双语科普故事 | AI工具预测网络健康谣言是否会对现实世界造成伤害

The flood of misinformation online inevitably produces adverse consequences in key measures of public health—and death from COVID among unvaccinated people stands out as perhaps the most prominent example. That cause-and-effect relationship—that scrolling through endless postings about hydroxyquinoline, ivermectin and vaccine conspiracies can lead people astray—seems more than obvious. But it isn't straightforward to determine scientifically.

网络上充斥着大量的误导信息，不可避免地对公共卫生的关键指标产生了负面影响——其中，未接种疫苗人群因新冠死亡是一个尤为突出的例子。这种因果关系，即人们不断浏览关于羟氯喹、伊维菌素以及疫苗阴谋论的大量帖子后很容易被误导，是显而易见的。但从科学角度去确切证实这一因果关系却并非易事。

Clear linkages between misinformation and adverse consequences have proved very difficult to find—partly because of the complexity of analyzing the workings of a public health system and partly because most social media companies do not usually allow independent outside parties to analyze their data. One exception to keeping data off-limits is Reddit, a platform that has begun to emerge as a place where, with the company’s blessing, social media research can flourish. Now studies using Reddit posts may be putting scientists closer to finding the misinformation missing link.

事实证明，要找到错误信息与不良后果之间的明确关联极为困难——部分原因在于分析公共卫生系统运作机制的复杂性，部分原因是大多数社交媒体公司通常不允许外部独立机构分析其数据。Reddit是数据不公开的例外。在该公司支持下，该平台已逐渐成为社交媒体研究的热土。如今，利用Reddit帖子开展的研究，或许能让科学家更接近揭示错误信息与不良后果之间的缺失环节。

A new analytical framework that combines elements of social psychology with the computational power of a large language model (LLM) could help bridge the gap between online rhetoric and real-world behaviour. These results were recently posted to the preprint server arXiv.org and presented at the Association of Computing Machinery CHI Conference on Human Factors in Computing Systems in Hawaii this week.

一个融合了社会心理学与大型语言模型计算能力的新框架，有助于缩小网络言论与现实世界行为之间的差距。这些研究结果近期已发布到预印本服务器arXiv.org，并于本周在夏威夷举行的计算机协会人机交互会议上进行了展示。

Eugenia Rho, a computer scientist at Virginia Tech and senior author of the new study, wanted to pin down whether a tie exists between people's behaviour and the type of language encountered on a site such as Reddit.

弗吉尼亚理工大学的计算机科学家、这项新研究的资深作者尤金妮娅・罗，旨在探究人们的行为与在Reddit等网站上接触到的语言类型之间是否存在联系。

Along with her PhD student Xiaohan Ding and their colleagues, Rho began her research by first tracking down thousands of Reddit posts from banned forums opposing vaccines and COVID-19 prevention measures. Next, the team trained an LLM to recognize the “gist” of each post—the message's underlying meaning, as opposed to the literal words it is composed of. “That's sort of the secret sauce here,” says Valerie Reyna, a psychologist at Cornell University and a co-author of the study.

罗与她的博士生丁晓涵（音译）及其他同事，首先从被封禁的论坛中追踪了数千条反对疫苗及新冠预防措施的Reddit帖子，以此开启了研究。接下来，团队训练大型语言模型识别每篇帖子的 “主旨”—— 即信息的潜在含义，而非字面意思。“这可以说是这项研究的关键创新之处，” 康奈尔大学心理学家、该研究的合著者瓦莱丽・雷纳指出。

“Fuzzy-trace theory” suggests that people pay more attention to the implications of a piece of information than to its literal meaning. This helps explain why people are more likely to remember an anecdote about someone getting robbed than a dry statistic about crime rates or why gamblers are more apt to place a bet when folding is framed as possibly losing them money rather than potentially gaining it. “People are more moved by certain kinds of messages than others,” says Reyna, who helped pioneer fuzzy-trace theory in the 1990s.

“模糊痕迹理论”指出，相较于信息的字面意思，人们更关注其隐含意义。这解释了为什么人们更容易记住某人遭抢劫的轶事，而不是枯燥的犯罪率统计数据。同时，这也解释了为什么赌徒在弃牌被描述为可能输钱时，比被描述为可能赢钱时更倾向于下注。“某些类型的信息更触动人心，”20世纪90年代参与开创模糊痕迹理论的雷纳说道。

This careful choice of wording enhances persuasiveness. “Over and over and over, studies show that language in the form of a gist is stickier,” Rho says. Her team's analysis found that in the context of social media, this seems to be especially true for causal gists, or information that implies a direct link between two events. A post might link vaccinations with getting sick in a specific format, using wording that packs a lot of rhetorical punch. For example, one Reddit user posted, “Had my Pfizer jab last [Wednesday] and have felt like death since.” Rho's team found that every time the causal gists in anti-COVID posts grew stronger, COVID hospitalizations and deaths spiked nationwide, even when the Reddit forums were subsequently banned. The researchers pulled their data from nearly 80,000 posts spanning 20 subreddits active between May 2020 and October 2021.

这种对用词的谨慎选择增强了说服力。罗指出：“研究表明，以主旨形式呈现的语言更具吸引力或说服力。” 她的团队分析发现，在社交媒体环境中，对于因果主旨，即暗示两个事件存在直接联系的信息，这种效应尤为明显。帖子可能会采用特定格式，将疫苗接种与生病联系起来，并使用具有强烈感染力的措辞。例如，一位Reddit用户发帖称：“上周三打了辉瑞疫苗，从那以后就感觉自己快不行了。” 研究发现，每当反新冠疫苗帖子中的因果主旨愈发强烈时，全国范围内因新冠住院和死亡的人数就会激增，即便相关Reddit论坛随后被封禁。研究人员从 2020 年 5 月至 2021 年 10 月期间活跃的 20 个子版块的近 8 万个帖子中提取了数据。

But such a framework might not make equally good predictions across the board. “When there is no discernible gist, the approach might be less successful,” says Christopher Wolfe, a cognitive psychologist at Miami University in Ohio, who was not involved in the study. This could be the case for studying the behaviour of people seeking treatment for common health issues, such as breast cancer, or trying to view sporadic ephemeral events, such as auroras.

但这样的框架未必能在所有领域都做出同样准确的预测。“当无法明确把握主旨时，这种方法的预测效果可能会打折扣，”俄亥俄州迈阿密大学的认知心理学家克里斯托弗・沃尔夫（未参与此项研究）表示。例如，在研究人们针对常见健康问题（如乳腺癌）寻求治疗的行为时，或者尝试观察诸如极光这类偶尔发生、转瞬即逝的事件时，这种情况就可能会发生。