How To See If Someone Used AI | Spot Tells That Hold Up

You’re reading a post, grading a paper, or reviewing a draft at work, and a thought pops up: was this written with a chatbot? You’re not alone. AI writing tools are everywhere, and people use them in different ways. Some use them for brainstorming. Some use them to polish grammar. Some paste in a prompt, accept the first output, and submit it.

This article gives you practical ways to spot AI use without turning it into a witch hunt. You’ll learn signals that carry weight, quick checks you can run, and fair ways to handle the situation if stakes are real. You’ll also learn why “AI detectors” can mislead you, so you don’t treat a score like proof.

What counts as AI use in writing

“Used AI” can mean a lot of things. One writer might ask a model for an outline, then draft everything by hand. Another might generate full paragraphs, then edit them into their own voice. A third might submit raw output with light tweaks.

So your job isn’t mind reading. Your job is to decide whether the text shows signs of automated drafting, and whether that matters in your setting. A teacher may care about original thinking. A manager may care about accuracy and liability. A publisher may care about voice match and sourcing.

If you need a simple rule of thumb: treat AI use like any other tool use. If the result meets your standards for truth, originality, and process, it may be fine. If it fails those standards, the tool details become relevant.

Seeing if someone used AI for writing with a reality check

The strongest signal isn’t a detector score. It’s whether the writing behaves like it came from a person who knows the details. AI text can sound smooth while being oddly empty. It can also state claims with total confidence, even when the claims are wrong.

Start with a reality check that takes five minutes. Pick three factual claims, three named things (people, products, places, laws, tools), and three numbers with units. Then verify them. If the piece can’t survive that check, you’ve learned something useful, even if you never “prove” tool use.

Look for detail that only the writer could know

Human writing often carries grounded signals: a dated meeting note, a file name, a version number, an observed mistake, or a decision made under a constraint. AI can invent these, yet it often avoids specifics to stay safe.

If the text says “research shows,” ask what was read, which pages, and what was pulled from them. If it describes an “experience,” ask for the steps taken, the time spent, and what went wrong on the first attempt. People who did the work usually remember friction.

Check whether the piece owns its choices

Strong writing makes choices. It picks one definition, one angle, one set of criteria, one set of trade-offs. AI drafts often sound balanced in a way that feels slippery: they present options, then refuse to choose.

Ask one direct question: “What’s your main claim in one sentence?” Then ask: “Why that claim, not the other one?” A writer who owns the idea can answer fast, even if the prose needs edits.

Watch for a voice that never changes

Many AI drafts keep a steady, polite tone from start to finish. That can be fine. Still, real writers often shift gears. They speed up in a punchy section. They slow down for a careful caveat. They show mild annoyance at a common mistake. They show taste.

If the voice never changes, never shows preference, and never gets concrete, it may be an AI baseline draft that wasn’t pushed into a real point of view.

Fast signals that suggest automated drafting

No single clue proves anything. Use clusters. If you see several signals in the same piece, it’s worth deeper checking.

Over-smooth structure with thin substance

AI often produces tidy sections that look complete: intro, steps, recap. The issue is that each section can be light on actionable detail. You get “steps” that don’t tell you what to click, what to measure, or what success looks like.

A quick test: pick one step and ask, “If I did only this step, what changes on my screen or in my results?” If the step can’t answer that, it may be padding.

Repetitive phrasing in different sentences

Models recycle patterns. You may see the same sentence shape again and again with swapped words. You may also see repeated lead-ins like “It’s a good idea to…” without a method attached.

Try this: highlight repeated phrases. If the same phrase shows up across unrelated paragraphs, the draft may be auto-built from patterns.

Generic examples that avoid real constraints

A human example often has friction: limited time, missing data, a strict rubric, a broken file, a picky editor, a budget cap. AI examples tend to float above the mess. They read like they were written for “anyone,” which means they fit no one.

Ask for one example tied to the writer’s real context: their class topic, their team’s tools, their project’s constraints. Vague examples are easy to fake. Real ones are harder.

Odd certainty about facts, quotes, and links

AI can mix details from different sources into one neat sentence. It can also invent citations that sound plausible. A common tell is a reference that doesn’t exist when you search it, or a quote that can’t be found on the page it claims to cite.

Pick one quote or one cited claim. Search the exact wording. If it’s not there, the draft may be careless at best, auto-generated at worst.

Clean grammar that clashes with the author’s past work

If you have prior writing from the same person, mismatch matters. A student who has always struggled with sentence control suddenly submits a flawless essay with mature pacing and zero typos. That doesn’t prove AI, yet it raises the odds.

Look for changes in: average sentence length, use of idioms, use of contractions, and how the writer handles tricky topics. Voice is hard to fake consistently.

Lists that feel padded

AI loves lists. It can produce ten bullets on demand, even when five would do. Look for bullets that repeat the same idea in new words, or that state obvious points without adding criteria, thresholds, or a test you can run.

A quick fix test: ask the writer to cut the list to the top three and justify why those three made the cut. Ownership shows up in the explanation.

Shallow handling of counterpoints

When a topic needs nuance, AI drafts often name the “other side” in one sentence, then move on. A person who has wrestled with the topic usually gives a sharper counterpoint, then explains why they still choose their stance.

Ask for one counterpoint that the writer personally finds strong, and ask what evidence would change their mind. If they can’t answer, the “balanced” section may be a template.

How to test text without accusing the writer

If stakes are real, your process matters. A blunt accusation can damage trust and can be wrong. Use neutral checks that any writer can pass, no matter what tools they used.

Ask for the working materials

Draft history can tell you a lot. Version history shows how the text grew. A single paste of 1,200 words is different from a draft built over time with visible revisions. Even with AI use, you may see prompts, edits, and rewrites that show effort and ownership.

If the writer claims they drafted it themselves, ask for: an outline, notes, a research list, or a rough draft. People who wrote the work usually have some trail, even a messy one.

Request a short oral walkthrough

Ask the writer to explain the piece in their own words, then zoom in: “Why did you pick this claim?” “What backs this number?” “What would you change if you had one more day?” A real author can usually answer quickly.

If the person stalls on every “why,” it can mean they didn’t own the ideas. That can happen with AI drafts, but it can also happen with rushed writing. Either way, you’ve found the quality gap.

Swap one constraint and ask for a rewrite

Pick one section and change the constraint: new audience, tighter word limit, different example, or a new data point. Then ask for a rewrite in 10–15 minutes. Someone who owns the ideas can adapt. Someone who relied on a paste may struggle to adjust without re-running a tool.

This test is fair because it doesn’t punish tool use. It tests understanding and authorship.

Signal cluster	What you may see	Quick check that adds confidence
Source fragility	Links that don’t match claims, vague “studies say,” missing titles	Verify 3 claims against the exact cited pages
Specificity gap	Advice fits any case, few proper nouns, no numbers with units	Ask for one concrete scenario with real constraints
Voice mismatch	Polished tone that doesn’t match earlier work	Compare 2 past samples for sentence length and phrasing
Pattern repetition	Same sentence shapes, repeated lead-ins, circular bullets	Highlight repeats; count how often they show up
Over-balanced stance	No crisp claim, no choice made, lots of safe hedging	Ask for the top recommendation in one sentence
Missing process	“Do research” without steps, no criteria, no decision rules	Ask what was checked, what was excluded, and why
Over-clean errors	Few typos but odd factual slips, mixed terminology	Check terms against a glossary or a standard reference
Instant completeness	Perfect headings and pacing, no rough spots at all	Request early notes that show idea growth

Detector tools and where they mislead

AI “detectors” look for statistical patterns in text. They can help as one input, yet they can also fail in both directions. A careful human writer can get flagged. A skilled AI user can slip through with edits, paraphrasing, or mixed drafting.

Even the companies building AI models have said text detection is hard at scale. OpenAI shut down its own AI text classifier after noting accuracy limits. You can read the notice on OpenAI’s AI text classifier page.

So treat detector scores like a smoke alarm in a kitchen. It can tell you to check the stove, but it can’t tell you what caused the smoke. Use detectors to decide when to verify facts and authorship, not to declare guilt.

What tends to raise false positives

• Non-native English writing with simple sentence patterns
• Highly structured writing like lab reports or legal notes
• Text that follows a strict template or rubric
• Short passages where patterns are noisy

What tends to raise false negatives

• AI text heavily edited by a person
• AI output stitched with human paragraphs
• Text rewritten through multiple tools
• Samples too small to score well

Provenance and draft history: the cleanest evidence when available

The most reliable way to know how something was made is provenance: attached data that records creation steps. For text, that’s not always available. Still, you often have something close: version history, tracked changes, exported metadata, or a workflow trail in an editor.

On the standards side, the NIST report on synthetic content risk (AI 100-4) describes approaches like provenance data tracking and watermarking. That framing is useful even when you’re dealing with plain text, because it pushes you toward auditable signals instead of vibes.

In day-to-day work, provenance is simple: show how the draft grew. If a writer can show a trail of thinking, revisions, and source checks, that beats a detector score every time.

Check	Where it works best	What to watch for
Version history	Docs with built-in revision logs	One big paste, then light edits
Tracked changes	Editing workflows with reviewers	Sudden full-section rewrites without notes
Source spot-check	Claims with data, dates, or definitions	Confident claims that don’t match the source text
Voice comparison	Known authors with prior samples	Shift in idioms, pacing, and precision
Constraint rewrite	School or work submissions	Writer can’t adapt without re-generating
Oral explanation	Any setting with live interaction	Writer can’t justify choices or sources

Workflow you can reuse in five steps

If you want a repeatable process, use this sequence. It’s fair, fast, and hard to game.

Step 1: Set the rule in plain language

Decide what is allowed in your setting. Is AI okay for outlining? Is it okay for grammar fixes? Must the writer disclose tool use? If the rule is fuzzy, disagreements start before you even look at the text.

Step 2: Run a three-by-three check

Mark three factual claims, three specific nouns (names, places, tools), and three numbers. Verify each one. If the piece is opinion-only, swap in three quotes, three references to class material, or three claims that can be traced to a reading list.

Step 3: Ask for ownership artifacts

Request an outline or notes and one paragraph explaining the main choice the writer made. Ownership shows up in trade-offs: what they cut, what they kept, and why.

Step 4: Ask for one constrained rewrite

Choose one section and tighten one constraint: fewer words, one fresh data point, or a new audience. Set a timer. If the writer understands the content, they can shift it without rebuilding the whole piece.

Step 5: Use a detector only for triage

If you still feel unsure, run a detector and treat the output as a cue to verify more, not as a verdict. Save inputs and outputs if you need an audit trail.

Common failure modes in AI-written text

Even when you don’t care about tool use, these patterns help you judge risk and credibility.

Blended facts that read clean but don’t hold up

Models can merge two related ideas into one sentence that sounds right. The result reads smoothly, yet it may be wrong in a subtle way. This is why source spot-checking works so well.

Definitions without boundaries

AI drafts often define terms in broad, safe language. Strong definitions draw boundaries: what counts, what doesn’t, and a short test you can apply. If a piece never draws boundaries, it may be shallow research or automated drafting.

Advice that ignores real constraints

When you know the context, you can feel the gap. The draft says “gather data” but doesn’t say which data exists, who owns it, or what to do when it’s missing. People with hands-on experience tend to name those constraints early.

How to handle the situation with care

If you’re a teacher, editor, or manager, your response sets the tone. Treat this like a quality issue first, not a character trial.

• Lead with standards. Point to the rule: originality, citation quality, or process proof.
• Ask for artifacts. Notes, drafts, sources, and version history reduce guesswork.
• Give a clear redo path. If the work fails the standard, say what a passing submission looks like.
• Separate tool use from copying. Copying a source is different from using AI to draft. Handle each with the right policy.

Where this matters most: school, hiring, and publishing

The same checks work in many settings with small tweaks.

School writing

Ask for process proof: outlines, reading notes, and a short explanation of how the writer chose sources. A student can still use AI, yet process proof shows whether they understood the material.

Hiring tasks

If you use take-home writing tasks, add a live follow-up. Ask the candidate to explain choices and revise one section on the spot. This tests real skill while staying fair to candidates who use tools for grammar or structure.

Publishing and client work

Set expectations in writing: voice match, source quality, and who owns fact-checking. If a draft arrives polished yet flimsy on sources, treat that as the issue. A clean paragraph isn’t proof of truth.

Limits you should accept up front

You can’t reliably prove AI use from text alone in every case. A person can write in a flat tone. A model can be edited into a human voice. So aim for a fair standard: factual accuracy, clear sourcing, and ownership of ideas.

If you apply the workflow in this article, you’ll catch most risky cases: pasted drafts with no ownership, weak sourcing, and empty generalities. You’ll also avoid the trap of trusting a single score and accusing someone who did the work.