How to Detect AI-Generated Text: Technical vs. NLP Approaches
Two fundamentally different approaches to AI detection exist today. Here's how they work, where they fail, and how to combine them for the best results.
The Two Schools of AI Detection
AI text detection has split into two distinct approaches, each with different strengths and weaknesses. The first is NLP-based detection — analyzing writing patterns, word choice, and sentence structure to determine statistical likelihood of AI authorship. The second is technical fingerprint detection — scanning for concrete evidence like hidden characters and HTML attributes that AI interfaces inject into copied text.
Understanding both approaches is crucial because they catch different things. Using only one method leaves blind spots that the other method covers.
NLP-Based Detection: How It Works
Tools like GPTZero, Originality.ai, Copyleaks, and Turnitin's AI module use machine learning models trained to recognize statistical patterns in AI-generated writing. They measure properties like perplexity (how predictable each word is given the context) and burstiness (how much variation exists in sentence complexity and length).
AI models tend to produce text with consistently low perplexity — each word flows predictably from the previous context. Human writing is typically more variable, with sudden topic shifts, unusual word choices, and irregular sentence lengths. NLP detectors exploit these statistical differences.
Strengths: Can detect AI text even after it's been cleaned of technical markers. Works on text from any source, including AI APIs without web interfaces. Can analyze writing style regardless of how the text was delivered.
Weaknesses: Produces probability scores, not proof. False positives are a documented problem — human-written academic text is sometimes flagged as AI. Lightly edited AI text often passes detection. The accuracy varies significantly by writing style, topic, and language.
Technical Fingerprint Detection: How It Works
Technical detection scans the raw text data for concrete, verifiable evidence of AI tool origin. This includes invisible Unicode characters (zero-width spaces, byte order marks, word joiners), proprietary HTML attributes (ChatGPT's data-start, data-end), and formatting patterns (smart quotes, em dashes, Unicode ellipsis) that AI interfaces inject.
These markers are deterministic evidence: a ChatGPT data-start attribute either exists in the text or it doesn't. There's no probability involved, no gray area, no chance of a false positive for technical markers with critical severity.
Strengths: Zero false positives on critical markers. Results are binary and provable. Fast (milliseconds vs. seconds for NLP). Works entirely client-side with no data upload required. Identifies the specific AI tool used (ChatGPT vs. others).
Weaknesses: Only detects text that retains copy-paste artifacts. If someone cleans the text or manually retypes it, technical markers are gone. Doesn't work on API-generated text that never passed through a web interface.
The Optimal Detection Workflow
The most reliable approach combines both methods. Start with technical fingerprint detection to find definitive evidence. If markers are found, you have proof. If no technical markers are found, that doesn't mean the text is human-written — it could have been cleaned or generated through an API.
Next, run the text through an NLP-based detector for stylistic analysis. If both methods agree (both say AI, or both say human), you have high confidence. If they disagree — say, no technical markers but high NLP probability — the text may have been AI-generated and cleaned, which is useful context.
This layered approach eliminates the blind spots of each individual method: technical detection provides proof when available, and NLP analysis provides coverage when technical markers have been removed.
The Future of AI Detection
As AI models improve and more users learn to clean technical markers, NLP-based detection faces an arms race — each improvement in detection is met with improvements in evasion. Technical detection, however, is less affected by this dynamic. As long as AI tools use web interfaces that process and render text, there will be technical artifacts to detect.
OpenAI has discussed implementing cryptographic watermarking — embedding statistical patterns at the model level that survive editing. If implemented, this would represent a third detection approach: one that combines the reliability of technical detection with the persistence of NLP detection. Until then, the two-method approach remains the gold standard.