AI Checker
for ML Engineers

Take a sample from the corpus and inspect it before changing the dataset. AI scores and sentence highlights can help you decide where to spend manual review time, look into provenance and record the reason for a decision.

For ML engineers

AI text detection for corpus and dataset review

One collection may mix authored documents, recurring templates, generated passages and edits whose origin is unclear. AiChecker gives you a text-level signal to inspect while building a review protocol. Explore chosen samples in the browser, then test a proposed filtering rule on data that represents the collection you're actually working with.

Inspect suspected synthetic passages

Keep the sample complete and record its source and original boundaries. Read the highlighted words with the known provenance beside them. A suspicion of generation and evidence of how the text entered the collection are different things to record.

Review boilerplate and ingestion effects

Go back to the source document. Did extraction remove a heading, join paragraphs or repeat the warranty text? Compare that page with what you scanned. Template repetition and parsing errors each need handling in the pipeline.

Build a manual review sample

Draw samples across sources, languages, lengths and score ranges, including low scores. Keep disagreements visible and allow an uncertain label. Looking at high scores alone cannot tell you what the filter misses.

A worked example

See how to use an AI score in a review

A score tells you where to look more closely. This example shows how to turn a highlighted passage into a concrete question or revision.

Illustrative report · fixed sample scores

Support corpus: extracted warranty passage

79%Overall AI score

Look at the repeated wording and the original source before giving this sample a provenance label or excluding it.

Text with sentence highlights

Select a sentence to view its review note. The percentage is an illustrative sentence score.

Lower signalMixed signalStronger signal

Example review decision

Keep provenance uncertain for now. Record the boilerplate flag separately and inspect the source. Do not treat the sample's illustrative percentage as a ground-truth label.

Demonstration only. Scores and highlights are illustrative, not live scan results. Review notes are examples, not automated findings. The AI score is not a percentage of words written by AI or proof of authorship.

From text check to a useful next step

  1. Define and record the input

    Record the source ID, document boundaries, preprocessing steps, scan date and whatever provenance you have. Leave out data your handling policy does not allow you to submit.

  2. Inspect a stratified sample

    Look through low-scoring, borderline and high-scoring examples. Before assigning labels, compare the highlights with the source and look for repeated boilerplate.

  3. Evaluate the proposed rule

    Measure errors on an independently labelled holdout covering the languages and document types you intend to filter. Record exclusions and uncertainty first, before carrying a threshold over to the larger collection.

Find samples worth investigating.Keep labels grounded in provenance.

Your review checklist

  • Source IDs and preprocessing choices are recorded.
  • The sample includes the collection's relevant languages and document types.
  • Boilerplate and duplicates receive a separate review.
  • Provenance or a stated annotation protocol supports the review labels.
  • The holdout evaluation counts false positives and synthetic text missed by the filter.
  • Uncertain samples are still distinguishable from confirmed labels.

Questions about AI checks for ML engineers

Can an AI checker help review a text dataset?

It can help you choose examples to read. Keep the source context with each sample, and compare the result with independently established provenance before deciding on labels or exclusions.

Can I use the AI score as a ground truth label?

No. A documented origin or a stated annotation protocol is the basis for a label. Leave an uncertain category available when the evidence falls short.

How should I choose a filtering threshold?

Try candidate thresholds on an independently labelled holdout that matches your collection. Weigh the cost of a human text wrongly removed and a synthetic text missed, rather than borrowing a universal cutoff.

Do benchmark results transfer to my corpus?

Not automatically. The language, domain, document length and preprocessing may differ. Test on representative data before making the score part of a filtering rule.

Does this page provide a corpus scanning API?

This is a browser text checker for manual inspection. It does not document or promise an API, an automated ingestion pipeline or a test for training-data contamination.

Can I explore a sample for free?

There's a daily free allowance in the checker. Stay inside the current input limit, and submit only material your data-handling policy allows.

Start with the text you are reviewing

Use the checker above to inspect your text. Your current free scan allowance is shown with the tool.

Check a corpus sample

Go beyond the limits

300,000 scan words per month
Batch scanning
Unlimited amount of deep scans
Document upload
Credit / Debit Card
Secure Card CheckoutPowered by stripe

Your subscription will be billed at $5 per month starting on ….
You can cancel anytime. If not canceled, your next billing date will be ….