmd+Shift+V on a Mac), or with "Paste as AI Text" from the right-click menu, is recorded as AI text and shown in colour. Typing and normal paste count as the writer's own text. As you edit AI text, its colour gradually shifts toward normal text, so heavy rewriting is visible, and a stats bar shows the human versus AI split of the document.

It is not a detector. It does not guess anything from your style, and it doesn't generate text. It records how text entered the document, based on how you put it there.

In referee terms, this is closer to golf than football. Golfers call penalties on themselves, and the system works because the player is expected to report honestly and the record is there for anyone to check. That brings us to the limits.

Where authorship tracking fails, and why it needs good faith

Authorship tracking can be faked, and anyone selling it as unfakeable is overselling. Every form of tracking depends on the writing actually happening in the tracked space.

• Retyping. Anyone can read an AI answer on one screen and type it by hand into another. Keystroke-level monitoring may notice unusually steady typing; a simple provenance record won't.

• Off-record work. A referee only sees the pitch. Text drafted somewhere else and then pasted in shows up as a paste, but the tool can't know what happened before it arrived.

• Mislabelling. In a self-declared tool like Assist Writer, pasting AI text with a normal Ctrl+V records it as the writer's own. The record is only as honest as the person making it.

So authorship tracking depends on good faith. That is a real weakness, and I'd rather say it plainly than pretend otherwise. But notice the shape of the failure. A detector fails on honest people: it flags someone who did nothing wrong, and they have little way to answer. Tracking fails mainly for dishonest people: someone who sets out to deceive can produce a misleading record, but an honest writer's record reflects what they actually did. For the person trying to show their work, that asymmetry is the whole point.

AI detection vs authorship tracking side by side

Here's the comparison in one place, including watermarking, which is covered in the next section.

No row is perfect. The useful question is which failure you can live with. For a teacher deciding whether to open a misconduct case, wrongly accusing honest students is the most expensive failure there is. For a writer who wants to be open, a record they control is worth more than any score someone else runs on them.

Provenance vs detection for writing: where does watermarking fit?

Watermarking sits between the two. Like provenance, it is applied at the moment text is created. Like detection, it is checked later by looking at the finished text.

Google DeepMind's SynthID Text works by subtly adjusting the word choices a model makes as it generates, leaving a statistical pattern a matching detector can look for. Google open-sourced it in 2024 and it is available in the Hugging Face Transformers library. Google's own developer documentation is candid about the limits: detection is probabilistic, the watermark is less effective on factual responses, and detector confidence can be greatly reduced when text is thoroughly rewritten or translated.

In the match analogy, watermarking is a chip inside the ball. It tells you whether that particular ball was used. It tells you nothing about a ball from another manufacturer, and if someone repaints and reshapes the ball enough, the chip stops answering. A student using an open-source model, or any model that doesn't apply the watermark, leaves no trace at all.

So when people ask whether provenance or detection is the future of AI writing transparency, the honest answer is that watermarking won't settle it on its own. It is one signal from one side. Recording the writing process, whether the writer does it or a monitored editor does, works regardless of which AI tool the text came from, because it watches the document rather than the model.

How to show which parts of an essay were written by AI

If your school, client or employer allows some AI use and wants to see where it went, the cleanest evidence is a record made while you write. After the fact is too late: you can't build a referee's notes from a photo. Here is how to do it in Assist Writer.

1. Open a document. You can start in the Assist Writer editor without creating an account. In demo mode your documents stay in your browser. A free account stores up to five documents in the cloud.

2. Write your own parts normally. Typing, and normal paste with Ctrl+V, count as your own text. If you're pasting your own notes from elsewhere, use normal paste.

3. Paste AI text as AI text. When you bring in something from ChatGPT, Claude or Gemini, paste it with Ctrl+Shift+V (Cmd+Shift+V on a Mac), or right-click and choose "Paste as AI Text". It appears in the AI colour, either a five-colour spectrum or solid blue, depending on the theme you pick.

4. Rewrite in your own voice. Edit the AI text as much as you need. Its colour gradually shifts toward normal text as you change it, so a reader can see which AI passages you reworked heavily and which you left mostly as they were.

5. Check the stats bar. It shows the human versus AI split for the whole document. If your assignment sets a limit, this is where you check you're within it before you submit.

6. Share the record. On screen, the colours speak for themselves. With Pro, you can export to PDF, Word (.docx), RTF or Markdown with the author colours and the human/AI stats kept, so the reader sees the record without opening the editor. Details are on the Assist Writer Pro plan and pricing page: $4.99 a month or $52.99 a year.

7. Add a short written note. Colours show where AI was used. A sentence or two explains how: "I used Claude to suggest counterarguments for section 3 and rewrote them; the summary in section 5 is Gemini's with light edits."

The shortcuts, the right-click menu and the two editing modes (Markdown with live preview, or rich text) are all covered in the Assist Writer quick guide to AI paste and shortcuts.

What the record looks like in practice

Picture a hypothetical 1,500-word history essay. The student types the introduction and most of the argument. They ask ChatGPT for three possible counterarguments and paste the answer with Ctrl+Shift+V. They delete one, rewrite one almost entirely and keep the third close to the original. Later they paste a Gemini summary of a source as AI text and trim it.

Opened in the editor, the essay shows plain text for everything typed, a faded colour on the heavily rewritten counterargument, a strong colour on the one left mostly intact and on the summary, and a stats bar with the overall split. A teacher can see at a glance what a detector could only guess at, and can ask a much better question than "did you use AI?": "Why did you keep this counterargument as it was?"

What the record does not prove

Be straight with yourself and your reader about this. A provenance record shows what you marked. It does not prove that unmarked text is your own, because the tool can't see what happened outside it. Its value comes from consistency and openness: you used it the whole way through, the timing makes sense, your disclosure note matches the colours, and you can talk about your own sentences when asked. Combined, those are far stronger than any single score.

Which should teachers, editors and writers rely on?

My position: use detection, if at all, as a reason to have a conversation. Use authorship tracking as the evidence that conversation draws on. Never use a detector score as the verdict.

For teachers

If AI use is allowed in some form, ask students to keep a provenance record from the start and to hand in a disclosure note with it. That turns the question from "catch the cheat" into "show your process", which is fairer to the honest majority. If your institution licenses a monitored writing space, explain clearly to students what it records before they start.

If you still run a detector, read a high score as "worth asking about", then look at drafts, version history and the student's own explanation. Keep in mind the Stanford finding on non-native writers before acting on any flag.

For editors and content teams

Detectors are a poor fit for professional writing, where AI-assisted drafts are often allowed and the real question is how much human work went in. A provenance record per piece answers that directly, which also helps when a client contract limits AI content.

For students and freelance writers

Start the record before anyone asks for it. A record created during writing is believable in a way that a reconstruction after an accusation never is. If you're already facing an accusation, the Assist Writer blog has a separate guide on proving you wrote something after you've been falsely accused.

What if a detector and a provenance record disagree?

This will happen. A detector flags a paragraph; the record shows it was typed.

Look first at the kind of writing. Formulaic passages, such as definitions, method sections, lists of facts or anything written in a second language, are exactly what detectors over-flag. Then look at the record's consistency: was the document drafted over time, do the AI-marked passages match the disclosure note, can the writer explain the flagged paragraph in their own words? If all of that holds, the record should outweigh the score. The detector saw a photo; the record was there.

If the record looks thin, for instance a whole document arriving in one normal paste with no history, then the record isn't telling you much either, and a conversation about the work is the only fair next step.

Common mistakes with both approaches

• Reading a detector percentage as "how much AI was used". It is a model's confidence about style, not a measurement of anything that happened.

• Treating a provenance record as a lie detector. It shows what was recorded, and it depends on the writer using it honestly.

• Starting the record after the fact. A record built after an accusation has much less weight than one kept from the first sentence.

• Pasting AI text with a normal paste by habit. In Assist Writer, Ctrl+V means "my own text". Get used to Ctrl+Shift+V for anything that came from an AI tool, or use the right-click menu when unsure.

• Assuming watermarking will solve it. It only covers text from models that apply it, and it weakens with rewriting.

• Requiring monitoring without explaining it. People reasonably object to being watched without knowing what's collected. Say what is recorded, who sees it and why.

The bottom line

One approach judges the photo; the other keeps the referee's notes. Detection has a place in spotting wholesale AI text, but it is weakest in mixed documents, where most real writing now sits, and when it fails, it fails on honest people. Authorship tracking depends on good faith. For anyone who wants to be open about their AI use, it is still the only approach that lets them show their work instead of waiting for someone else to guess.

Frequently asked questions

What's the difference between AI detector and authorship tracking evidence?

An AI detector analyses finished text and estimates how likely it is that a language model wrote it, based on style patterns. Authorship tracking records how text entered a document while it was being written: typed, pasted, or pasted from an AI tool. A detector gives a probability that someone has to interpret; tracking gives a record of events that anyone can read and question.

Are AI detectors accurate enough to prove someone used AI?

Not on their own. The best paid detectors do reasonably well on long, unedited, fully AI-written text, according to a 2025 Jisc review. But paraphrasing weakens them, a 2023 Stanford study found they flagged over half of non-native speakers' essays, and Turnitin has said its errors cluster where human and AI writing meet. A score is a reason to ask questions, not a verdict.

Can authorship tracking be faked?

Yes. Every tracking method depends on the writing happening inside the tracked tool. Someone can retype AI text by hand, draft elsewhere and paste it in, or, in a self-declared tool, paste AI text as their own. Tracking therefore relies on good faith. Its strength is that an honest writer's record reflects what they did, whereas a detector can wrongly flag someone who did nothing wrong.

Is Assist Writer an AI detector?

No. Assist Writer does not guess authorship from writing style and does not generate text. It is a browser-based editor that records how text entered the document: text pasted with Ctrl+Shift+V (Cmd+Shift+V on Mac) is marked as AI text and shown in colour, typing and normal paste count as your own, and a stats bar shows the human versus AI split.

Does a provenance record prove I didn't use AI?

It shows what you recorded, not everything you did. A tool can't see work done outside it, so unmarked text is not automatically proven to be yours. A record carries weight when it was kept from the first sentence, the document grew over time, the AI-marked passages match your written disclosure, and you can explain your own sentences when asked. Together those are far stronger than any detector score.