How to Test Text for AI in Under Five Minutes

In June 2026, Grammarly's parent company Superhuman bought GPTZero, an AI detector with 19 million registered users and roughly $30 million in annual recurring revenue. Checking text for AI has become checkout-lane software, something millions of people now do between meetings. The technique, for most of them, hasn't caught up with the habit.

Done right, the whole job takes five minutes. To test text for AI properly: use a sample of at least 300 words, run it through a detector that reports confidence levels, read the sentence heatmap rather than just the headline percentage, corroborate with a second engine or a human read, and write down what you found.

What a five-minute test settles, and what it can't

A quick test tells you how a text reads statistically: whether its word choices and rhythms look more like model output or human drafting. That's genuinely useful for triage. It is not authorship proof, and the ceiling is worth knowing before you start. In a 2025 ACL study, expert readers who use ChatGPT daily beat automated tools, catching AI text with 99.3% accuracy even against evasion tricks. Software is faster and cheaper; it is not infallible, and our records of detectors getting it wrong should calibrate anyone's expectations.

It also can't tell you why. A high score doesn't distinguish "pasted from a chatbot" from "drafted by a person who writes in smooth, formal patterns," and it says nothing about intent or permission. Those answers live in drafts, version history and conversations, not in any percentage.

So run the test to decide where to spend attention. Don't run it to decide someone's fate in five minutes.

How to test text for AI, step by step

1. Collect a sample of 300 words or more

Detection statistics stabilize with length. Below roughly 300 words, scores swing wildly between runs and tools, which is why short samples cause a disproportionate share of false alarms. Take the whole document when you can. If you only care about one suspicious section, take that section plus its surroundings, not a lone paragraph.

Resist the urge to stitch together several short pieces to hit the threshold. Gluing five emails into one "document" mixes contexts and genres, and the resulting score describes the collage, not any of the parts. Scan short items separately and look for a consistent pattern across them instead.

2. Run it through a detector that reports confidence

Any scanner can emit a percentage; you want one that also tells you how much to trust it. Our free AI detector takes up to 5,000 characters with no sign-up, runs multiple engines on the same text, and labels every result with an explicit confidence level. Working from a PDF or DOCX? Upload the file directly rather than copy-pasting, which mangles formatting and sometimes drops whole sections; on our platform one credit covers 100 words, and a free account includes 300 credits a month.

Whatever tool you choose, the feature to insist on is the same: visible uncertainty. If you're still fuzzy on what these tools even measure, our plain-English explainer on AI checkers is the two-minute prerequisite.

3. Read the confidence level before the percentage

Two scans can both say "72% likely AI" and mean different things. On a 2,000-word essay at high confidence, that's a strong signal. On a 250-word cover letter at low confidence, it's barely a lean. Reading confidence first is the single habit that separates people who use detectors well from people who get burned by them.

4. Check the sentence heatmap for patterns

The heatmap is where verdicts become explainable. Flags clustered on formulaic passages, intros, summaries, boilerplate, suggest style, not synthesis; that pattern is exactly how human writing gets falsely flagged. A hard block of flagged paragraphs mid-document, sounding nothing like the rest, tells a different story. Even Turnitin, with its heavily marketed document-level numbers, discloses roughly 4% false positives at the sentence level, so individual red sentences prove little; patterns are the evidence.

If your tool offers no sentence-level view at all, that's informative too: you're holding a scanner built for verdicts rather than evidence, and it's worth switching before a result actually matters.

5. Corroborate with a second method

The RAID benchmark showed commercial detectors degrading in different ways under different attacks, which is an argument for never trusting one engine's opinion. Cross-check with a second tool, or better, with ground truth: earlier writing you know is the author's, compared side by side. Where tools agree with each other and with the writer's known voice, confidence is earned. Where they split, the honest reading is "uncertain," not whichever number suits you.

6. Write down the result and its limits

Tool, date, word count, score, confidence, one sentence of interpretation. Save the PDF report if the scanner offers one. Five minutes of documentation converts a fleeting impression into evidence you can defend months later, which matters because a surprising number of these tests eventually surface in disputes.

Reading the outcome

Two rows earn a comment. "High score, low confidence" is the one that burns people, because the big number is what sticks in memory and the small caveat is what mattered. And "tools disagree sharply" feels like a malfunction but isn't: detectors train on different data and draw their lines in different places, so edited and hybrid text genuinely lands on different sides of different thresholds. Disagreement is the tools telling you the truth about their limits.

Mistakes that quietly ruin the test

Four habits account for most bad calls. Testing an 80-word fragment and believing the answer. Chasing a specific number, when 0% and 100% are both unattainable certainties; paraphrased and "humanized" text pushes scores down without changing authorship, a problem we dissect in the truth about undetectable AI text. Convicting on a single tool's output. And treating the test as the investigation instead of its first step.

A fifth habit deserves its own sentence: testing text you've already made up your mind about. Confirmation bias turns a 55% score into "proof" when you expected AI and into "noise" when you didn't. Decide in advance what result would change your view, then run the scan.

Avoid those, follow the six steps, and you'll extract everything a five-minute test honestly offers, which is more than most people get from an hour of arguing about a percentage. The full reasoning behind how we score, calibrate confidence and flag short samples is public on our methodology page.