## Model-Specific Fingerprints

Each AI model has unique tells & avoidance patterns — how to ID which one wrote the text

| Model    | Uses                                                                                                                                               | Avoids                                      |
|----------|----------------------------------------------------------------------------------------|---------------------------------------------|
| ChatGPT  | - •Uses 'delve' constantly<br>- •Lists/bullet points heavy<br>- •Drama starters ('This changes everything')<br>- •'It's not X, it's Y' patterns<br>- •Em dash 10× increase (GPT-3.5→4o)<br>- •Rule of three compulsion | - •Definitive claims<br>- •Simple language<br>- •Unpredictable structure |
| Claude   | - •'You're absolutely right!'<br>- •Uses 'quiet' uniquely<br>- •Heavy em dash usage<br>- •XML tag affinity<br>- •Flowing paragraphs over lists<br>- •Context before answer   | - •delve<br>- •tapestry<br>- •landscape<br>- •beacon |
| Gemini   | - •Simpler vocabulary<br>- •Uses 'might' frequently<br>- •Skips preambles<br>- •More italics than others<br>- •Straight-to-answer structure<br>- •Shorter responses        | - •Flowery language<br>- •Long intros<br>- •Personality quirks |
| Grok     | - •Internet slang<br>- •Pop culture references<br>- •Sarcasm/humor<br>- •Casual endings<br>- •Flowing prose over bullets<br>- •Unpredictable organization                  | - •Corporate speak<br>- •Hedging<br>- •Formal structure |
| DeepSeek | - •'comprehensive/crucial'<br>- •Chain-of-thought visible<br>- •Business tone<br>- •Compact output<br>- •'Below is' phrasing<br>- •Step-by-step reasoning                      | - •Creative flourishes<br>- •Long explanations |

## Detailed Model Characteristics

### ChatGPT

The Enthusiastic Overexplainer

**Sycophancy:** 56.71%  
**Readability:** Grade 12 (Flesch-Kincaid)

Signature Patterns

- •Uses 'delve' constantly
- •Lists/bullet points heavy
- •Drama starters ('This changes everything')
- •'It's not X, it's Y' patterns
- •Em dash 10× increase (GPT-3.5→4o)
- •Rule of three compulsion

Actively Avoids

- •Definitive claims
- •Simple language
- •Unpredictable structure

### Claude

The Measured Hedger

**Sycophancy:** 57.44%  
**Readability:** Literary register

98% wrongly admitted mistakes when challenged (Claude 1.3)

Signature Patterns

- •'You're absolutely right!'
- •Uses 'quiet' uniquely
- •Heavy em dash usage
- •XML tag affinity
- •Flowing paragraphs over lists
- •Context before answer

Actively Avoids

- •delve
- •tapestry
- •landscape
- •beacon

### Gemini

The Practical Middle Ground

**Sycophancy:** 62.47%  
**Readability:** Most clinical tone

Highest sycophancy rate at 62.47%

Signature Patterns

- •Simpler vocabulary
- •Uses 'might' frequently
- •Skips preambles
- •More italics than others
- •Straight-to-answer structure
- •Shorter responses

Actively Avoids

- •Flowery language
- •Long intros
- •Personality quirks

### Grok

The Edgy Contrarian

**Sycophancy:** 0%  
**Readability:** Grade 7-8 (Flesch-Kincaid 61.4) - Most readable

0% overlap with any other model family (CMU classifier), 100% no-agreement in Copyleaks

Signature Patterns

- •Internet slang
- •Pop culture references
- •Sarcasm/humor
- •Casual endings
- •Flowing prose over bullets
- •Unpredictable organization

Actively Avoids

- •Corporate speak
- •Hedging
- •Formal structure

### DeepSeek

The Concise Analyst

**Sycophancy:** 0%  
**Readability:** Policy paper style

74.2% classified as OpenAI output despite independent development

Signature Patterns

- •'comprehensive/crucial'
- •Chain-of-thought visible
- •Business tone
- •Compact output
- •'Below is' phrasing
- •Step-by-step reasoning

Actively Avoids

- •Creative flourishes
- •Long explanations

## Model Sycophancy Rates

How often AI models agree with users regardless of correctness

Highest Sycophancy

Gemini: 62.47%  
Most likely to agree with user

Regressive Sycophancy

Claude: 18.31%  
Leads users to wrong answers through agreement

✓ Most Objective

Grok: 0%  
Least likely to agree just to please

**What is Sycophancy?** The tendency of AI models to agree with users' statements regardless of factual accuracy. Claude 1.3 wrongly admitted mistakes on **98% of questions** when challenged by users.