The OpenAI Foundation, founded by Elon Musk in 2015, recently released the latest model of its GPT-2 text generator, which it states can be dangerous in the wrong hands.
Of course, the new model is barely more convincing than its predecessor, to be able to easily fool a person.
The foundation decided to release the first part of the model in February, as part of a phased process. The reason for the delay in releasing the complete model was security concerns , as its creators believe it could be used for malicious purposes by hackers or terrorists .
As OpenAI admits, GPT-2 could be used to create misleading articles, online impersonation, automate the production of violent or fake content for social media, and automate the creation of spam and phishing.
Until recently, this model was the pinnacle of text generation, until Google created BERT ,which can predict user searches to bring more relevant results.
According to OpenAI, people find the text production from the new 1.5 billion-parameter GPT-2 model “convincing,” but only slightly more so than the 774 million model released in August.
However, OpenAI says GPT-2 can be used very convincingly for bad use. CTEC estimates that GPT-2 can be modified to create propaganda supporting white supremacy, Marxism, Islamism, or anarchy.
In a letter accompanying the release of the new GPT-2 model, OpenAI says it believes there are malicious actors out there who have the resources and incentives to use it to their advantage.
However, the foundation believes that the risk from lower-level threats, such as financially motivated cybercriminals, is not as immediate. There is also no evidence that GPT-2 has been used so far.
OpenAI has developed a detection model that has a success rate of around 95% for text detection. While that rate is high, it argues that systems need to be strengthened with methods that rely on metadata, human judgment, and public education.
