The surveillance advertising economy has a dirty secret hiding in plain sight: it is breathtakingly bad at its stated job. A user documents a litany of personalization failures — ads for products already purchased, language settings ignored despite explicit account data, insurance offers for policies already held with the same bank — that collectively paint a picture not of omniscient corporate surveillance but of an extraction machine so focused on hoarding data that it never bothered to build competent inference. The examples are specific and damning. Spotify and YouTube Music, armed with years of skip-pattern data, still cannot detect that this user skips every electronic track with vocals. Google Chrome, logged into a Google account with English locale settings on every layer, routes the user to a Finnish-language landing page. A bank sends car insurance offers to a customer who already holds car insurance with that bank. An airline with a passport scan on file sends visa information for the wrong citizenship. These are not edge cases requiring exotic machine learning — they are failures of basic conditional logic. The ad-targeting failures are almost comic in their specificity. A single click on an industrial 3D-printing resin ad — a product category the user has zero history with — floods subsequent browsing with 3D-printer content. A one-time search for a specific oven installation manual produces a homepage of oven recommendations. The algorithm's model of human behavior appears to be: whatever you last touched, you must want more of forever. This is not personalization. It is a parrot with a database. The structural diagnosis here is that the adtech stack was never optimized for user utility — it was optimized for impression volume and data liquidity. The incentive is to collect as much signal as possible and sell access to that signal to advertisers, not to actually close the loop on whether the signal produces useful targeting. Advertisers pay for reach and clicks, not for relevance. The user is the product, but the product doesn't even need to work well for the buyer. Language-handling failures reveal something even more fundamental: major platforms treat locale as a geography lookup rather than a user attribute. Moving to Germany temporarily overrides years of English-language preference data. PlayStation's store oscillates between Finnish and English within a single session. These are not ML problems — they are basic software engineering choices that prioritize server-side defaults over client-side identity. The data exists. It is simply not used. Netflix asking a 15-year single-user account to select a profile on every launch is the purest distillation of the problem. The platform has the data. The inference is trivial. The feature exists to serve Netflix's multi-profile architecture, not the user's experience. Across every example, the pattern holds: systems are built to serve the platform's operational model, and user data is a byproduct that gets warehoused but rarely deployed for the user's benefit. The uncomfortable conclusion is that the privacy debate and the utility debate are two sides of the same coin. Users surrender enormous amounts of personal data under the implicit promise that it will at least make their digital experience better. It doesn't. The data is collected because it can be monetized in aggregate, not because anyone has built the engineering discipline to act on it at the individual level. The surveillance works. The intelligence doesn't.
