{"aiVersion":"1","content":{"id":"cmswj45a50004ih5nz9imlxcj","slug":"how-pangram-checks-writing-a-simple-guide-for-lear-20260817","title":"How Pangram Checks Writing — A Simple Guide for Learners","level":"MEDIUM","publishedAt":"2026-08-17T01:02:19.134Z"},"topic":{"slug":"how-pangram-checks-writing-a-simple-guide-for-lear-20260817","category":"technology"},"article":{"paragraphs":["Pangram is a statistical detector that estimates which parts of a text were written by AI, AI-assisted, or human-written. The tool has been adopted by platforms and researchers, and Substack added Pangram on July 21, 2026 so readers can scan posts. Pangram's team reports a near-zero false-positive rate, and the company provides line-by-line highlights to show which sentences look machine-made.","Independent studies have tested Pangram alongside other detectors such as GPTZero, Turnitin and Copyleaks, and results have varied. A June 29, 2026 paper in the International Journal for Educational Integrity found Pangram aligned closely with ground-truth in many academic tests, while other work has shown that adversarially edited AI text can evade detection. Therefore, experts say Pangram is useful for screening but it should be used with care in high-stakes decisions.","Pangram's creators have published technical reports and model updates, and the company now offers an image-detection research preview. Although the image tool has performed well on some benchmarks, it has been shown to make errors when images are mixed with real photos. Therefore, while text flags can be informative, users are advised to double-check context and not rely solely on a single score."],"wordCount":194,"readTime":2},"vocabulary":[{"word":"statistical","example":"The model gives a statistical estimate for each sentence.","phonetic":"/stəˈtɪstɪkəl/","definition":"based on data or numbers"},{"word":"false-positive","example":"A false-positive label can upset someone wrongly accused.","phonetic":"/ˌfɔːls ˈpɒzətɪv/","definition":"when a test says something is true but it is not"},{"word":"adversarial","example":"Adversarial edits can hide AI writing.","phonetic":"/ˌæd.vəˈsɛr.i.əl/","definition":"designed to trick a system"},{"word":"benchmark","example":"The model was tested on several benchmarks.","phonetic":"/ˈbɛn.tʃmɑːrk/","definition":"a standard test to measure performance"}],"quiz":[{"answer":"July 21, 2026","question":"When did Substack add Pangram?"},{"answer":"near-zero false-positive rate","question":"What phrase is used to describe Pangram's low false positives?"},{"answer":"human-written, AI-assisted, or AI-generated","question":"What three labels does Pangram estimate for text?"}],"discussion":[{"question":"Do you feel more confident reading articles that are labeled human-written? Why or why not?"},{"question":"Have you ever used an online detector for your writing? What did you learn?"},{"question":"What steps would you take if a tool flagged your own text as AI-written?"},{"question":"Would you change how you write if platforms show AI scores on posts?"}]}