1. The Biology of Human Writing vs. Autoregressive Sampling
When a human sits down to write about a professional experience, their thoughts do not proceed along a single linear probability vector. A human author experiences cognitive friction:
- They stop and search for a rare word that captures an exact tactile memory.
- They write a convoluted run-on sentence, get frustrated, and follow it with an abrupt sentence fragment.
- They use idiosyncratic regional colloquialisms and personal verbal tics.
A Large Language Model (such as GPT-4o, Claude 3.5, or Llama 3) does none of this. Given a prompt, an LLM projects the sequence of existing tokens through transformer attention heads, calculating a probability distribution over a vocabulary of approximately 100,000 tokens. When generating words, sampling strategies like top-p (nucleus sampling) or temperature restrict choices to high-probability tokens.
The result is prose that is grammatically flawless, beautifully balanced, and statistically completely un-human. For a visual comparison, visit our educational AI vs. Human Writing Analysis page.
2. The Twin Pillars of Detection: Perplexity and Burstiness
Every modern AI text detector operates on two fundamental metrics:
Perplexity (Predictability)
Perplexity is the exponentiated cross-entropy of a text sequence. In simpler terms, it measures how surprised a language model is by the next word. If text consistently picks the 1st or 2nd most probable word according to standard language distributions, perplexity is low. Human writing contains frequent high-perplexity spikes (unusual metaphors, slang, rare vocabulary).
Burstiness (Sentence Variance)
Burstiness measures the variance in sentence length, grammatical structures, and syntactic rhythm. Human writers naturally alternate between long, complex sentences and punchy statements. Machines produce uniform, metronomic cadence where sentence after sentence sits between 16 and 22 words.
When a detector evaluates text, it plots sentences onto a two-dimensional coordinate system of perplexity versus burstiness. Text clustering in the low-perplexity, low-burstiness quadrant triggers high AI probability scores.
3. The "Delve" Effect: RLHF Lexical Artifacts
Beyond raw statistical curves, models exhibit distinct lexical fingerprints created during Reinforcement Learning from Human Feedback (RLHF). Human evaluators during the training process consistently rewarded completions that sounded "thorough, polite, and sophisticated."
As an unintended consequence, words that human evaluators rated highly became over-represented in the model's output weights. Words like delve, tapestry, beacon, testament, crucial, and underscore appear in LLM text at frequencies up to 400% higher than in contemporary American business writing.
To see how these habits manifest specifically on professional social platforms, see our guide: How to Know If Someone Used ChatGPT on LinkedIn.
4. Why 100% Accuracy Is a Mathematical Impossibility
Any vendor or service claiming "100% accurate AI detection" is misleading its users. Language is an open-ended, infinite combinatorial system. It is entirely possible for a human being—especially in formal contexts like legal drafting, grant writing, or corporate boilerplate—to write with low perplexity and regular sentence cadence.
Similarly, an advanced prompt engineer can command an LLM: "Introduce intentional sentence length variance between 3 and 45 words, include two conversational fragments, avoid academic transitional phrases, and use idiosyncratic punctuation."
This is why AI Feed Lens presents confidence intervals and probabilistic classifications rather than arbitrary binary declarations of guilt. For a critical look at edge cases, read our essay on Why AI Writing Detectors Can Be Wrong.
Frequently Asked Questions
What is the difference between perplexity and burstiness?
Perplexity measures word-level surprise (how predictable each word choice is given the prior words), whereas burstiness measures document-level variance in sentence length, structural cadence, and rhythm.
Can an AI detector flag human writing as AI?
Yes, this is known as a false positive. False positives most frequently occur when human writers adhere strictly to repetitive corporate templates, legal boilerplate, or academic styles that naturally lack emotional burstiness.
How does AI Feed Lens evaluate burstiness in short feed posts?
AI Feed Lens applies sentence-segmentation algorithms calibrated for micro-text. It normalizes for social formatting (single-line breaks) to measure true clause complexity and syllable diversity across the post.
Does AI watermarking exist in web text?
While research into cryptographic token watermarking exists, no major commercial LLM currently enforces client-visible text watermarking. Therefore, statistical detection (perplexity and burstiness) remains the gold standard.
Can an extension check text in real-time as I scroll?
Yes! AI Feed Lens does exactly this using lightweight DOM parsing, displaying a non-intrusive badge next to posts on LinkedIn directly as you scroll. Read our guide on Detecting AI Writing in the Browser.
SEE STATISTICAL AI DETECTION IN ACTION
Test your own writing or inspect social posts with AI Feed Lens. Experience the power of client-side linguistic evaluation right inside Chrome.