15 Forensic Markers of AI-Generated Content (Part 2)
Session 1.6 · ~5 min read
The first eight markers covered structural and rhetorical patterns. Markers 9 through 15 target behavioral patterns: the ways AI avoids commitment, simulates empathy, and hedges against every possible objection. Together, all 15 markers form a diagnostic checklist you can apply to any text.
Marker 9: The Safety Disclaimer
"It's worth consulting a professional before making any decisions." "Please consult your doctor before starting any new exercise routine." "This is not financial advice." These disclaimers appear in AI text regardless of context. An article about picking houseplants will advise you to "do your own research" as if choosing between a fern and a succulent carries legal liability. The pattern comes from RLHF safety training, which penalizes any response that could be interpreted as giving risky advice.
Marker 10: Parallel Structure Overuse
Every paragraph begins the same way. "One approach is..." "Another approach is..." "A third approach is..." Or: "First, consider..." "Next, examine..." "Finally, evaluate..." Parallel structure is a legitimate writing technique, but AI applies it mechanically, producing pages where every section follows an identical syntactic template. Human writers vary their paragraph openings naturally. AI does not, because parallel structure is a safe, easily-generated pattern.
Marker 11: The Non-Answer Answer
Ask AI a direct question with a clear answer, and it will sometimes produce 500 words that avoid committing to that answer. "What programming language should I learn first?" generates a response covering Python, JavaScript, and Rust, explaining the merits of each, and concluding with "The best language depends on your goals." This is not helpful. The reader asked for a recommendation. The AI produced a brochure.
The non-answer answer exists because RLHF punishes wrong recommendations more than it rewards correct ones. Recommending Python when the reader should learn JavaScript scores worse than recommending nothing. So the model recommends nothing, at length.
Marker 12: Context-Free Confidence
"Studies show that remote workers are 13% more productive." Which studies? Conducted by whom? In what year? With what sample size? Under what conditions? AI produces citations-shaped sentences with no actual citations. The confidence of the claim is inversely proportional to its specificity. The more certain the language, the less likely you are to find a source behind it.
Context-free confidence is the opposite of expertise. Real experts cite their sources and acknowledge limitations. AI sounds certain because certainty is a pattern, not a conclusion.
Marker 13: The Empathy Prefix
"I understand this can be frustrating..." "It's completely normal to feel overwhelmed..." "Many people struggle with this..." These prefixes simulate empathy before delivering generic advice. The pattern comes from customer service training data and RLHF optimization for "helpful" responses. Real empathy is specific: "I spent six months trying to get this right and failed three times before it worked." AI empathy is a template applied before any substantive content.
Marker 14: Temporal Vagueness
"In recent years..." Which years? "As technology continues to evolve..." Evolve how? "The landscape has changed significantly..." Changed from what to what? AI uses temporal language that sounds current without committing to any specific timeframe. This is because the model's training data spans years, and it cannot determine what is "recent" relative to the current date. The result is prose that sounds timely but is actually timeless in the worst sense: it could have been written at any point in the last decade.
Marker 15: The Closing Platitude
"The future is bright for those who embrace these changes." "By implementing these strategies, you'll be well on your way to success." "The journey of a thousand miles begins with a single step." AI closes with platitudes because RLHF rewards optimistic, encouraging endings. The closing platitude adds no information. It exists to make the reader feel good about having read the article, which is a lower bar than making the reader actually better informed.
The Complete Diagnostic Checklist
| # | Marker | Quick Identifier |
|---|---|---|
| 1 | Comprehensive Guide Opening | "In this comprehensive guide..." |
| 2 | Tricolon Abuse | Three-item parallel phrases, repeatedly |
| 3 | False Bridge | "But here's the thing..." |
| 4 | Premature Summarization | "In summary" before the article is half done |
| 5 | Hollow Metaphor | Metaphor that breaks after one comparison |
| 6 | Over-Attribution | "Studies show" with no citation |
| 7 | Enthusiasm Spike | Sudden exclamation in flat prose |
| 8 | Synonym Cycling | Same idea, three different words in three sentences |
| 9 | Safety Disclaimer | "Consult a professional" in non-critical context |
| 10 | Parallel Structure Overuse | Every paragraph starts with the same syntax |
| 11 | Non-Answer Answer | 500 words that avoid giving a direct recommendation |
| 12 | Context-Free Confidence | Specific-sounding claims with no sources |
| 13 | Empathy Prefix | "I understand this can be frustrating..." |
| 14 | Temporal Vagueness | "In recent years..." (which years?) |
| 15 | Closing Platitude | "The future is bright for those who..." |
Human or well-edited"] D -->|"4-8"| F["Moderate:
AI-assisted, some editing"] D -->|"9-15"| G["Heavy:
Minimal editing"] D -->|"16+"| H["Unedited AI:
Standard slop"]
The checklist is a tool, not a verdict. Human writers occasionally exhibit some of these patterns. The diagnostic power is in density and combination. A text with 3 markers might be human-written with some careless habits. A text with 12 markers in 1,000 words is almost certainly unedited AI output.
Further Reading
- A Survey of AI-generated Text Forensic Systems (arXiv)
- How Does AI Detection Work? A Complete Guide (Link-Assistant)
- GPTZero: AI Detection (GPTZero)
- AI Content Detector (Copyleaks)
Assignment
- Create your personal "AI Detection Checklist": a one-page reference document listing all 15 markers.
- Format it as a table with four columns: Marker | Description | Example | Fix (how to rewrite the marker into human-quality prose).
- Test your checklist on 3 different texts: one you know is AI-generated, one you know is human-written, and one you are unsure about.
- Record the marker counts for each. Does the checklist correctly differentiate them?