In 2009, Google published a paper in Nature claiming that flu trends could be predicted weeks in advance by analysing search query volumes. The paper was celebrated as a landmark demonstration of big data's predictive power. By 2013, Google Flu Trends was overestimating flu prevalence by a factor of two. By 2015, the project was quietly wound down.
Can the internet predict the future? In a narrow, probabilistic sense: sometimes, partially, for certain types of events. In the broader sense the question usually implies: no. And the gap between those two answers matters enormously.
Prediction systems are very good at telling you what happened yesterday, dressed up as a forecast for tomorrow. The internet doesn't see the future. It sees patterns in the past and extrapolates—confidently, fluently, and with a success rate that is always better than chance and never as good as claimed.
What Predictive Systems Actually Do
Every major platform runs predictive models. Netflix predicts which show you'll watch next. Spotify predicts which song will keep you listening. Amazon predicts what you'll buy. These are pattern matching systems trained on historical data—and they work well because human behaviour is, within certain domains, highly repetitive. What they cannot predict is novelty. They cannot predict a behaviour they haven't seen before. They cannot predict Black Swan events—sudden discontinuities with no precedent in the training data. Google Flu Trends failed partly because it was trained on historical search patterns, and when media coverage of flu became unusually heavy, people searched flu-related terms for reasons unrelated to being ill. The model had no framework for that distinction.
The Trending Topics Illusion
When something "trends" on Twitter or TikTok, it can feel like the internet has identified something important before mainstream media caught up. Sometimes this is true. But these cases exist alongside a far longer list of trending topics that predicted nothing, amplified misinformation, and created false urgency around non-events. Trending is not prediction. It is amplification—surfacing what is generating engagement right now, which is not the same as what is important, true, or predictive of anything beyond the next news cycle.
The honest answer is modest: the internet is very good at telling us what a lot of people are thinking about right now. That is genuinely useful. It is not prediction. And confusing the two—letting the confident outputs of pattern matching systems substitute for careful analysis—is one of the more consequential epistemic errors of our moment.
