New Study Finds DeepSeek-Chat AI Content Highly Detectable—Raising Questions About Its Origins

A new study conducted by Originality.ai has found that text generated by DeepSeek-Chat is 99.3% detectable using the company’s AI content detection models.

Unlike previous LLM releases, DeepSeek-Chat did not trigger the usual drop in detection accuracy upon its release. This suggests the potential for it to be a distilled version of an existing LLM.”

— Jonathan Gillham

COLLINGWOOD, ONTARIO, CANADA, January 30, 2025 /EINPresswire.com/ -- A new study by Originality.ai has found that text generated by DeepSeek-Chat, a recently released large language model (LLM), is 99.3% detectable using the company’s advanced AI content detection models.

This suggests that DeepSeek-Chat’s output is

highly distinguishable from human-written text, reinforcing the effectiveness of AI detection technology.

The study analyzed 150 DeepSeek-Chat-generated text samples, comparing their detectability across multiple AI content detection tools.

Key Findings:

Originality.ai’s models achieved 99.3% accuracy in detecting DeepSeek-Chat content, outperforming competitors

- GPTZero: 97.3% accuracy

- RapidAPI’s AI Content Detector: 80.7% accuracy

DeepSeek-Chat could possibly be a distilled version of OpenAI’s LLMs.

Originality.ai rigorously evaluates newly released LLMs against its AI detection models to measure efficacy. Typically, new models initially lower detection accuracy before engineers retrain the system to close the gaps in order to restore peak detection performance.

However, DeepSeek-Chat did not cause this expected drop in accuracy. This unusual result has led researchers at Originality.ai to believe that DeepSeek-Chat could possibly be a distilled version of OpenAI’s ChatGPT or another existing LLM.

Adding weight to this theory, Bloomberg and the BBC have reported that OpenAI and Microsoft are investigating whether OpenAI’s technology was used or obtained in an unauthorized manner in relation to DeepSeek.