Copyleaks 研究发现近 60% 的 GPT-3.5 输出包含某种形式的抄袭内容

New York, NY – February 22, 2024 – According to a 2023 报告到 2026 年，几乎 90% 的在线内容将由人工智能生成。由于人工智能内容饱和，人们担心数据污染和不可避免的 模型崩溃 raise concerns about AI-generated text’s overall quality and reliability.

此外，人们对原创性的担忧也开始更加广泛。 几起诉讼 regarding AI infringing on copyright and potentially plagiarizing, educational institutions and enterprises across the globe are questioning the authenticity of AI text: Where did it originate from? Is it safe to use as original content

归根结底，AI抄袭了吗？

To find out, Copyleaks, the leader in plagiarism identification, AI-content detection, and GenAI governance, conducted an analysis to determine the degree to which AI-generated content is original and free of potential plagiarism.

要进行此分析：

Copyleaks asked GPT-3.5 to write 1,045 个输出，平均 412 个字 在所有输出中， 26 subjects: Physics, Chemistry, Science, Psychology, Law, Economics, Biology, Business Studies, Engineering, Accounting, Geography, Mathematics, Computer Science, Sports, World History, Philosophy, English Language, Art, Physical Education, Statistics, Social Science, Nature, Music, Sociology, Humanities, Theater. 

Copyleaks gauged the specific outputs with the highest levels of identical text (A one-for-one copying of someone else’s text that is passed off as your own), minor changes (Content with minor alterations to the source material, such as altering a verb within a sentence (e.g., slow to slowly), and paraphrasing (Putting someone else’s idea into your own words without crediting the original source) across all 26 subjects. 

Key findings from the analysis include:

• 59.7% of GPT-3.5 outputs contained some form of plagiarized content. 45.7% of all outputs contained identical text, 27.4% contained minor changes,  and 46.5% had paraphrased text. This also highlights that GPT-3.5 isn’t manufacturing “brand new” text; most of the content hails from a previous source, raising issues around plagiarism, copyright, and intellectual property.

   

• The individual GPT-3.5 output with the highest percentage of plagiarism was in Physics, where 27.0% of the text was identical. This was followed by an individual Chemistry output where 24.7% of the text was identical.

   

• The analysis also examined Similarity Scores. The Similarity Score is a Copyleaks-specific scoring method aggregating the rate of identical text, minor changes, paraphrased text, and more. A score of 0% signifies that all of the content is original, whereas a score of 100% means that none of the content is original.
   

• The subject with the highest average Similarity Score was Physics at 31.3%, followed closely by Psychology at 27.7% and Science at 26.7%. The subjects with the lowest average Similarity Score were Theater at 0.9%, Humanities at 2.8%, and English Language at 5.4%.

   

The insights provided by the analysis can help educational institutions and organizations put emphasis on certain subjects when checking for plagiarism, allowing them to tailor their approach as needed to ensure all potential risks and concerns are addressed,” said Alon Yamin, CEO and Co-founder of Copyleaks. “For example, Physics, Chemistry, Mathematics, and Psychology might require a more in-depth look to identify plagiarized text, while other subjects, including Theater and Humanities, may require less scrutiny.”

Yamin added: “Furthermore, the data underscores the need for organizations to adopt solutions that detect the presence of AI-generated content and provide the necessary transparency surrounding potential plagiarism within the AI content. Full-spectrum protection that includes AI and plagiarism detection ensures compliance with copyright and licensing and empowers authenticity and originality within all content.”