Powered by Pangram
Every article on AI News Audit is analyzed with Pangram, an AI detector. We now use version 3.3. It reads an article in short segments and rates how much of each one looks AI-generated, with a confidence level for every call.
The pipeline
How Pangram reads an article
Split into windows. The article is divided into short, overlapping passages called windows. Each window is scored on its own, so an AI passage inside an otherwise human article is still picked up.
Label every window. Each window gets a label (such as Human Written or AI-Generated), an AI score from 0 to 1, and a confidence of High, Medium, or Low.
Combine into a verdict. The window results are combined into one rating for the whole article (Human, Mixed, or AI), along with the share of the text in each category.
The verdicts
The three labels
Written by a person, with no sign of AI.
A mix: some passages read as human, others as AI.
Generated primarily by an AI model.
On this site
The scores you'll see
Each article card and detail page shows a few numbers, all based on the window analysis:
AI Likelihood
The article's overall AI score, averaged across its windows and shown as a percentage. Higher means more of the article reads as AI.
Max AI
The highest score of any single window. This helps flag one AI passage even when the rest of the article looks human.
AI Content Fraction
The share of the text, from 0 to 100%, that was classified as AI-written.
Confidence
Each window also gets a confidence of High, Medium, or Low, showing how sure the model is about its label. New in v3.3.
What changed
New in version 3.3
- Finer segment analysis. Version 3.3 scores many short, overlapping windows instead of giving a single overall number, so one AI passage in an otherwise human article is caught.
- Confidence levels. Each window now reports High, Medium, or Low confidence along with its score.
- Per-window labels. Each passage gets its own label, such as Human Written or AI-Generated, instead of one overall probability.
- A breakdown by share. Version 3.3 reports how much of the article is human versus AI, not just one number.
Field definitions follow Pangram's published text-classifier API. The reported false-positive rate of 0.001 is for news articles.
AI scores by Pangram v3.3.