Is Pangram a Reliable AI Detector?
AI Detectors

Is Pangram a Reliable AI Detector?

Shadab Sayeed
Written by Shadab Sayeed
July 06, 2026
Calculating…

AI detectors are often used in schools, colleges, and student communities, but the important question is simple: can we actually trust them? To answer that, I tested Pangram on a dataset where the real writer was already known. Each text was marked as either written by AI or written by a human, and then Pangram's verdict was compared with that truth.

The short answer from this test is that Pangram looked reliable overall, but it was not perfect. It made very few false accusations against human writing, which is a big plus for students. The bigger weakness was that it sometimes missed AI-written text and called it human.

Also Read: [STUDY] Is Pangram AI Detector Similar to Turnitin?

What was tested?

The dataset had 160 writing samples. Out of these, 82 were AI-written and 78 were human-written. That is a fairly balanced dataset, which makes the test more useful because one side does not completely dominate the result.

  • Total samples: 160
  • AI-written samples: 82
  • Human-written samples: 78
  • Correct Pangram verdicts: 149
  • Wrong Pangram verdicts: 11

Also Read: StealthWriter vs Pangram: Can a "Humanized" Rewrite Really Escape Detection?

Pangram correct and wrong verdicts
Pangram gave the correct verdict for 149 out of 160 samples in this dataset.

The main result

Pangram reached an overall accuracy of 93.1%. In simple words, it was correct about 93 times out of every 100 samples in this test. That is a strong result, especially for a tool that is trying to judge writing style instead of checking a fixed fact.

But students should not confuse high accuracy with perfection. A detector can be good on average and still be wrong for a specific essay. This matters because one incorrect result can create a serious problem for a student if the result is treated as final proof.

Also Read: [STUDY] Can Pangram Detect Undetectable AI?

Metric Result What it means
Overall accuracy 93.1% How often Pangram was correct overall
AI precision 98.6% When Pangram said AI, how often it was right
AI recall 87.8% How many AI-written samples Pangram caught
Human clear rate 98.7% How many human-written samples Pangram correctly cleared
Pangram reliability metrics
The detector was strongest when it flagged something as AI or cleared human writing.

Where Pangram did well

The most impressive part of this test was Pangram's low false positive rate. A false positive means a human-written sample gets wrongly called AI. In this dataset, that happened only 1 time out of 78 human samples.

That matters a lot for students. A tool that constantly accuses real human writing of being AI-generated would be dangerous in academic settings. In this test, Pangram was much better at avoiding that problem than many students might expect.

  • Pangram correctly cleared 77 out of 78 human-written samples.
  • It wrongly flagged only 1 human-written sample as AI.
  • Its false positive rate was only 1.3%.
Pangram performance by original writer
Pangram made more mistakes on AI-written samples than on human-written samples.

Where Pangram struggled

Pangram's main weakness in this dataset was missing AI-written text. Out of 82 AI samples, it correctly detected 72, but missed 10. That gives it an AI recall of 87.8%.

This means Pangram was strong, but some AI content still slipped through. For students, this is an important point. Passing an AI detector does not automatically prove that a text is human-written. It only means the detector did not find enough signs to label it as AI.

  • False negatives: 10 AI-written samples were marked as human.
  • False negative rate: 12.2% of AI samples were missed.
  • The missed AI samples usually received very low Pangram scores in this dataset.

What the score pattern showed

The Pangram scores in this file were not spread across many values. They mostly appeared as very strong signals. A score of 100 matched Pangram's AI verdict, while low scores matched Pangram's human verdict. This made the detector's decision style look quite strict in this dataset.

Most AI-written samples received a score of 100, and most human-written samples received a score of 0. This explains why the overall accuracy was high. The detector was usually very confident in one direction or the other.

Pangram score distribution for AI and human samples
Most AI samples were pushed to 100, while most human samples stayed close to 0.

Did text length affect the result?

I also checked whether Pangram made more mistakes on shorter or longer texts. The error rate did not show one simple rule. Texts between 301 and 500 words had the most errors by count, but that is also where many samples were located. The over 500 word group had no errors, but it only had 8 samples, so we should not overread that result.

Pangram error rate by text length
Text length did not create a clear, simple pattern in this dataset.

So, is Pangram reliable?

Based on this dataset, Pangram looks reliable as a screening tool. It got a high overall accuracy, it was very good at avoiding false accusations against human writing, and when it said something was AI, it was almost always correct in this test.

However, it should not be treated as a final judge. The detector still missed 10 AI-written samples, and it wrongly flagged 1 human-written sample. That means the safest conclusion is not "Pangram is always right." The better conclusion is: Pangram performed strongly in this test, but its verdict should be used with context.

Student-friendly takeaway: Pangram seems useful for spotting AI-written text, but no AI detector should be the only evidence used against a student. The writing process, drafts, notes, sources, and teacher judgment still matter.

Final verdict

My answer is: yes, Pangram looked reliable in this test, but not perfect. Its biggest strength was protecting human writing from false AI labels. Its biggest weakness was that some AI-written samples passed as human. For students, that means Pangram can be helpful, but it should be treated as a signal, not as unquestionable proof.

About the Author
Shadab Sayeed

Shadab Sayeed

CEO & Founder · DecEptioner
Dev Background
Writer Craft
CEO Position
View Full Profile

Shadab is the CEO of DecEptioner — a developer, programmer, and seasoned content writer all at once. His path into the online world began as a freelancer, but everything changed when a close friend received an 'F' for a paper he'd spent weeks writing by hand — his professor convinced it was AI-generated.

Refusing to accept that, Shadab investigated and found even archived Wikipedia and New York Times articles were being flagged as "AI-written" by popular detectors. That settled it. After months of building, DecEptioner launched — a tool built to defend writers who've been wrongly accused. Today he spends his days improving the platform, his nights writing for clients, still driven by that same moment.

Developer Content Writer Entrepreneur Anti-AI-Detection