> For the complete documentation index, see [llms.txt](https://evidano.gitbook.io/ailyze-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://evidano.gitbook.io/ailyze-docs/analyze-data/detect-ai-written-text.md).

# Detect AI-Written Text

Estimate how likely a text is AI-generated, see which passages drive the score, scan a folder of documents at once, and read the result responsibly.

This page shows you how to use **AI Detection** to estimate how likely a piece of writing is to be AI-generated, passage by passage, and how to check many documents at once. It is for teachers, editors, and researchers who need a first signal about a text's origin and want to read that signal with the right caution. Nothing you paste or upload leaves your computer: the detector downloads to your browser once and runs there.

## Before you start

* Any modern browser. **AI Detection** on your account page and the public **Free AI Detector** page are the same tool, free to use, and need no account.
* English text. The detector was trained on English only; on other languages its score means nothing, and the page warns you when it sees one.
* Enough text. Very short passages are withheld from scoring; the floor and the per-scan limits are in [Plans, Limits & Credits](/ailyze-docs/reference/plans-limits-and-credits.md).

## Step 1: Load a text

From your account page, click **AI Detection**. Paste or type into the editor, or click **Upload file** to read a text, Markdown, PDF, Word, CSV, or Excel file in the browser. On your first visit the **Analysis** pane shows **One-time setup** while the detector downloads; it is cached afterwards and works offline. You can type while it loads.

Three sample texts under the editor, **Human-written sample**, **AI-generated sample**, and **Journalism sample**, show what each kind of result looks like before you paste your own.

<figure><img src="https://1178311500-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fi4GBXPMUDxszIpUVOARU%2Fuploads%2Fgit-blob-a9d5b38adbea30600054336e852f64bf5b38a2b3%2Fanalyze-detect-ai-written-text-01-editor.png?alt=media" alt="AI Detection page with the human-written sample in the editor and the Analysis pane beside it"><figcaption><p>The editor on the left, the Analysis pane on the right. The setup bar disappears once the detector is cached.</p></figcaption></figure>

## Step 2: Scan

Click **Scan**, or press Ctrl + Enter. With **Auto-scan** on, the text is rescanned shortly after you stop typing, and only the passages you changed are rerun. Click the **Auto-scan** control to switch to a timer, change the delay, or turn it off. Long documents pause auto-scan; **Scan it anyway** runs them on demand.

## Step 3: Read the result

The **Overview** tab shows:

* A gauge with the **AI-generated content probability**, which is the word-weighted average of every passage's score, and a verdict that runs from **Very likely human-written** through **Inconclusive** to **Very likely AI-generated**. Below the word floor it reads **Not enough text for a reliable result**.
* A confidence pill from **Low confidence** to **High confidence**. Hover it to see why: how long the sample is, whether the passages agree with each other, and how far the score sits from the middle.
* A composition bar that splits the words into **Human**, **Mixed**, and **AI** shares, and the **Most AI-like passage**.
* **Across the document**, a strip of coloured bands, one per sentence group. Click a band to jump to that text in the editor, where the same colours highlight the passages.
* **Score distribution**. Mass at both ends with a dip between them suggests mixed authorship, so read the highlighted passages rather than the single percentage.

The **Breakdown** tab lists **Passages** and **Sentence Groups** with their scores, most AI-like first. The **Style** tab shows writing-style statistics measured from your text alone, such as sentence-length variation and stock AI phrasing; they are feedback on the prose and not part of the score. The **History** tab compares scans from this session and keeps recent scans in this browser only.

## Step 4: Export or print

Click **Export** in the Analysis pane for **Copy summary**, **Download annotated report (HTML)**, **Download raw scores (JSON)**, **Download summary (text)**, or **Print or save as PDF**. Every export carries the accuracy warning with it.

## Step 5: Scan many documents at once

Click **Bulk detector** under the editor, or under **AI Detection** on your account page, to open the **Bulk AI Detector**. Drop a folder of documents or click **Choose files**, then click **Scan documents**. Each file gets a quick scan: an overall score, a verdict, and a confidence rating, listed most AI-like first in a table with the columns **Document**, **Words**, **AI**, and **Verdict**. Files that could not be read, or that were too short, stay in the list with the reason. Open a row to run the full scan on that one document, with the usual highlighting and tabs.

<figure><img src="https://1178311500-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2Fi4GBXPMUDxszIpUVOARU%2Fuploads%2Fgit-blob-8f7cef7ddacdc9540fbee75efa01191931777043%2Fanalyze-detect-ai-written-text-02-bulk.png?alt=media" alt="Bulk AI Detector page with the drop zone, Choose files, and Scan documents buttons"><figcaption><p>Nothing is uploaded. Every file is read and scored in the browser tab, and the results are discarded when you close it.</p></figcaption></figure>

**Export** offers **Download summary (CSV)**, **Download summary (Excel)**, and **Annotated reports**, a ZIP of HTML reports for the documents you opened. Batch results are not saved, so export before you close the tab.

## Step 6: Adjust the thresholds

The gear icon in the Analysis pane opens **Detector settings**: where the **Human** and **AI** bands start (**Count as human below** and **Count as AI above**), the word floor for flagging short passages, the **Highlighting** style (**Heat map**, **Three bands**, or **Off**), and **Detector precision**, which chooses between the standard download and a larger, more accurate build for machines with GPU acceleration.

## How to read the score responsibly

* The score is a probability from a classifier, not proof. It is wrong in both directions.
* False positives are markedly higher for writers whose first language is not English, and for technical, templated, legal, or heavily edited prose.
* AI-drafted text that was rewritten in a personal voice can score as human. A low score never rules out AI.
* Translations, quotations, references, and code skew the result.
* Never use the score as the sole basis for an academic-misconduct, hiring, or disciplinary decision. Treat a high score as a reason to look at drafts and version history and to talk to the writer.

The **Model limitations** link under the Analysis pane repeats these points inside the tool.

## Next steps

* Decide where the tool fits among the ten workflows: [Choose Your Workflow](/ailyze-docs/getting-started/choose-your-workflow.md).
* Check the accepted file types and the per-scan limits: [Supported File Types](/ailyze-docs/reference/supported-file-types.md).
* Writing up your own findings so they read as yours: [Reporting Findings](/ailyze-docs/methods-and-best-practices/reporting-findings.md).