Memes are Dividing Us: How Facebook uses AI to combat misinformation, disinformation, and polarizing memes.
In an era where entire industries are suffering from a lack of consumers due to the global COVID-19 pandemic, tech and social media platforms are facing massive influxes of user activity and content. Users on these platforms are debating and receiving information related to the upcoming election, the Supreme Court nomination, climate change, vaccines, and COVID-19, to name a few. This surge of information and users has also led to an increase in hate speech and misinformation, that may directly affect users’ opinions on these topics. The question remains, however, as to who is really regulating the content being placed on these platforms, and how users are being affected by the collection of this mass data.
Content moderation is now a hotly contested issue as it continues to get more exposure. Some companies have taken a more direct approach without regard for the impact that content moderation may have on free speech, while some have been more hesitant to do so. For example, after the heinous murder of George Floyd, there was a national outrage followed by large scale protests, and in some instances, rioting. President Trump’s response to this was sending out a message of “When the looting starts, the shooting starts.” We saw this message get flagged on Twitter for promoting violence, while Facebook declined to take any action. Mark Zuckerberg appeared on Fox News, claiming that he didn’t want his company to be an “arbiter of truth” on political issues.
We cannot ignore the consequences that inadequate content moderation, with regards to misinformation and hate speech, can have on society. While there may be consensus with regards to content moderation for things such as child and animal abuse, there remains a strong divide when it comes to political opinions. We have large portions of society who no longer believe in tangible facts and statistics and believe such information to be “fake news.” We have a complete erosion over political discourse that was highlighted in the recent presidential debate. There is a very fine line when it comes to effective content moderation and the preservation of freedom of speech for the benefit of society. Whether or not private tech companies are best suited to solve this problem and distinction remains to be seen.
Facebook’s Approach and Use of AI in Content Moderation
First, within its Community Standards Enforcement Report, Facebook explains how they use a mix of artificial intelligence (AI) and human moderators to enforce its content policy standards, which aim to regulate Hate Speech, Fake Accounts, Spam, and Bullying and Harassment, among other topics. Facebook has begun to lean more and more on the use of AI for regulation as a direct result of the pandemic and the inability of human workers to physically come into the workplace. Human moderators cannot simply perform their jobs at home because they deal with sensitive and graphic information which requires a more controlled environment for review. In the report, Facebook states how it took action against over 22 million posts concerning hate speech between April and June of this year. Ninety-five percent of those posts were initially flagged by AI.
Using AI to Combat COVID-19 Misinformation
Facebook must also tackle the challenges arising from AI involvement with content moderation of videos, photos, and yes, memes, which are becoming the primary vehicles for spreading misinformation. Facebook claims it has put warning labels on over 50 million posts related to COVID19. The danger related to the spread of false COVID-19 misinformation is very apparent. The United States leads the world in COVID-related deaths with over 211,000 deaths. Misinformation on Facebook may have played a direct role in some of these deaths as users were (and continue to be) fed endless amounts of false truths related to face masks, treatments, severity of the virus, and who is impacted by it. If people were able to only have their views influenced by science and fact, then there is a real possibility that the country could have a significant drop in the spread of this deadly disease.
The company and its engineers speak on these challenges in separate blog posts titled AI advances to better detect hate speech and Using AI to detect COVID-19 misinformation and exploitative content. Facebook explains that AI models have difficulty going through a video or meme because the image may be duplicated and modified slightly while it is spread over the platform. This makes it harder for an AI model to detect whether a photo is an acceptable image or whether it has been morphed into a misleading one. Below is an example from the Facebook blogs.
AI Has a Solution!
Facebook’s answer to this problem is an AI model called SimSearchNet. The AI runs on a database which includes every image uploaded to Instagram and Facebook. It also checks the images against task-specific human-curated databases. The system checks billions of images against these databases available and alleviates some burden from human fact checkers who can focus on the detection of new misinformation rather than identifying near duplicates of a singular image (which could range in the millions).
Using AI to Combat Hate Speech within Memes
While COVID-19 is a challenge for society that promotes the use of content moderation, other issues, such as hate speech, are much trickier to regulate. The danger of moderating hate speech is that it effectively limits an individual’s First Amendment right of freedom of speech. In contrast, the danger of not moderating hate speech is the destruction of public discourse, leading to a more polarized society. There also remains the issue of bias within AI, which may regulate hate speech and misinformation for one group of individuals based on certain parameters, while allowing other speech to get posted without hindrance. Tie these issues with the aforementioned problems of duplication, and the hurdle that AI faces becomes even larger.
The Hateful Memes Challenge: A Possible Solution
A major step that Facebook has taken to combat hate speech is by creating the Hateful Memes Challenge and Data Set. The dataset contains upwards of 10,000 examples of created memes. The company explains that the model processes the image and the text separately, and attempts to understand the unique relationship between the two. This is a complex multimodal understanding which is still being developed. In order to expedite the research in this area, Facebook has released the Hateful Memes dataset to the border research community, in an associated Hateful Memes Challenge with a $100,000 prize pool. Further information of Facebook’s AI work in the area can be found in this research paper.
The Limits of AI
So why is this a problem? It would seem that AI has done a great job of controlling what is put out on the platform. However, it is important to remember that AI has severe limitations. Content moderation as a whole poses many risks for error, and when you speak about automated systems, the risks multiply ten-fold. AI can mislabel true information based on its decision-making while allowing misinformation to squeeze by the cracks. With limited human employees available for review and a larger volume of content as a result of COVID-19, this can lead to larger scale mislabeling.
AI is not an Absolute Fix for Existing Problems
The harmful repercussions of widespread misinformation is something that we have witnessed within society over the past four years. We have neighbors, friends, and family members whom we refuse to speak to because of our differing political opinions. Political and public discourse is on the verge of a complete breakdown and the argument can be made that it has already been eroded within society as a result of online misinformation.
However, I believe that we are not past a point of no return. Facebook’s SimiSearchNet and Hateful Memes Challenge are just some examples of how humans and machines can work efficiently in the pursuit of effective content moderation. They show how humans and AI can work together to limit biases and mislabeling thereby more effectively walking the fine line between content moderation as a tool which benefits society, and not as one that silences constitutional freedoms.
For a deeper dive into other work that has been done by Santa Clara Law’s High Tech Law Institute click here. A link to Professor Chien’s class The Business, Law, Technology, and Policy of Artificial Intelligence can be found here. The class is a wonderful opportunity for any student who wishes to see more clearly the intersections of Artificial Intelligence, Law and Public Policy.
About the author of this post:
Inderjit Pannu is a 2L focusing on Artificial Intelligence and Intellectual Property Law at Santa Clara University School of Law.
