# AI Detector: Check If Text Was Written by AI

Source: https://iachecker.com/ai-detector
Site: IA Checker — https://iachecker.com

## Summary

Check text, a URL or a document for AI writing. Sentence-level signals, the full verdict scale, our published false-positive record and what a detection score does not prove.

## What an AI detection score is, and what it is not

This page explains the evidence behind a result: the signals the engine measures, the full verdict scale, how the current model performs on our published corpus, and the situations where a score should not be trusted on its own.

## The signals behind a score

Every report shows the same three measured signals, each scored from 0 for human-leaning to 100 for AI-leaning. They are reported separately on purpose. A single high signal is a weaker case than three that agree, and you can see which one is carrying the verdict.

- **Rhythm uniformity** (Sentence cadence variance): Human drafts vary sentence length in bursts: a long clause, then a short one, then a fragment. Generated prose tends to settle into a narrow band. A high score means the cadence is unusually even, which is common in AI output and also in heavily copy-edited text.
- **Punctuation flatness** (Punctuation range): The engine looks at how many punctuation devices the writing actually uses: semicolons, parentheses, dashes, question marks. A narrow range scores high. Note that formal registers legitimately flatten punctuation, which is why this signal rarely decides a verdict alone.
- **Formulaic phrasing** (Patterns & structure): This aggregates connector vocabulary, document structure and contextual grounding: whether the text names specifics or stays at a level of generality that any model could produce. It is usually the most informative of the three, and the hardest to fake in either direction.

## Reading the verdict

The result is not a percentage with a pass mark. It lands on one of nine tiers, and three of them sit deliberately in the middle because that is where a lot of real writing belongs.

- **Human-written**: Signals agree, strongly, on human authorship.
- **Almost certainly human**: Signals lean human with one minor counter-signal.
- **Most likely human**: Human-leaning, but the sample is not decisive.
- **Humanized AI**: Patterns consistent with generated text that has been rewritten to read more naturally.
- **Part human, part AI**: Segments diverge: some passages read generated, others do not.
- **Too close to call**: The measured signals are split. The honest answer is that we do not know.
- **Most likely AI**: AI-leaning, but short of a confident call.
- **Almost certainly AI**: Several independent patterns are compatible with generated writing.
- **AI-generated**: Signals agree, strongly, on generated authorship.

The three middle tiers are the point of the scale. A detector that only answers human or AI has to force every ambiguous sample into one of them, and ambiguous samples are the ones where a wrong answer costs somebody their grade or their job application.

## Our published detection record

Most detectors market an accuracy figure and never publish the false-positive rate, which is the number that actually matters when a person is being accused. Here is our latest run in full, on a 20-text regression corpus scored by Advanced 3.2. It is sized to catch a break between model versions rather than to settle a global accuracy claim, and we would rather return an uncertain verdict than a false accusation.

- Human texts that stayed classified human: 10 of 10
- AI texts classified AI: 5 of 10
- AI texts that landed in a hybrid tier: 5 of 10
- Human and AI inversions: 0

## Checked against more than one engine

One engine is one opinion. A detector that only ever agrees with itself gives you no way to tell a confident result from a lucky one, so a report can put our verdict side by side with other detectors on the same text and show you where they disagree.

- **Originality.ai** (Integrated): Runs on the same text as your IA Checker result, so a disagreement between the two is visible instead of hidden.
- **ZeroGPT** (Integrated): A widely used free detector. Seeing it agree, or not, is often the fastest way to judge how much weight a score deserves.
- **Pangram 4** (Integrated): A specialist detection engine available as a third opinion inside the suite. It is not counted in the corpus run above, which was scored by Advanced 3.2 alone.

## What you can run through it

Paste text directly, give the checker a URL to scan, or upload a file up to 20 MB. PDF, DOCX, TXT, Markdown and CSV are read natively. Images and scanned pages go through OCR first, so a photographed page or a screenshot of a document can be checked as well.

Longer samples produce better evidence. Under roughly a paragraph the engine flags the result as uncalibrated, because style signals on a few sentences are close to noise. If you only have a short extract, treat the output as a prompt to look closer rather than as a finding.

## What a detection score does not prove

No AI detector, including this one, can prove who wrote a text. It measures statistical properties of the writing. Those properties correlate with generated text, and they also correlate with several kinds of entirely human writing, which is where the harm happens.

The best documented failure is bias against people writing in a second language. Non-native English writing tends toward more regular cadence and a narrower vocabulary, which is exactly what a style-based detector reads as machine-like. Published studies have repeatedly found elevated false-positive rates on this group across commercial detectors.

Three other patterns produce false positives reliably: technical and legal registers, where formulaic structure is the convention rather than a tell; text that has been through heavy editing or a grammar tool, which smooths the same variance a detector looks for; and short samples, which do not carry enough signal either way.

The inverse is just as true. A generated draft that a person has genuinely reworked will often read as human, because at that point much of the writing is human. That is not a hole to be patched. It is what the measurement means.

## Which model runs your check

Advanced 3.2 is the model behind free checks, and it is the one the benchmark above was run against. Ultra 3.2 is available on paid plans and applies the same signal vocabulary with a wider feature set. Pangram, a specialist third-party detection engine, is integrated as a second opinion inside the suite.

The signal names and the verdict scale stay the same across models on purpose, so a result you read last month is still readable today and a report can be compared across plans.

## Using a result responsibly

If you are assessing someone else's work, the score is where the conversation starts, not where it ends. Open the sentence-level view and look at which passages carry the signal. Ask about drafts, sources and process. A person who wrote the text can talk about it; that has always been the strongest evidence available, and it still is.

Never use a score on its own as grounds for a grade penalty, a rejection or a disciplinary step. An institution that does this will eventually act against someone who did nothing wrong, and the detector output will not survive scrutiny when they appeal.

If your own writing was flagged and you did write it, the sentence-level breakdown is what to bring: it shows which passages triggered which signal, which is far more useful in a conversation than the headline number. Keep your drafts and version history where you can.

## FAQ

### How accurate is this AI detector?

On our latest published run, 10 of 10 human texts stayed classified human and 5 of 10 AI texts were classified AI, with the other 5 held in a hybrid tier and no inversions. That corpus is 20 texts, sized as a regression check rather than a global accuracy measure, and we publish it with that limit stated. Treat any detector advertising a single high accuracy percentage without a published corpus with caution.

### Can an AI detector prove someone used ChatGPT?

No. A detector measures statistical properties of writing, not authorship. It can tell you that a text has patterns common in generated writing. It cannot establish who produced it, and it should never be the sole basis for an accusation.

### Why was my own writing flagged as AI?

The most common causes are writing in a second language, a formal or technical register where formulaic structure is the convention, heavy editing or grammar-tool passes that smooth out sentence variance, and samples too short to carry signal. Open the sentence-level view to see which passages triggered which signal.

### Are AI detectors biased against non-native English writers?

Published research has repeatedly found elevated false-positive rates on non-native English writing across commercial detectors. The reason is structural: second-language writing tends toward regular cadence and a narrower vocabulary, which is what style-based detection reads as machine-like. This is a known limit of the method, and it is why a score must not stand alone in an academic integrity procedure.

### Does rewriting AI text beat a detector?

Substantially reworking a generated draft will usually move the result toward human, because at that point much of the writing is human. Light paraphrasing often lands in the Humanized AI tier instead, which is a distinct verdict rather than a pass.

### Is this AI detector free?

Yes. Checks on the Advanced 3.2 model are free and need no account to start. Ultra and the wider suite, including humanization and the document tools, are on paid plans listed on the pricing page.

### Can I check a PDF or a Word document for AI content?

Yes. Upload a PDF, DOCX, TXT, Markdown or CSV file up to 20 MB and it is read before the check runs. Scanned pages and images are put through OCR first, so a photographed document works too.

### Is my text stored or used for training?

Submitted text is not used to train models. The privacy page sets out what is retained, for how long and why.

## Related

- AI detector for teachers (https://iachecker.com/ai-detector/for-teachers): Exams and essays: how to use a score without treating it as proof.
- AI in CVs (https://iachecker.com/ai-detector/cv): Why résumés false-positive, and what to read instead of a cutoff.
- Extract text from PDF (https://iachecker.com/tools/extract-text-from-pdf): Turn a PDF into text before you run a check.
- AI Humanizer (https://iachecker.com/ai-humanizer): Rewrite a mechanical draft so it reads naturally, without losing meaning or facts.
- IA Checker vs GPTZero (https://iachecker.com/compare/ia-checker-vs-gptzero): Sourced, dated comparison of the two detectors, including where GPTZero wins.
- Pricing (https://iachecker.com/pricing): What is free, what Ultra adds, and which plans include the rest of the suite.

## Limits

Detection results are indicators, not proof of authorship. False positives
are possible, particularly on short, edited or heavily formatted text. A
person should stay in the decision loop.

## Elsewhere on this site

- Pricing: https://iachecker.com/pricing
- Full index for assistants: https://iachecker.com/llms.txt