Fake news is reaching a fever pitch in America—it’s taking center stage as the matériel in an absurd and unlikely Cold War. In almost every news cycle this pernicious term comes up as a talking point. Inevitably debate follows within tech, waxing penitent, about how we might solve fake news. I work in a branch of machine learning known as text analytics. The text analytics space deals in teaching computers to read and draw insights from natural language. Natural language, so we’re clear, is anything and everything written online: Your distant Aunt’s emoji filled Facebook posts, typo ridden hotel reviews on TripAdvisor, angry op-eds on Breitbart. We’re doing pretty well here. The best models can sense complex themes, gradients of emotion, and even intent with a high percentage of precision. So, naturally, we’ve been invited into the praetorium.
The most interesting thing about fake news is its virality. My colleague, Seth Redmore, describes fake news as a “brain virus,” which illustrates the point well. It calls on a notion known as “Memetic Warfare,” detailed in a study from 2010 by the Federation of American Scientists:
Much like a virus moves from body to body, memes move from mind to mind. Just as genes organize themselves into DNA, cells, and chromosomes, so too do replicating elements of culture organize themselves into memes, and co-adaptive meme complexes or “memeplexes.”
Paul Horner, the impresario of the Facebook fake news mill, emphasized the role of memetic power in a 2016 interview with the Washington Post. “[They] just keep passing stuff around,” he said. “Nobody fact-checks anything anymore.” Such is the battle cry of the First Meme War. So, how do we fight back against this barrage of alternative facts and late night tweets?
Is Text Analytics a Solution?
Teaching a computer program to tell you when this person is appreciative or that situation is unfortunate is straightforward. Language bears generally accepted sentiment, what you might call weight. In the sentence, “We have a narrow window to accomplish this,” the words “narrow window” might bear a negative sentiment weight of -1.38. A computer decides this sentiment weight by referencing some giant corpora of sentiment bearing phrases and words. While the science behind this is complicated, it’s relatively straightforward when compared to sussing out credibility. I say “credibility” because fake versus real is a false dilemma. Trying to solve for fake versus real risks alienating credible news that might be breaking or outside the mainstream. Further, if you’re judging based on centrality to the bulk message you’re going to end up penalizing divergent conversations—which are often very interesting and truthful. For example, maybe a group of academics are mad about curtailed academic freedom but nobody else cares.
So let’s focus on credibility for now. Credibility carries with it the burden of context. Teaching computers context is notoriously difficult. So, let’s eliminate what we can’t do. Computer intelligence is narrow intelligence, and it can’t handle complex social problems. But computers can follow rules. A team could build out a comprehensive tagged dataset, something the computer references in real time whenever an article is published on the internet. If n properties of the message, its spread, and its variations all match cases where fake news previously presented then the article is flagged for further review. This isn’t so dissimilar to what Facebook is proposing as a solution right now, only baking in text analytics makes it more efficient.
But, is it worth it? I don’t think so. Text analytics and machine learning constitute an exciting and vibrant field. There’s ample utility for applications like sentiment analysis, predictive analytics, and NLP. Regardless, computers aren’t even fractionally as intelligent as infants. Yet, we’re being propositioned to outsource basic civics and common sense to a machine learning model. This is a stark example of using technology to treat a symptom rather than an illness. The solution doesn’t end with technology; the solution begins with education. The First Meme War isn’t going to be won with weapons like text analytics. It’s going to be won by soft power. Emphasizing subjects like Government and Logic and History in high school will far outpace any supplemental technology. Take it from a text analytics expert—machine learning is nothing when compared to the power of human learning.