Copyleaks Research Reveals AI Has Unique Stylistic Fingerprints

Copyleaks Identifies 74.2% Stylistic Overlap Between DeepSeek-R1 and OpenAI’s Model
March 3, 2025

AS FEATURED ON

Forbes Logo

IN THIS REPORT

Key Takeaways

  • Distinct AI Signatures: Each AI model exhibits a unique and identifiable writing style, even when attempting to mimic other styles or addressing various topics.

  • High Identification Accuracy: Copyleaks’s methodology can accurately determine the originating AI model with a precision rate of 99.88%, even when analyzing content from previously unseen AI models.

  • Implications for Intellectual Property and Transparency: Identifying the source of AI-generated content is crucial for protecting intellectual property rights, maintaining transparency about AI-generated material, preventing misuse of AI technologies, and building trust in online content.

  • Potential for Enhanced AI Governance: This research opens avenues for better governance and compliance in AI usage. It allows for tracing AI-generated content back to its source, promoting responsible AI adoption. 

Copyleaks Research Identifies Which AI Wrote What

Just as every human has a unique fingerprint, recent Copyleaks research, featured in Forbes, reveals that AI also has its own distinct, identifiable signature.


The Copyleaks Data Science team developed a precise method for identifying which AI model wrote a particular text, even when the AI deliberately tries to imitate another writing style. 


But does it matter if we can identify what AI wrote what? Indeed, it does. 

The Why: Do AI Fingerprints Matter?

AI-generated content is everywhere, and it’s saturating the internet more and more. That’s why identifying the source of the content becomes crucial, like tracing the origin of a painting or a piece of music. Having this ability with genAI allows us to: 

  • Protect the rights around intellectual property

  • Maintain transparency about what’s AI-generated 

  • Prevent misuse of AI technology 

  • Build trust in online content

The Method: How Copyleaks Identified AI’s Writing Style

The researchers approached this challenge like digital detectives. Instead of relying on a single method, they created an ensemble of three different AI “investigators” looking for different text patterns and characteristics. These “investigators” only make a final judgment if and when all three of them agree on a conclusion, similar to how the court systems require unanimous jury verdicts. 


This method not only helped distinguish content created by major AI models with 99.88% accuracy but also detected the unique “stylistic fingerprints” of unseen AI models, revealing their distinct writing styles and potential relationships to known AI systems. 

The Subjects: The Big Four

The study focused on the four major models in the AI market: 

Claude Logo
Gemini Logo
ChatGPT Logo

Interestingly, each AI model maintains its distinct writing style even when asked to mimic different voices or address varying topics. It’s similar to how you might recognize your favorite author’s style, whether writing a horror or a romance novel. 

The Findings: Unexpected Family Resemblence

When the researchers at Copyleaks tested their system on an AI model that it had not been trained on, they uncovered some intriguing findings:

  • Some newer models, like phi-4 and Grok-1, showed completely unique writing styles that didn’t match any of the major AI models.

     

  • The Mixtral model showed some similarities to both OpenAI and Llama’s writing styles but still maintained its own distinct voice.

     

  • Most surprisingly, the DeepSeek-R1 model wrote in a style remarkably similar to OpenAI’s models, raising questions about potential shared development approaches. 
Data Bar Graph - Predictions by model-writer (combined)

What These Findings Mean for the Future

The research findings present new possibilities for the future of AI content, including:

Better Transparency

We could see automated systems that clearly identify which AI-generated specific content, similar to how we attribute human authorship. 

Improved Security

This could help detect unauthorized use of AI models or attempts to pass off AI-generated content as coming from a different source. 

Evolution Tracking

As AI models continue to develop, we can better understand how their writing styles evolve and influence each other. 

Fair Use Protection

Organizations can better protect their AI models from unauthorized use or copying. 

Looking Ahead

While this research represents a significant breakthrough, there’s still more work to be done. 

Future studies will likely explore: 

  • How these AI fingerprints show up in different languages

  • More detailed analysis of what makes each AI’s writing style unique

  • Expanding the system to recognize even more AI models

  • Understanding how these writing patterns might change as AI evolves

 

The ability to identify AI authors with such high accuracy marks is an important step forward in making AI more transparent and trustworthy. As AI continues to play a larger role in content creation, tools like this will become increasingly valuable for maintaining integrity in the digital world. 

View the full data report from the Copyleaks Data Science Team.

View the Full Report