AI Detection Metrics Explained
Understanding how we measure AI-generated content in news articles
How does AI detection work?
We use Pangram, a state-of-the-art AI detector, which analyzes text on two levels:
- Chunk-level: Breaks the article into chunks and analyzes each one for AI use.
- Article-level: Uses those chunk-level results to give an overall judgment for the entire article.
Content Categories
Human
Written entirely by humans
Mixed
A combination of human and AI content
AI
Generated primarily by AI
AI Likelihood
Overall article score
The overall probability that the entire article contains AI-generated content.
Interpretation: Higher percentages indicate greater likelihood of AI generation
Max AI Likelihood
Highest chunk score
The highest AI likelihood score found in any single chunk of the article. Useful for spotting a "hidden" AI paragraph, even if the rest looks human.
Use: Identifies specific AI-generated sections within the article
Average AI Likelihood
Average chunk score
The average AI likelihood score across all chunks of the article.
Use: Shows overall AI content consistency across the article
AI Delta
Variation in chunk scores
The difference between the maximum and average AI likelihood scores. High delta means some parts have more AI content than others.
High Delta: Mixed human/AI content
Low Delta: Consistent content (AI or human)
AI Content Fraction
Percentage of AI-generated content
Estimates what percentage of the total text was written by AI. More specific than other scores — it tries to answer "how much" of the article is AI.
Examples:
• 0% = Fully human-written
• 50% = Half AI, half human
• 100% = Fully AI-generated