No. 115/2023

29 HUMBOLDT KOSMOS 115/2023 analysis to discover how it came about,” says Roberts. But everything turned out differently. Molly Roberts’ supervisor at Harvard was Gary King, a world-leading specialist in quantitative methods. “Gary wrote to tell me and another graduate student at Harvard, Jennifer Pan, that he had found all these Chinese blogs and there were far too many of them to read them all,” Roberts explains in a video chat from her apartment in California, where she lives with her husband and children. “He wanted to know whether I would be interested in finding out how you could use artificial intelligence to discover a way of structuring this mass of data.” CRITICISM DISAPPEARS FROM THE WEB The blog posts King had downloaded from the Chinese Internet were all about workers’ protests. Roberts was supposed to ascertain whether the texts were positive or negative. To do so, she initially had to train the artificial intelligence. “The first thing we did was to assign numbers to frequently used terms,” Roberts explains. “Then we read the texts, evaluated them and labelled them: here someone is writing something positive about the protest, there it’s being portrayed negatively.” On the basis of these examples, the computer learned to interpret words and word contexts according to the instructions. “The fascinating thing about AI is that it can recognise patterns on a deeper level that we humans often can’t access.” In this respect, according to Roberts, the machine is ahead of the human brain. Six months into her work, Roberts and her collaborators noticed something. Because the evaluation of one of the blog posts was not completely clear, she typed in the URL to search for additional information on the original page. But the post wasn’t there anymore. “Sorry, this entry is no longer available” appeared on her screen. The researcher started checking other URLs and discovered that especially the posts that had been positive about the protests had disappeared. It soon became clear that the Chinese Propaganda Ministry must have censored the posts. Roberts realised what she had hit on: “Without searching for it, we had found a mechanism for measuring Internet censorship,” she says, still sounding really excited when she talks about it. Roberts wrote a programme that regularly pinged the pages to see whether they were still online. No easy task. If the computer pings too often, at some stage the sender, i.e., the computer, will be blocked. “I first had to learn quite a lot about programming,” says Roberts, laughing. And she AI RECOGNISES PATTERNS IN TEXTS THAT HUMANS OFTEN DON’T DISCOVER. Photo: picture alliance/AP Images › managed that, too. But it also meant that she had to decide what she wanted to work on in the future: international trade relations or censorship? LOCAL PROTESTS ARE THE MOST PROVOCATIVE “I woke up in the morning and the first thing I thought about was censorship,” says Roberts. That’s when she knew she had to change her plans.” In 2013, together with Gary King and Jennifer Pan, she published a paper in the American Political Science Review, describing how the Chinese government, whilst allowing criticism, nonetheless severely restricted the population’s collective expression. She completed her doctorate on the subject a year later. “It’s reports about local protests that are subject to most censorship in China,” says Roberts, summarising her main findings at the time: when people go onto the streets to protest about expropriation, for example, or police brutality. Roberts was able to follow live how posts on such topics disappeared from blogs and forums. “We often think censorship functions via punishment, when people are imprisoned, for instance, because of something that they have written,” she says. That is certainly true, she states, not just in China but in other authoritarian states like Russia and Iran, too. But in all these countries you can also observe what a major role access to information plays. In Russia, the government purposely spreads misinformation to cover up what is really happening in the “