Personalization
Personalization is the practice of adapting what a system shows or does to the individual user, based on a model of that person's preferences, behaviour and context. It governs the order of a news feed, the products on a store's home page, the songs queued next, the ads displayed, and increasingly the tone and content of an AI assistant's replies. The core move is the same everywhere: gather signals about a user, infer what that user is likely to want, and use the inference to choose among many possible outputs.
User modeling is the foundation, and it draws on two broad families of signal. Content-based signals describe the items and the user's relationship to their attributes: someone who reads many articles about astronomy is shown more astronomy. Collaborative signals ignore item content and use patterns across people: users whose histories overlap with yours liked something you have not seen. Amazon's item-to-item collaborative filtering, described by Greg Linden, Brent Smith and Jeremy York in 2003, scaled this idea to a huge catalogue by comparing items rather than users. Real systems add context such as time, device and location, and explicit signals (ratings, follows, settings) alongside implicit ones (clicks, dwell time, skips, purchases), which are far more plentiful but noisier.
Personalization works by exploiting regularities, and it has a built-in feedback loop. What a user is shown determines what the user can click, and those clicks become the training data for the next round. A system can therefore reinforce its own early guesses, narrowing exposure over time. Eli Pariser popularised the worry in his 2011 book The Filter Bubble, arguing that individually tailored feeds could isolate people from differing views. The empirical picture is more mixed than the slogan: studies of social media exposure, including a 2015 analysis by Facebook researchers published in Science, found that people's own choices of friends and clicks limited exposure to cross-cutting content alongside the ranking algorithm, and findings vary by platform and topic. The better-supported concern is narrower: optimising purely for predicted engagement favours what is clickable over what is valuable, and systems need deliberate diversity and exploration to counter it.
The cold-start problem is the practical limit on personalization. A new user, or a new item, has no history, so collaborative methods have nothing to compare. Systems fall back on popularity, ask for explicit preferences during onboarding, or use content features and context until data accumulates. A related pitfall is mistaking a short-term signal for a stable preference: a user who buys a one-off gift may be pursued by that category for weeks.
Privacy is the central cost. Behavioural data can reveal health conditions, sexual orientation, finances or political leaning that the user never stated. Anonymisation is hard: in 2008 Arvind Narayanan and Vitaly Shmatikov showed that supposedly anonymous Netflix Prize ratings could be re-identified by matching them with public movie reviews, and a planned follow-up contest was cancelled in 2010 after privacy complaints and a lawsuit. Regulation reflects the tension. The EU's General Data Protection Regulation, in force since 2018, defines profiling, requires a legal basis for processing, grants rights of access and objection, and limits solely automated decisions with significant effects. Technical responses include on-device models, federated learning and differential privacy, each of which trades some accuracy or complexity for less centralised data.
Common misconceptions: personalization is not merely a convenience, and it is not mind-reading. It also shapes prices, ad targeting and what information is surfaced, so it can produce unequal experiences that are hard for users to detect because each sees only their own version. And a model predicts behaviour, not wishes; people often click on things they would prefer not to want. Giving users controls, such as the ability to see why something was recommended, reset a profile or switch to a non-personalized view, is a recognised mitigation. The entry on recommendation engines covers the algorithms that do much of the work, and the entry on AI assistants covers personalization through memory and instructions.
Read personalization as a trade of data for fit: each additional signal can make a system more useful and more intrusive at once, so its quality is judged not only by how well it predicts, but by what it leaves out and what the user can control.