The Responsible Disclosure Spectrum: From ‘AI-Assisted’ to ‘AI-Generated’

“AI-assisted” and “AI-generated” get used like they mean the same thing. They don’t. One describes a tool that helped you do your own work; the other describes work the tool mostly did — and most of what people actually produce lands somewhere in the messy middle.

That middle is where disclosure gets hard. If you can’t sort your own writing into a clean box, how are you supposed to label it honestly? The answer isn’t a better box. It’s a spectrum, and a few honest questions about where you sit on it.

Key takeaways

• “AI-assisted” and “AI-generated” sit at opposite ends of a spectrum, not in two tidy categories — most real work lands in between.
• What you disclose should track how much of the thinking, structure, and final wording came from you versus the model.
• A useful test: could you defend every claim and rebuild the argument without the tool? The more often the answer is no, the closer you are to “generated.”
• Vague disclosures (“AI was used”) help no one. Name what the tool did so the reader knows what they’re looking at.
• Disclosure and AI detection are separate problems. A clear note protects your honesty; it doesn’t change a detector’s score.

Why two labels aren’t enough

People reach for “AI-assisted” when they want to sound responsible and “AI-generated” when they’re describing someone else’s shortcut. The labels carry a moral charge they shouldn’t.

Think about a single essay. You might brainstorm with a model, write the draft yourself, paste it back for an editing pass, then accept half the suggestions. Which label fits? “Assisted” undersells the editing pass. “Generated” is just wrong — you wrote it. Neither word captures the truth, so you end up either overclaiming or underclaiming.

That’s the trap. A binary forces you to round, and rounding is where dishonesty sneaks in — usually without anyone deciding to be dishonest.

The spectrum, from light to heavy involvement

It helps to picture five rough points along a line. The boundaries blur, and that’s fine. The goal is to locate yourself, not to win an argument about definitions.

1. Mechanical assistance

Spell-check, grammar fixes, a synonym here and there. The model touches your surface but not your meaning. This is the level almost nobody expects you to disclose, and most AI policies say so outright.

2. Editing and polishing

You wrote it; the tool tightened it. It smoothed a clunky transition, trimmed a run-on, suggested a stronger verb. Your ideas, structure, and evidence are untouched. Light disclosure is reasonable here, especially in formal or academic settings.

3. Collaboration

Now the lines blur. You and the model traded ideas. Maybe it proposed an outline you reshaped, or drafted a paragraph you rewrote until it sounded like you. The thinking is shared. This is the heart of the spectrum, and it’s where honest, specific disclosure matters most.

4. Heavy generation with human direction

The model wrote most of the prose. You steered it with prompts, picked what to keep, fact-checked, and stitched sections together — but the sentences are largely its sentences. You’re an editor and director, not the author of the words. Disclosure should say so plainly.

5. Fully AI-generated

You prompted, the model produced, you pasted. Little or no human shaping of the actual content. If you’re publishing or submitting this, “AI-generated” is the only honest label.

Most work you’ll ever produce sits at points 2 through 4. The skill isn’t memorizing the levels — it’s being honest about which one you’re at this time.

A test that cuts through the labels

When the labels fight you, ask a process question instead: Could I defend every claim and rebuild the argument without the tool?

• If yes, easily — you’re near the assisted end. The tool helped you say what you already understood.
• If no, not really — you’re near the generated end. You’re carrying ideas you can’t fully stand behind.

This works because it shifts the question from *what software did I open* to *whose understanding is on the page*. A grammar tool doesn’t threaten your authorship. A model that supplied an argument you couldn’t reconstruct does. The test scales with how much of the thinking was actually yours, which is the thing disclosure is supposed to communicate.

Mini-scenario: two students, same tool, different disclosures

Maya and Devin both use the same AI model for a history essay.

Maya outlines her thesis, writes a full draft from her notes, then runs it through the model once to fix awkward phrasing and tighten her conclusion. She accepts maybe a third of the edits. Her honest disclosure: *”I used an AI tool to copy-edit phrasing and tighten my conclusion. The research, thesis, and argument are mine.”* That’s point 2, accurately described.

Devin prompts the model to “write a 1,000-word essay on the causes of the French Revolution,” lightly rearranges the paragraphs, swaps a few words, and submits. If he writes the same disclosure as Maya, he’s lying — not by inventing a fact, but by borrowing her label. An honest version: *”This essay was primarily AI-generated from my prompts; I edited and reorganized it.”* That’s point 4 or 5.

Same tool. Same word count. Completely different places on the spectrum — and completely different honest disclosures. The label isn’t determined by the software. It’s determined by the work.

How to write a disclosure that actually informs

The fix for vague disclosure is specificity. Name the task, not just the tool.

Weak: *”AI was used in this document.”* That tells a reader nothing — it could mean a spell-check or a ghostwritten draft.

Better: *”I used an AI model to brainstorm angles and tighten transitions in the final section. All research, claims, and the argument are my own.”*

A good disclosure answers three things: what the tool did, what stayed human, and how much. Get those across in a sentence or two and you’ve located yourself on the spectrum without needing the labels at all. For more on where the rules come from and how to read them, browse our guides on AI writing and ethics.

A few habits that make this easier:

• Keep version history. Saved drafts and document history are the strongest evidence of where your work actually came from. They beat any after-the-fact label.
• Disclose at the right altitude. A blog footer, an academic AI-use statement, a note to a client — match the formality to the context.
• Ask when the policy is silent. Treat a missing rule as a reason to email, not a green light. A written go-ahead from the person grading or paying you outweighs any general guideline.

Disclosure is not the same as passing a detector

Here’s a confusion worth killing. Disclosing AI use and getting flagged by an AI detector are two separate problems, and solving one does nothing for the other.

Detectors don’t read your disclosure. They score the statistical texture of your text and return a probability, not proof. They get it wrong in both directions — OpenAI even retired its own AI Text Classifier in July 2023 because it wasn’t accurate enough, and a 2023 Stanford study (Liang et al.) found detectors disproportionately flag non-native English writers whose work is entirely their own.

So you can disclose honestly and still get flagged. You can also write something fully human and get flagged. The disclosure protects your *honesty*; it can’t move a *number*. Those are two different jobs, and it’s worth keeping them apart in your head.

Frequently asked questions

Where exactly is the line between AI-assisted and AI-generated?

There isn’t a single line — that’s the point of treating it as a spectrum. A rough cutoff most people find fair: if the model produced sentences or paragraphs that survived into your final draft largely intact, you’re in “AI-generated” territory for those parts. If the model only nudged your own sentences, that’s “assisted.” When you’re unsure, describe what happened instead of picking a label.

Do I have to disclose AI if I only used it for grammar and spelling?

Usually not. Most policies treat a grammar suggestion like a spell-checker’s red underline. The gray zone starts when a tool rewrites whole sentences or restructures your argument. If the tool changed your meaning, wording, or organization in a way a reader would notice, one honest line removes the ambiguity later.

Isn’t any disclosure just admitting I cheated?

No. Disclosure describes a process; it doesn’t confess to a violation. Using a permitted tool and saying so is the transparency most policies actually ask for. What gets people in trouble is unapproved use that’s then hidden — not the note describing approved use.

If I disclose honestly, can a detector still flag my work?

Yes, and the two are unrelated. Detectors return a probability, not proof, and sometimes flag writing that’s entirely human. Your disclosure plus saved drafts are what protect you in a dispute. A flag is a prompt to explain your process, not a verdict on your honesty.

What disclosure language should I actually use?

Be specific about the task, not just the tool. Instead of “AI was used in this document,” write something like “I used an AI model to brainstorm angles and tighten transitions; all research and argument are my own.” Naming the role tells the reader where on the spectrum you sit.

The short version

Stop trying to cram your work into “assisted” or “generated.” Locate it on the spectrum, ask whether you could rebuild the argument without the tool, and then describe — plainly and specifically — what the model did and what stayed yours. That’s all honest disclosure ever was. And when you want to see how your edited draft actually reads to a detector, run a quick check on a draft before you submit it.