Managing the high volume of misinformation

William Meisel: Author, Computer Intelligence: With Us or Against Us?

The Web and social media were once praised for allowing the wide expression of opinions and friendly communication. Today, these resources are being exploited by actors with destructive agendas. The campaign by Russia to impact US elections through misleading web sites and social media posts is a major example of such misuse. Research in 2019 at the University of Oxford found evidence of “social media manipulation campaigns” by governments or political parties in 70 countries. China and Russia were the most active, but the report indicated that India, Iran, Pakistan, Saudi Arabia, and Venezuela have used Facebook and Twitter “to influence global audiences.”

In other cases, disinformation is simply designed for financial gain. Sensational headlines with fabricated news draw individuals to the web sites hosting the articles, generating advertising revenues. Web sites with financial motivation also draw viewers with violent videos and other objectionable material.

Artificial Intelligence (AI) technology can reduce misuse of the Web and social media. Humans reviewing postings at services such as Facebook, YouTube, Twitter, and Google are overwhelmed by the size of the task. Further, the delay in humans evaluating and making a decision on a posting allows viral dissemination of objectionable material over social media before the posting is taken down. Interference with the 2016 election included overwhelming numbers of false or divisive postings, some automated and some by hundreds of individuals hired to create such postings as a full-time job.

In addition, human evaluation is necessarily subjective. Two people may disagree on the acceptability of a potentially objectionable posting, particularly if it could be considered “free speech.” Humans can also be accused of making decisions based on personal biases.

AI can help address the limitations of review by humans. The major AI technique of machine learning using deep neural networks (“deep learning”) is a statistical technique that summarizes the implications of large data sets. In the case of identifying objectionable postings, the data used to train the machine learning is the classification by human reviewers of hundreds of thousands—if not millions—of postings. Deep learning analyzes the data and develops a model that in effect predicts the most likely decision of a human reviewer on a posting, along with the confidence in that prediction. The resulting neural network in effect measures the net conclusion of potentially conflicting reviews of similar postings and indicates the degree of conflict in those reviews.

Deep learning models have been criticized for it being difficult to understand how they make a decision. This objection to deep learning is perhaps overstated, in that humans also have difficulty explaining subjective decisions. A deep learning model can explain a decision to some degree by providing examples of similar human-labelled postings that led the software to come to its conclusions. In a difficult case, a human reviewer of an AI-based decision can see how similar postings were handled by other reviewers to decide if the computer recommendation seems consistent with past practice. This reduces the arbitrariness of a single human’s review.

The advantage of an automated technique is that it can review every posting quickly and block it almost instantly. If the confidence is high, no further review is needed. If confidence is below some threshold, the posting can be reviewed by human reviewers and potentially re-posted.

Another advantage of such an approach is that it is more objective than the review of a single human since it is based on multiple reviews of similar postings. The automated conclusion is defensible as being consistent and less biased than an individual human review.

Companies such as Facebook’s YouTube have employed such techniques. In the Christchurch massacre, the attacker streamed video of his attack and its horrific consequences. Segments were quickly reposted on YouTube. YouTube tried to remove all Christchurch postings, using machine learning to find potential segments for human evaluation. Humans couldn’t keep up, so YouTube eventually blocked all postings detected as candidates by machine learning, providing time for humans to review them and preventing their spread during the review.

Social media and web sites that have provided so much value in the past are under attack. Deep learning is an AI tool that can help us preserve those benefits efficiently and objectively.

William Meisel, PhD, is the author of the recently published book, Computer Intelligence: With Us or Against Us? Dr. Meisel is a long-time researcher and analyst in the area of AI, with an early book on computer pattern recognition while a professor at USC and a decade as the founder and president of a company developing automated speech recognition. He is currently president of consulting organization, TMA Associates. (Contact info@tmaa.com.)