OpenAI, an artificial intelligence research and development organization, announced in February that it developed a sophisticated fake news-writing AI, GPT-2, with MIT. At the time, it was pointed out that if such AI is abused, fake news or spammy information may overflow and cause confusion. For this reason, OpenAI only released a simplified version of GPT-2.
However, OpenAI has finally released it on the web, saying that research is still underway and there is no strong evidence that it is misused even if the full-featured full version is released.
GPT-2 is a text generation system with the ability to create text content that matches the front and back with minimal information. It was learned from more than 8 million text documents collected on the website. When writing a news article, he has the skill of completing the article by giving it a title and the first sentence.
GPT-2 can adjust the output sentence. For example, it is possible to make a system that automatically generates a certain discriminatory idea by advocating it. However, OpenAI explains that it has released the full functional version as there is no indication that the limited version released in March was abused in the production of fake news.
The research team is also said to be developing a system that can detect sentences generated by GPT-2 with 95% probability. However, it is said that this is not a fully automatic operation, but some need to borrow human power, such as searching for fake photos or videos.
OpenAI says it will participate in the Responsible Publishing Codes Project for Machine Learning and discuss this with the community while monitoring how GPT-2 will be used in the future. Related information can be found here .