Sberbank: GigaCheck

2024: GigaCheck Presentation

Sber and SberDevices on June 27, 2024 introduced the GigaCheck AI detector to determine texts written by the neural network.

GigaCheck allows you to establish with high accuracy who wrote the text - a person or an artificial intelligence system. Over the past month, 6% of the 220,000 articles on news resources that analyzed ours have been written by generative neural networks.

The solution is highly accurate and fast, provided via API and embedded in any third-party application.

The key metric TPR&FPR (True positive rate & False positive rate) = 0.01-94.7% means that with a false positive rate of no more than 1%, the accuracy of GigaCheck is 94.7%. The indicator of the closest competitor is no more than 85%.

As of June 2024, GigaCheck gives a generalized assessment of authorship. If one text uses both fragments written by man and fragments generated by artificial intelligence, then authorship is determined by the predominant content. In the near future, GigaCheck will also be able to determine which text fragments are generated by LLM.

The work of the GigaCheck text detector will be appreciated by experts involved in verifying the authenticity of scientific publications, dissertations and diplomas. Owners of publics in social networks using the detector will be able to quickly detect comments and posts created by generative models. Editors and everyone who accepts the work of copywriters will also receive an effective tool for checking and evaluating it.

{{quote 'author = said Sergey Markov, Managing Director of SberDevices Experimental Machine Learning Systems Directorate. | With GigaCheck, we analyzed 220,000 texts of news resources over the past month and found that on average 6% of them were created by LLM, and in some large publications more than a third of all published content. Until now, the likelihood of creating such a working text detector was considered extremely low. Many experts regarded this task as practically unsolvable. Developing such a service requires serious resources and large amounts of data. But even unlimited resources may not be enough. For example, one of the recognized leaders in the development of artificial intelligence systems, OpenAI, announced its detector for English-language texts in early 2023, but over time was forced to close access to it due to low accuracy. The most important is the achievement of our team, which, despite all the difficulties, managed to develop its own effective technology,}}

For June 2024, you can test the capabilities of GigaCheck freely available on the site and through Telegramboat-.