English
The current writing ability of artificial intelligence is not enough to write world-famous books, but it is enough to hide the traces of machines and confuse the eyes of ordinary people. Most importantly, AI can mass-produce text, concoct fake news, and quickly farm comments. How to find AI pretending to be human? Researchers at Harvard University and MIT-IBM Watson AI Lab thought of using AI to identify AI.
They developed a new tool called the "Giant Language Model Tester" (GLTR) to identify text generated by artificial intelligence. The discrimination is based on the principle that artificial intelligence uses statistical patterns in text to automatically generate text, regardless of the actual meaning of words and sentences. That is, if the words in an article are predictable, the article is most likely written by artificial intelligence, and GLTR can flag whether the words in the article are predictable.
To develop the thickest shield, test it with the strongest spear. Last week, OpenAI, an AI research organization co-founded by Elon Musk, launched an algorithm that can automatically generate highly realistic passages. As long as massive amounts of data are fed into the algorithm's "learning mode", the algorithm can count word frequencies, combine high-frequency words, and generate new passages. The Harvard team invented GLTR based on Open AI's publicly released code.
GLTR can indicate the most likely word combinations based on word frequency. The most predictable words are green, the less predictable are yellow and red, and the most unpredictable are purple. When the text generated by the algorithm written by Open AI was tested, it was found that most words were predictable; while the text of news and science summaries written by humans was more diverse.
So, what kind of words are the "human pen" that AI cannot imitate? Journalists from The Paper put popular passages from "Hamlet" into this tool to retrieve.
Resume submission
Business cooperation
media inquiries
WeChat QR code
