AI VS HUMAN WRITING:
HOW TO SPOT THE DIFFERENCE
Language models and human writers approach text generation from completely different foundations: statistical token prediction versus idiosyncratic human experience.
The Mathematical Difference: Perplexity and Burstiness
To understand the difference between human and synthetic writing, linguists and computer scientists look at two primary metrics:
1. Perplexity (Token Predictability)
Perplexity measures how "surprised" a language model is by a sequence of words. Because AI generates text by predicting the mathematically most likely next token, synthetic writing almost always exhibits low perplexity. Humans, by contrast, regularly use unusual analogies, unexpected vocabulary, and non-linear phrasing.
2. Burstiness (Sentence Variation)
Burstiness refers to the rhythm and variation in sentence length and structure. Human writers naturally vary their pace: a long, complex compound sentence is often followed by a short one. Like this. AI drafts, on the other hand, tend to cluster around uniform sentence lengths (18–24 words) and identical cadence.
Anatomy of a LinkedIn Post: AI vs Human
Examine these two fictional posts reflecting common patterns encountered on LinkedIn:
Here are 5 key pillars to unlock 10x agility before Q4 ends:
1. Embrace continuous iteration
2. Cultivate psychological safety
3. Leverage cross-functional alignment...
Agree? Repost to help a colleague navigate the future of work! 🚀"
- Cliché opening trope ("In today's fast-paced...")
- Manufactured platitudes ("stepping stone to synergy")
- Uniform sentence lengths with formulaic numbered listicle
- Comment/engagement bait callout ("Agree? Repost!")
Zero philosophical takeaways. Just happy the alerts stopped firing at 3 AM."
- Concrete empirical details and numbers
- High burstiness (mix of long narrative and blunt fragments)
- Informal, self-deprecating humor
- No synthetic calls to action or generic advice lists
Understanding False Positives & False Negatives
Stylistic signals are indicators of probability, not forensic proof of authorship:
What causes a False Positive?
When a human writes in a highly structured, corporate, or formal style (such as standard compliance updates, corporate press releases, or non-native English speakers relying on standard textbook phrases), their text may exhibit low variance and be flagged as AI-like.
What causes a False Negative?
If a user prompts an LLM with specific idiosyncratic stylistic constraints or heavily edits the output by inserting personal anecdotes and fragmented sentences, the resulting burstiness increases, potentially causing an AI post to be classified as human.
Continue Reading About AI Content Detection
Deepen your understanding of computational detection with our comprehensive guides: