Test If Something Was Written By AI | Signals That Hold Up

You might be checking a student essay, a freelancer draft, a scholarship statement, or a blog post you’re about to publish. The goal isn’t a gotcha. The goal is a fair call you can explain: accept it, request revisions, ask for proof of work, or review it more closely.

Below is a practical workflow you can repeat. It blends human review, source checks, and careful use of detector tools.

Why AI Authorship Is Hard To Prove

Text is slippery evidence. A strong writer can sound polished. A rushed writer can sound flat. A model can be coached to sound personal. A person can copy a template and sound generic.

Many tools don’t create from scratch. They rewrite, expand, or smooth existing text. That blurs “generated” and “edited,” so a detector can flag honest work that used AI for cleanup.

OpenAI has noted that classifier-style tools can miss AI text and can flag human text, especially when writing is plain or predictable. Treat scores as hints, not verdicts, and read their limits in the OpenAI AI classifier announcement.

Fast Screen: A 2-Minute Read-Through Checklist

Start with a quick pass. You’re looking for patterns that show up before you run any software.

Language Patterns That Raise Suspicion

• Even cadence. Sentence lengths stay similar, with few rough edges.
• Safe phrasing. Lots of mild claims, few concrete details.
• Over-clean structure. Each paragraph follows the same mold.
• Thin specificity. Names, dates, numbers, and citations are scarce or vague.

Content Patterns That Often Show Up In AI Drafts

• Untraceable claims. “Studies show” with no working source.
• Instant breadth. Many subtopics, little depth in any one.
• Strange confidence. The voice sounds sure, yet avoids checkable details.

None of these proves authorship. They tell you whether a deeper check is worth your time.

Step-By-Step: A Repeatable Way To Check A Text

Use this workflow when the stakes are real: grades, hiring, publication, or policy decisions.

Step 1: Set The Rule Before You Judge

Define what counts as acceptable work for this task. Is AI banned, allowed with disclosure, or allowed for grammar only? Write your rule in one sentence and apply it consistently.

Step 2: Ask For Proof Of Process

Don’t ask for a confession. Ask for artifacts that honest writers can share quickly.

• Outline or bullet plan made before drafting
• Notes or reading list used while writing
• Earlier drafts showing real revision
• Version history from the writing app, when available

Step 3: Verify Facts And Citations

Pick 5–10 factual claims and verify them. Check names, dates, definitions, and numbers. If citations exist, open them and confirm the source matches the claim and the wording.

Step 4: Test For Ownership With Targeted Questions

Choose one paragraph and ask two narrow follow-ups that require the writer to extend that point. A real author can usually expand it in their own voice. A copy-paste user often restates the paragraph without adding substance.

Step 5: Look For Seams

AI-assisted writing can contain abrupt shifts in tone, reading level, or terminology. Read only the first sentence of each paragraph and see if the line of thought stays coherent or loops back in new wording.

Step 6: Use Detectors Like A Smoke Alarm

Detectors can be useful when you treat them as a pointer, not a judge. Use them to decide where to re-read closely.

• Test multiple chunks, not just the full document.
• Compare results across two tools.
• Save screenshots or exports with timestamps.
• Never claim “proof” from one score.

If you want a broader technical view that includes provenance and watermarking, NIST lays out approaches for synthetic content in Reducing Risks Posed by Synthetic Content (NIST AI 100-4).

What Common “AI Tells” Get Wrong

Some popular tips sound convincing, yet they trigger false positives.

“Too Polished Means AI”

Strong writers exist. Editors exist. Many people draft carefully and proofread. Clean grammar is not a reliable signal.

“Repetition Means AI”

Repetition shows up in weak human writing too, especially in school assignments that follow a rigid prompt. Treat repetition as a quality issue first.

“A Detector Said 98%”

A percentage is not a chain of evidence. If you can’t point to specific passages and explain what’s wrong with them, you don’t have a solid case.

Signals That Hold Up Better In Real Reviews

These checks tie the writing to verifiable work, not just style.

Verifiable Sources And Accurate Quotes

Ask for the source list used during drafting, then spot-check two items in full. Real citations lead to real pages that match the claim. Fake citations often look plausible, yet the title, author, or outlet won’t line up.

Draft Trail With Real Revision

Human drafts tend to show messy growth: reordered sections, rewritten claims, and notes in the margins. AI copy-paste often shows one clean drop-in with light edits.

Clear Method For Any Numbers Used

If the text includes figures, ask how they were calculated. A real author can usually show the spreadsheet, notes, or steps. AI-generated numbers often lack a traceable path.

Citation Formatting Oddities

Watch for references that look neat yet don’t behave like real sources. Titles may be slightly off, author names may be missing, and publication dates may be inconsistent across the paper. If a reference list includes journals or reports, open two items and check that the outlet and page exist and that the text matches what’s quoted.

Over-Perfect Neutrality

Many AI drafts avoid taking a stance. They list pros and cons, then stop. If the task asks for an argument, a recommendation, or a decision, ask the writer to pick a position and defend it with sources and specific reasoning tied to the prompt.

Table 1: Evidence Checklist You Can Use In Any Setting

Use this table after your first pass. It helps you collect signals without leaning on vibes.

What You Check	What You Look For	What It Suggests
Source links	Working URLs that match the claim	Real research trail, or a broken one
Named details	Specific dates, names, places, numbers	Concrete work, or vague filler
Internal consistency	Terms stay stable across sections	Single voice, or stitched text
Prompt fit	Direct answers to the rubric or question	Task ownership, or generic output
Revision trail	Earlier drafts with real edits	Human drafting pattern, or one-shot paste
Detector agreement	Similar flags across tools and chunks	Worth re-reading closely, not a verdict
Follow-up answers	Clear expansions with sources or details	Author ownership, or evasive replies
Quote accuracy	Quoted lines match the source wording	Careful reading, or sloppy fabrication

Ways To Check If A Text Was Written By AI With Less Drama

Your process matters as much as your result. A blunt accusation can do damage. A calm review process keeps the situation fair.

Use A Two-Layer Review

Layer one is quality: does the work meet the task? Layer two is authenticity: does the writer show ownership of the work? This keeps you from treating “AI suspicion” as the whole story.

Ask For A Short “How I Wrote This” Note

One paragraph is enough. Ask for tools used, sources used, and one revision choice they made. Honest writers can answer fast.

Use Live Writing When Stakes Are High

For hiring or formal disputes, a short supervised writing sample can settle things quickly. Use a prompt that matches the original task and a time box that feels fair.

Match The Remedy To The Evidence

If the text is generic but not dishonest, request more specific detail and better sources. If the text contains fake citations or invented facts, treat that as a separate issue from AI use.

Table 2: What To Do After You Spot A Red Flag

This table helps you pick a next step that fits what you actually found.

Red Flag	Next Step	What You Save
Vague claims with no sources	Request citations for 5 factual claims	List of claims and returned sources
Citations that don’t match	Ask for source PDFs or page numbers	Screenshots of the mismatch
Tone shifts mid-document	Ask for earlier drafts and edits	Draft timeline or change history
Detector flags one section	Review that section line-by-line	Marked passages and notes
Writer can’t explain choices	Ask 3 targeted follow-up questions	Q&A transcript
High stakes decision pending	Use a supervised writing sample	Prompt, time, and sample text

How To Record Your Decision So It Holds Up

Keep a clean record. You don’t need a giant file. You need clarity.

• Save the version you reviewed, with date and file name.
• Note the rule you applied (AI banned, AI allowed with disclosure, and so on).
• List the passages you flagged, with short reasons tied to evidence.
• Record what you asked the writer to provide and what they returned.
• Store detector outputs as screenshots, not just a number in an email.

Limits And Fair Warnings Before You Decide

AI checking is messy. Models change fast. Detectors change fast. So treat your decision as best judgment from evidence, not as a lab result.

When the evidence is thin, choose low-harm actions: request better sources, ask for a rewrite with clearer details, or use a short writing sample. For publishing, the best filter is editorial: require sources, require named details, and require drafts for paid work.

Be careful with non-native English writers and writers with disabilities. Their text can trip detectors and can read “formulaic” without any AI use. Lean on sources, drafts, and follow-up answers instead of style alone.