Algorithms
The invisible structure of modern life
Description
The word algorithm carries a strange weight in 2026. A few decades ago it was mathematical jargon the kind of term that lived in textbooks. Today it is a noun the general public uses, often with suspicion, almost always with imprecision. People say the algorithm fed them a video, rejected their loan, decided what news they saw. The word stands in for an entire layer of modern infrastructure a layer most of us cannot see. To decode what algorithms actually are is to recover a literacy the last twenty years of consumer software have quietly eroded.
Algorithms are not new and not mysterious. Long division is an algorithm. So is the procedure a librarian uses to shelve books, the rule a postal sorter follows for bins, the steps a chef walks through when poaching an egg. An algorithm is a finite sequence of operations that produces a result. What changed in fifty years is not the concept but the substrate. Algorithms used to run inside human heads, on paper, or in mechanical contraptions. They now run on machines that operate at billions of cycles per second, on data sets larger than any person could read in a lifetime.
That shift in scale is what we want to look at. We are not going to walk through pseudocode. We are going to follow the trajectory of an old idea as it moved from arithmetic into commerce, from commerce into recommendation, from recommendation into decisions about people. The goal is not to indict algorithms or defend them, but to see them clearly.
The question we're asking: what are algorithms, how did they come to mediate so much of daily life, and what does that mediation cost?
What we'll see: the long lineage of the idea, the recommender turn, the move into consequential decisions, and the fight over accountability.
Table of contents
01What an algorithm actually is
The word comes from a ninth-century Persian mathematician, Muhammad ibn Musa al-Khwarizmi, who worked at the House of Wisdom in Baghdad. His treatise on Indian numerals introduced the decimal system to the Arabic-speaking world, and through Latin translations, to Europe. His Latinised name, Algoritmi, became attached to the procedures his book described. For most of the next thousand years, the term referred to step-by-step methods for arithmetic. Euclid's procedure for finding the greatest common divisor, written around 300 BCE, is still taught as the canonical example.
The modern sense arrived with Alan Turing's 1936 paper on computable numbers, which formalised the notion of a procedure a machine could follow without human judgment at any step. Once electronic computers existed, algorithm design became a mature discipline. Sorting a list, finding the shortest path through a graph, matching patterns in text each problem accumulated a library of competing procedures, with trade-offs in speed, memory, and accuracy.
02The recommender turn
The next move was from search to recommendation. Search assumes a user with a question; recommendation assumes a user with no specific question and tries to figure out what they would want if they thought to ask. Amazon was an early mover. By the late 1990s its product pages featured a strip showing items frequently bought by customers who bought this one a feature built on collaborative filtering, predicting preferences from users with similar histories. Collaborative filtering needs to know nothing about the products. It only needs to know who bought what.
Netflix turned the technique into a public spectacle. From 2006 to 2009 it ran a competition offering one million dollars to any team that could improve its recommendation accuracy by ten percent. The competition produced genuine progress on the underlying mathematics matrix factorisation methods that could find latent dimensions in user behaviour without anyone having to label them. By the time the prize was awarded, Netflix had shifted strategy. What mattered was not which movies you would rate highly but which ones would keep you watching. The model that wins is the one optimised for the metric that pays the bills.
03When algorithms decide about people
Recommendation is consequential at scale but rarely decisive in any single instance. The harder cases involve algorithms that make decisions about a specific person whether they get a loan, whether their CV passes a first screen, whether they are flagged at a border, whether a judge sees them as a flight risk. These applications are grouped under algorithmic decision-making, and they have spread across institutions over the past fifteen years. The appeal is straightforward. A model trained on historical outcomes is faster than a human reviewer, cheaper at scale, and theoretically free of the inconsistency that plagues human judgment.
Credit scoring was an early adopter and remains the cleanest illustration. FICO scores have been used in American consumer lending since the late 1980s, and the model that produces them is an algorithm. Lenders moved from credit scores to broader machine learning models in the 2010s, often trained on alternative data social media activity, smartphone metadata, online behaviour. The risk was that the same signals could amplify the patterns of historical exclusion the traditional system had produced. A model trained on past lending decisions inherits the biases of those decisions, even when input variables look neutral.
04The accountability question
The bias debate forced a question algorithm design had not been asked. When a procedure makes a consequential decision about a person, who is responsible, and what does the person on the receiving end have a right to know? The traditional answer was that vendors owned their code as a trade secret and users got whatever the contract specified. That worked when the software was a spreadsheet. It worked less well when the software was deciding whether someone got parole. The first regulatory response treated algorithms as black boxes to be audited by independent parties testing for disparate impact without seeing the source code.
The European Union moved fastest. The General Data Protection Regulation, in force since 2018, includes provisions some legal scholars read as creating a right to explanation when automated decisions significantly affect a person. The AI Act, finalised in 2024, classifies AI systems by risk level and imposes transparency, documentation, and human oversight requirements on systems used in employment, credit, education, and law enforcement. Whether these rules will produce real accountability or be absorbed as compliance overhead is an open question. The legal architecture is in place. The institutional capacity to enforce it is still being built.
05Conclusion
Algorithms have always been part of human life. What changed is that they now run on machines that operate faster than any institution can supervise, on data no individual can review, with effects that compound across populations the system was never explicitly told to govern. The story of the past thirty years is the story of a technical concept becoming infrastructure quietly, without anyone announcing the moment it happened. The procedures behind a search result, a loan decision, or a video recommendation are doing work that used to be done by editors, loan officers, and friends with taste. Some of that substitution is gain, some is loss, most is unexamined. The shift was not voted on, and the institutions that might have voted on it did not see it coming until the consumer software was already in everyone's pocket.