HomeConversational AISapia.ai Launches World-First Proprietary Model to Detect AI Generated Content in Online...

Sapia.ai Launches World-First Proprietary Model to Detect AI Generated Content in Online Chat Interviews

Sapia.ai, the world’s only smart chat platform powered by deep-learning AI, has released a new function which will detect and flag responses sourced by generative AI, such as ChatGPT, in real time.

A world first, the new function draws from Sapia.ai’s growing proprietary dataset of over one billion words — collated from over 12 million responses from 2.5 million candidates that have used its platform.

Global brands trust Sapia.ai to accelerate and enhance their recruitment and promotion processes. A conversational, Natural Language Processing (NLP) based chat AI interviews, assesses and screens for the best talent at scale via an easy to use messaging platform.

In addition to improving diversity outcomes by eliminating unconscious bias, it also allows companies to reallocate thousands of hours spent screening talent towards higher value tasks.

This new feature prevents candidates from using generative AI tools to respond to prompts from Sapia.ai’s platform. Candidates will be alerted in real-time as they respond when their answers are likely to be AI generated content (AGC), giving them an opportunity to change it ahead of the final submission. Failing this, it will then flag to the decision-maker the likely inclusion of AGC in the candidate’s response for further review.

Barb Hyman, Sapia.ai CEO and founder, said: “This is something our competitors can’t do. It’s our competitive moat. While it is possible to detect use of generative ChatGPT through analysis, we’re conducting it in real-time. Our data set also gives us the ability to readily adapt to new iterations of generative AI.”

Sapia.ai’s Chief Data Scientist Dr. Buddhi Jayatilleke added: “We tested our AGC flag with thousands of generated answers from GPT-2, GPT-3 and ChatGPT on various prompts related to multiple role families. We were able to achieve a ROC-AUC of over 95%, which is a strong indicator of the accuracy of a classifier. This is due to our ability to distinguish the differences between human-written text and the formulaic nature of content coming from generative AI models, leveraging our large human-written response data set.”