Pangram Step by Step: How to Check if AI Wrote a Text

👁 119 reads

Pangram is an AI text detector that, instead of showing one percentage for the entire document, reveals which fragments look human-written, which appear AI-assisted, and which seem fully generated. The check itself takes a few seconds. What matters much more is how you read the result so you don’t wrongly accuse someone who wrote the text themselves.

These steps are based on Pangram’s official knowledge base and the model description on the producer’s site.

Before You Start: Set Your Rules

Before you check someone else’s first text, decide what you’ll do with the result. In a school or newsroom, it’s good if the rules are known in advance: the detector serves to point out fragments for discussion, not for automatic grading. The people whose texts you’re checking should know this is the case. It’s fairer and limits conflicts when the result turns out to be a false alarm.

It’s also worth checking a few of your own, self-written texts. You’ll see what the result looks like for a certainly human text in your style and language.

Step 1. Your First Text to Check

You log in to Pangram and open the panel at pangram.com/dashboard. Then:

• paste or type the text into the text field,
• click the Check for AI button,
• wait a few seconds for the analysis.

The text must have at least 50 words. Shorter fragments don’t give the model enough signal and the result would be unreliable. For short texts, Pangram shows a warning about limited confidence anyway. You can upload longer works as a file, for example PDF or DOCX.

Step 2. How Pangram Reads Text

This is the key to understanding the result. According to the producer’s description, the model divides the document into fragments and evaluates each separately. That’s why the result can say that the first half of a paper is human and the second half is AI-assisted. This is much more useful than one number for the whole, because it shows where it’s worth taking a closer look.

Pangram distinguishes fully human texts, lightly AI-assisted, moderately assisted, and fully generated. The “assisted” category refers to fragments that someone wrote and edited together with AI, so that you can’t attribute the words to a single author.

Step 3. Reading the Result

After analysis, you see how much of the text looks AI-written, and in the tabs on the right side the details: successive fragments with labels, a chart, and space for notes. Go through the text fragment by fragment, don’t just look at the number at the top.

The producer emphasizes that there is no universal “magic number” above which you must react. Low values often mean ordinary writing tools, like a spell-checker or suggestions in an editor.

Step 4. Detecting AI Assistance

AI assistance detection can be turned on and off. In the free version it’s off by default, which changes the result: lightly assisted text then shows as human, and moderately assisted as generated. If you’re comparing results with a colleague, make sure you have the same setting.

Step 5. History and Organization

Every check stays in the history. Pangram also lets you tag and group scans, which saves a lot of searching when checking papers from a whole class or texts from several authors. There’s also a Chrome extension that lets you check text without copying it to the panel.

The Most Important Rule: Results Are Clues, Not Proof

The producer reports very high accuracy and a low false-alarm rate, but those are results from their own tests. Independent reviews show that re-edited AI text can pass as human, and every detector sometimes makes mistakes. That’s why a Pangram result should open a conversation, not end it.

When evaluating authorship, it’s worth reaching for other sources:

• document version history, if the work was created in Google Docs or Word,
• earlier texts by the same person for style comparison,
• conversation with the author about the content, because whoever wrote the text themselves can discuss it.

Common Mistakes When Using the Detector

• Checking texts that are too short, under 50 words.
• Judging by one number without reviewing the fragments.
• Comparing results with different settings for assistance detection.
• Treating the result as a verdict without talking to the author.

How to Evaluate a Result in a Difficult Case

When the result is ambiguous, don’t check the same text multiple times hoping for a different result. Instead, compare it with earlier texts by the same person, look at the document’s version history, and talk with the author about the content. The detector provides one of several clues, and the decision belongs to a human.

A good habit is also recording a note with each scan: what you did with the result and how the conversation with the author ended. After a few months you have your own knowledge about how the detector performs in your field.

// read morePangram AI Detector: How It Works and PricingAI ToolsHow to Tell if AI Wrote a Text: 7 Signs and How to CheckGuidesPangram in Practice: School, Newsroom, Agency and HiringGuides

Kinetyra