How To Tell If Text Is Written By Ai | Red Flags List

If you’ve caught yourself wondering how to tell if text is written by ai, you’re seeing a shift. A lot of writing now arrives polished, fast, and oddly generic. The fix isn’t one “gotcha” trick. It’s a set of checks you can stack, then a short test or two that forces the writer to show real grounding.

This article is built for real use: grading, editing, hiring, and day-to-day review. You’ll learn what to spot on first read, what to test next, and what notes to keep when the call has consequences.

Fast Clues You Can Spot Before Running Any Tests

Clue	What You See	What To Do Next
High polish, low specificity	Clean grammar and tidy flow, yet few names, dates, numbers, or real constraints	Ask for one verifiable detail per paragraph
Even cadence	Sentence lengths and rhythm stay steady for long stretches	Request a rewrite in a sharply different style
Safe generalities	Broad nouns (“people,” “teams”) with no clear stance	Force a choice and a trade-off
Foggy sourcing	“Studies show” with no title, author, year, or link	Ask for full citations, then verify two
Confident but thin	Certainty where a careful writer would show limits	Ask what evidence would change the answer
Template blocks	Each paragraph follows the same shape: claim → tidy wrap	Check if any section breaks the pattern naturally
Shallow “lived” detail	Personal lines with no timing, tools, or step order	Ask for the exact steps and one mistake made
Glossy but wrong	Polished wording paired with a factual slip a subject reader catches	Spot-check core facts with primary sources
List padding	Bullets repeat the same point with new wording	Ask for one concrete case behind one bullet

Telling If Text Is Written By Ai With Quick Checks

Start with the cheap checks. You’re building a signal stack, not trying to “catch” someone. One clue can mislead. Three or four that point the same way are worth testing.

Check The Anchor Count

Anchors are verifiable pieces: a named policy, a paper title, a quote with a page number, a real date, a measured result, a tool version, a product model. A model-flavored draft can read smooth while dodging anchors.

• Underline anchors in each paragraph.
• If a paragraph has zero anchors, ask for one.
• If anchors appear, verify two, not ten.

Check For Rewording Loops

A common AI smell is repetition with new wrapping. A paragraph says the same thing three times: sentence, reword, wrap sentence. Humans do this too, yet it’s more common in “prompt to paste” writing.

Ask for a 20% cut with no lost meaning. Writers who know their point can tighten fast. Low-effort AI use often leaves the loop in place.

Check For Real Trade-Offs

AI drafts like balanced lists. Each option is “good and bad,” and nothing gets chosen. Ask for a decision: pick one approach, name one downside, then say why it’s still the pick. A real writer can commit and stay consistent.

Check The Source Trail

If the text claims authority, demand the trail: title, author, publisher, year, plus a link when it’s public. Then verify if the source says what the paragraph claims it says. A fake trail is a strong flag. A real source used loosely is still a problem for trust.

How To Tell If Text Is Written By Ai With Two Minute Tests

Once you’ve got a few flags, run two short tests. Skip detector scores at this stage. Use prompts that force grounding.

Test 1: Source-Backed Rewrite

Pick one paragraph. Ask for two citations that confirm two specific claims in that paragraph. Require a link and the exact line from the source (one sentence) that confirms each claim. If the writer can’t do it, the paragraph was built on patterns, not reading.

For publishing work, it also helps to know that AI text isn’t auto-banned in Search. Google’s Search Central post on AI-generated content and Search lays out the core idea: content gets judged by usefulness and trust signals.

Test 2: Show-Your-Work Steps

Pick one claim and ask for the steps that produced it. If it’s a “how-to,” ask for the order of steps and what fails at each step. If it’s a comparison, ask for the criteria and why each one was chosen. Humans tend to add friction points, like a dead end or a missing tool.

Test 3: Consistency Follow-Up

Ask a question that depends on an earlier sentence: “You said X in paragraph two. What’s the exception that flips your advice?” Low-effort AI use often drifts or contradicts the earlier text.

Test 4: Provenance Request

When it fits your workflow, ask for draft history or tracked changes. For media files, provenance standards can carry edit history through metadata. The C2PA technical specification explains how that kind of data can travel with a file.

Signals That Lean Toward Human Writing

No signal proves a person wrote the text. Still, some traits take effort to fake.

• Specific “why this order” notes. Not just steps, but why the order matters.
• Uneven texture. A tight line, then a longer one, then a punchy aside.
• Context cues that fit. A rubric line, a policy title, a dataset name, a tool version.
• Small imperfections. Minor quirks that don’t read like they were planted.
• A clear stance. A choice made with a known downside.

One warning: a careful human editor can smooth a draft until it reads “model-ish.” That’s why the tests above matter more than vibes.

When Ai Detectors Help And When They Mislead

Detector tools can be useful as one signal in a larger process. They can also misfire. Many detectors score patterns that models tend to produce. People can write in those patterns too, especially when they follow a template or write in a formal tone.

Use Detectors For Triage, Not A Verdict

Treat a score like a smoke alarm: it tells you where to check next. It doesn’t prove guilt. If you use a score as a verdict, you risk accusing honest writers and missing careful cheaters.

Check Multiple Chunks

If you do run a detector, test multiple sections from the same piece. Definition-heavy sections can score high. Story-heavy sections can score low. The pattern across the whole text matters more than one paragraph.

Ways To Reduce False Accusations

People write in many styles. A quiet writer can sound formal. A strong editor can smooth a draft until it reads “too clean.” If you treat style as proof, you’ll get burned. Aim for checks that hinge on verifiable work. Use the same prompts and judge outputs.

Ask For Process Notes, Not Confessions

Instead of “Did you use AI?”, ask, “How did you draft this?” Then ask for a short list of sources, a rough outline, and the edits made after the first draft. Most honest writers can share this in a minute. A pasted draft often comes with no trail.

Use A Baseline Sample When You Can

If you know the writer, compare to a small sample of their past writing from the same context. Look for consistent habits: sentence length swings, favorite words, and how they cite sources. Don’t treat difference as guilt. Treat it as a reason to run the two-minute tests.

• Keep the same rules for all writers.
• Give the writer a chance to revise with tracked changes.
• Write down what you verified, not what you “felt.”

A Practical Workflow For Teachers, Editors, And Hiring Teams

This workflow stays fair and easy to explain. It also leaves room for basic tools like spellcheck and grammar help.

Step 1: Set A Clear Rule

Write down what’s allowed. Some settings allow AI for outlines and ban it for final text. Some allow it with citation rules. Put it in the prompt, rubric, or policy.

Step 2: Read And Mark Anchors

On first pass, mark anchors: sources, dates, numbers, named methods, constraints. If the piece is all generalities, move to tests.

Step 3: Run Two Tests

Pick two tests from the earlier section. Keep them short. You’re checking grounding and consistency, not trying to trap anyone.

Step 4: Verify Two Claims

Choose two factual claims and verify them with primary sources. If the claims fall apart, the writing isn’t ready, no matter who typed it.

Step 5: Ask For A Revision With Trackable Changes

If you allow resubmits, ask for tracked edits or a brief change log. Writers who own the work can tell you what they changed and why.

Decision Checklist By Scenario

Scenario	Best Checks	What To Record
School essay	Anchor count, source-backed rewrite, show-your-work steps	Prompts used, two verified claims, revision notes
Application letter	Role-specific facts, one consistent story, trade-off commitment	Company details used, story consistency
Take-home task write-up	Criteria list, what failed first, final rationale	Steps, errors, tool versions
Blog draft	Source trail, fact checks, 20% cut test	Links checked, claims verified, edits logged
Customer email	Policy accuracy, correct case facts, tone match	Policy source, case notes
Product description	Spec accuracy, fit notes, warranty limits	Spec sheet source, verified numbers
Research summary	Full citations, claim-to-source matching, limit notes	Paper titles, section refs, mapping notes

Common Patterns In Ai-Written Text And A Clean Response

When a piece feels model-made, respond with a prompt that forces specificity. Don’t argue about “AI vibes.” Ask for work you can verify.

Pattern: Too Many Balanced Options

Each option gets the same weight. Ask for a single pick, one downside, and the reason it’s still the pick. Then ask the writer to link that choice back to the goal and constraint.

Pattern: Missing Local Context

The writing could fit any class, any company, any country. Ask for one local anchor: the rubric line, the policy title, the product model, the dataset name. If the writer can’t add it, treat the draft as generic.

Pattern: Polished Errors

A clean paragraph slips on one technical detail. Ask one narrow follow-up on that detail. If the answer drifts, you’ve learned that the writer may not understand the topic.

Quick Final Checklist

• Scan for anchors: names, dates, numbers, sources, constraints.
• Mark repetition: does the text reword itself in loops?
• Force a trade-off: ask for one pick and one downside.
• Run two tests: source-backed rewrite plus show-your-work steps.
• Verify two claims with primary sources.

If you still feel stuck on how to tell if text is written by ai, don’t chase perfection. Stack clues, test grounding, then decide based on what you can verify.