Machine Learning Times
Machine Learning Times
EXCLUSIVE HIGHLIGHTS
Three Best Practices for Unilever’s Global Analytics Initiatives
    This article from Morgan Vawter, Global Vice...
Getting Machine Learning Projects from Idea to Execution
 Originally published in Harvard Business Review Machine learning might...
Eric Siegel on Bloomberg Businessweek
  Listen to Eric Siegel, former Columbia University Professor,...
Effective Machine Learning Needs Leadership — Not AI Hype
 Originally published in BigThink, Feb 12, 2024.  Excerpted from The...
SHARE THIS:

1 month ago
Which AI Model Most Infringes on Copyrighted Content?

 
Originally published in AI Business, March 7, 2024. 

OpenAI’s GPT-4 reproduces the most copyrighted content from prompts among four popular large language models, according to new research from AI startup Patronus AI.

The startup, founded by former Meta AI researchers, also found that popular large language models from the likes of Meta, Mistral and Anthropic generated copyrighted content.

The startup tested OpenAI’s GPT-4, Anthropic’s Claude 2.1, Meta’s Llama 2 70B and Mistral’s Mixtral-8x7B-Instruct-v0.1.

GPT-4 reproduced copyrighted content, on average, in 44% of prompts crafted to test how a model regurgitates existing content. Mixtral-8x7B-Instruct-v0.1 produced copyrighted content on 22% of test prompts on average, while Llama 2 70B recreated content on 10% of the prompts.

The model that produced the lowest amount of copyrighted content was Anthropic’s Claude 2.1, with an average score of just 8%.

To continue reading this article, click here.

One thought on “Which AI Model Most Infringes on Copyrighted Content?

Leave a Reply