What the Algorithm Cannot Taste

There is a distinction that the language of data systematically erases. It is the distinction between measuring a thing and knowing it. A spectrometer can analyse the chemical composition of a ripe mango with extraordinary precision, cataloguing its sugars, acids, and volatile compounds down to the parts per billion. What it cannot do is tell you whether the mango is good. That judgment belongs to a tongue, to a memory, to a person standing in a kitchen with the fruit in their hand. The spectrometer produces information. The person produces knowledge of a different kind, embodied, contextual, and stubbornly resistant to being entered into a spreadsheet.

The ideology now dominant in Silicon Valley holds that this resistance is a temporary engineering problem. Given enough data, enough processing power, and enough time, every human preference can be modelled, predicted, and replaced by a more reliable computational proxy. Call it what it is: the belief that subjectivity is a bug, and that the correct version of human experience is one from which the human has been quietly removed. Smart kitchens learn your preferences and cook without you. Recommendation engines decide what you want to read before you have finished wanting it. Each of these is a genuine achievement. Each of them is also a step in the same direction: the progressive substitution of lived judgment with computed output.

The substitution is not neutral. When a platform's algorithm decides which music reaches listeners, it does not merely reflect existing taste. It shapes future taste by amplifying what already performs well and starving what does not yet have an audience. The metric becomes the standard. Virality becomes the definition of quality. This is not an accident of design. It is the logical consequence of treating preference as data: once you can measure something, you can optimise for it, and once you optimise for it, you have changed what it means. The map does not merely represent the territory. In this case, the map redesigns the territory to match itself.

The counterargument should be taken seriously. Computation has extended human capability in ways older methods of judgment could not. A radiologist who examines a thousand scans a year cannot develop the pattern recognition of an algorithm trained on ten million. A music streaming service that knows a listener's history can surface recordings a human curator would never have thought to recommend. These are real gains, and dismissing them as mere efficiency is a form of nostalgia that helps nobody. The question is not whether computation is useful. It is whether usefulness is the only thing that matters, and whether what computation cannot capture is worth preserving.

What computation cannot capture is the experience of encountering something for the first time without having been pre-sorted toward it. The discovery that happens when a reader picks up a novel a friend thrust into their hands, not because an algorithm identified it as compatible with their reading history, but because the friend believed it would change something. These experiences are not merely pleasant. They are the mechanism by which taste develops into something more than preference, into judgment, which is the capacity to evaluate rather than merely respond.

The smart kitchen is a useful lens. An oven that learns to cook a dish to the specifications its owner has historically enjoyed will produce a reliable result. What it will not produce is the meal that departs from those specifications in a way the owner had not known they wanted. It cannot cook towards a preference that does not yet exist. This is not a failure of processing power. It is a structural limitation of optimisation itself: to optimise for a target, you must already know what the target is. Every genuinely new experience is, by definition, outside the training data. The algorithm can reproduce the past with uncanny accuracy. It has no access to the future that has not yet been tasted.