AI Assistance for Misinformation Helps Experts the Most and Frequent Social Media Users the Least
Everyday, we are bombarded with news and media from many unknown
sources. Many times, a news story we read introduces information that
we cannot directly verify. Can we tell that the information may be
unreliable or biased by simply reading its content? This question is what we
set out to answer in our recent ICWSM paper using an online survey of 654 participants and a random sample of news stories from a wide-range of sources.
News can be unreliable for many reasons that go beyond factual inaccuracy. For example, unreliable news stories may be written in a sensational manner. These stories may make claims without any verifiable evidence or refer to unreliable sources. Some news stories contain faulty logical reasoning. We know from our prior work that many sources publishing unreliable information have a distinctive writing style. Artificial Intelligence (AI) is quite good at distinguishing between the stylistic differences of unreliable and mainstream sources. But, how well do humans perform in this task?
Our experimental study was set up as the following. Participants are randomly given a set of news articles to read and rate based on how reliable they think each article is. These news articles are sample from 2 groups of articles which are pre-labelled as unreliable or reliable by expert raters. We will use the expert rating to assess how well each study participant does at the task. Each participant is only given the article title and article content (no source or meta information is provided). After each article is given a rating between 1 and 10 by the participants, they can optionally provide a free text response as to what factors helped them make their rating decision. In addition to the main task, each participant fills out a demographic-based pre-study, which is used in our analysis.
Our study shows that overall participants can appropriately rate a reliable article as reliable and an unreliable article as unreliable. In fact, the comments from our optional free-text response question show that the participants are picking up on these known stylistic differences: “it doesn’t follow Associated Press style for capitalization”, “not a lot of evidence”, “A lot of negative emotions, Some language seems sensational”.
While participants did fairly well at the task, there were some mistakes made, with an average rating of 6.64 for reliable articles (where a 10 is perfectly reliable) and an average rating of 5.01 for unreliable articles (where a 1 is a perfectly unreliable article). This higher than expected rating for unreliable articles may indicate some uncertainty when judging reliability.
As mentioned earlier, AI has been shown to perform well in predicting reliability through writing style. While this high performance has only been shown in lab settings and it is debatable what this type of system will look like in a real-life environment, can it improve the participants’ performance?
To this end, we run additional tests with two different AI conditions. First, we show participants predictions from a state-of-the-art news article classifier in the form of probabilities (ex. “Our Smart AI says this article has a 95% chance of being reliable”) along with the news article. Second, we again show participants predictions in the form of probabilities, but additionally provide explanation for the predictions using the features used by the classifier (ex. lack of quotes, subjective language, etc.). We ensured that all predictions are correct, hence there is no case where the AI was misleading.
We find that each additional piece of information improves the ratings, showing that individuals follow the advice from AI to better distinguish between reliable and unreliable articles. This is a promising result, showing the potential of using AI to help readers be diligent about low quality sources and hopefully reduce the impact of potential misinformation provided by these sources.
Unfortunately, this increase in performance is not uniform across all the
participants of our survey. When breaking the results down using our pre-study data, we find several interesting results. First, individuals who read news multiple times a day do well in this task to start with and improve significantly with AI assistance. This result could be due to a form of expertise. On the other hand, individuals who trust news from social contacts and share news often on social media do significantly worse at the task. This group did much worse at rating unreliable articles and did not improve very significantly with AI assistance. One potential reason could be that these individuals are exposed to more unreliable information (or information of the same style as unreliable news articles) which had a type of normalizing effect. In addition we found that political leaning of the participant and the article did not have a significant impact on ratings, although this experiment was not designed to assess political leaning in news articles.
Our study is one of the few on this new topic. It provides us with
some positive and some negative results. While this study begins to address the usefulness of AI assistant in the context of news, there are many questions left to explore. How will AI impact decision making if it is presented repeatedly? How will it be perceived if it makes suggestions counter to the individual’s strongly held opinions? How do we adjust the feedback to the user’s level of bandwidth and ability? Who benefits most from seeing explanations and what types of features make sense to who?
Our hope is that this study can be used a stepping stone to developing tools that can provide helpful, transparent, and customizable assistance to news readers; ultimately helping society be better informed even in the presence of misinformation.