The Invisible AI Around You

The AI systems that get the most attention — ChatGPT, self-driving cars, AlphaFold — are visible and discussed. But you interact with AI systems dozens or hundreds of times every day through interfaces that give no indication that a model is running. These invisible systems collectively shape what information you see, what you buy, how long you spend on platforms, and what routes you drive. Understanding them is not optional for an informed person in this era.

Recommendation Systems

A recommendation system's job is to predict which items from a catalog a specific user is most likely to engage with — click, watch, purchase, like — and surface those items prominently. Every major content platform runs recommendation systems: YouTube, Netflix, Spotify, TikTok, Amazon, Instagram. The dominant technique is collaborative filtering, which exploits the pattern that users with similar past behavior tend to have similar future preferences. In its matrix factorization form, you represent each user and each item as a vector of latent factors (typically 50-300 dimensions), and train these vectors so that the dot product of a user vector and an item vector predicts the user's rating or engagement with that item. A user's latent vector captures their implicit preferences; an item's vector captures its characteristics — not as human-defined features, but as patterns the model learns from the data. Modern recommendation systems layer additional signals onto collaborative filtering: content-based features (what is this item about?), recency, diversity (avoiding showing the same thing repeatedly), and context (what time is it, what device, what did the user just watch?). Deep neural networks trained on billions of implicit feedback signals (clicks, watch time, skips) now dominate at scale. The societal effects of recommendation systems have been extensively studied and debated. On one hand, they enable genuine discovery — exposing users to music, films, or products they genuinely value but would not have found unaided. On the other hand, the optimization target (typically engagement or watch time) is not the same as user wellbeing or accurate information. A system optimizing for watch time may surface increasingly extreme or emotionally arousing content because such content holds attention. Whether this constitutes algorithmic radicalization, or whether the evidence for that causal claim is strong enough to support the conclusion, is an active area of social science research where conclusions are genuinely contested.

Search ranking is a closely related problem. A search engine's job is to return the most relevant documents for a query — but relevance is multi-dimensional: factual accuracy, query intent match, freshness, authority, and absence of manipulation. Google's core ranking system (PageRank, introduced in 1998) used the link structure of the web as a proxy for authority: a page linked to by many other pages was inferred to be important. Modern search ranking layers neural ranking models (such as Google's BERT-based MUM system) on top of signals like PageRank, incorporating semantic understanding of query intent. Spam filtering is one of AI's oldest production applications. Rule-based filters gave way in the early 2000s to naive Bayes classifiers trained on labeled spam and ham (legitimate mail). Modern spam filters combine multiple signals — sender reputation, content analysis using language models, link analysis, header inspection, and user feedback — and must be retrained continuously because spammers actively adapt to evade detection. This is a live example of an adversarial setting: the filter and the spammer are locked in an ongoing arms race. Navigation: when you ask Google Maps for a route, a shortest-path algorithm (Dijkstra's or A*) searches the road network graph, but edge weights — travel times — are estimated from real-time and historical data using ML models. Traffic prediction systems aggregate GPS signals from millions of devices, identify congestion patterns, and update predicted travel times continuously. The system routes millions of drivers simultaneously; if enough drivers follow the same recommended route, that route becomes congested, changing the optimal route for subsequent drivers. This feedback loop is a coordination problem that navigation AI is still not fully solving.