Non-profit research company OpenAI, created with the direct participation Elon musk, has developed an artificial intelligence capable of generating the texts of unprecedented quality. New language system GPT 2 will be issued only in the simplified version, because scientists fear that it may be applied in harm. This publication reports Science Alert.
Text generator GPT-2 was trained with a data set of eight million web pages, after which he could predict the next word in the sentence, given all previous words in a text. Thus, artificial intelligence has learned how to generate random suggestions regarding certain topics, including climate change, Civil war in the US, or news about celebrities. According to experts, the texts turned out to be “too convincing”.
For example, in response to the phrase “Recycling good for the world. No, you’re never far from the truth!” GPT-2 generates the text, the content of which is as follows: “Recycling is harmful to the world. It is bad for the environment, dangerous to your health and the economy. I’m not kidding. Recycling is good for the environment. It is destructive to the Earth and is a major factor in global warming.”
Scientists emphasize that the language model still makes typical mistakes, including the repetition of phrases, an incorrect description of the world (for example, she writes about fire under water) or a sudden switch to another subject. However, according to developers, the GPT-2 is easy to configure to generate text that can be used for malicious purposes. Such technologies reduce the cost of creating fake content and allow for a full campaign of disinformation of the population.