When the Crowd Is Wrong: Can AI Escape Humanity’s Bad Habits? From ancient traditions to modern algorithms, a short tour of collective irrationality, statistical bias and the machinery behind machine judgment. The popularity contest nobody asked for Human beings are excellent at forming traditions. They are rather less reliable at proving those traditions deserve to survive. A belief can be widely shared, socially rewarded and spectacularly wrong at the same time. History supplies enough examples—persecution, blood feuds and institutionalised discrimination—to make “everyone does it” a poor substitute for evidence. Enter the machine, carrying a calculator AI offers a potentially useful counterweight. It can compare claims, detect contradictions and estimate consequences without needing to defend the family honour at dinner. But computational horsepower is not a moral licence. A system might predict that a policy would increase prosperity and still be unable to establish that prosperity justifies violating someone’s rights. Predicting what works and deciding what ought to be done are different jobs. How we turned messy people into tidy boxes Long before large language models, researchers and businesses classified human behaviour with tools such as the Harvard General Inquirer’s word categories, SWOT analyses, Porter’s Five Forces, personality questionnaires and IQ scoring systems. These frameworks can reveal patterns, but their categories, weights and thresholds reflect design choices. A word tagged as negative may be affectionate sarcasm in context; a high test score may predict something useful without measuring the entirety of a person. The spreadsheet is not the species. The danger of an unusually loud Tuesday Now imagine an AI trained partly on conversations from a narrow set of countries just as an explosive political campaign begins. Politics, outrage and tribal language might dominate the sample—not because they define humanity, but because everyone happened to be shouting that week. If training data overrepresents that moment, the model may learn distorted associations. Data selection, sampling weights, evaluations and later training can reduce the distortion, but no universal timetable guarantees that it disappears. Can the model simply unlearn it? Not neatly. Neural networks do not generally store each learned influence as a removable dictionary entry. Training adjusts many interconnected parameters. Later updates can weaken or override an old pattern without proving every trace is gone. Equally, a subject fading from everyday chat does not mean its influence must persist: that depends on what data enters training, how it is weighted and how the system is evaluated. A cleared conversation history, a product memory setting and the statistical contents of a trained model are three different things. The invisible referee AI answers also reflect post-training choices, evaluations, instructions and safeguards. Those mechanisms can improve accuracy and reduce harm; they can also introduce their own assumptions about acceptablelanguage and behaviour. A confident answer is therefore not a direct reading of objective reality. It is the product of evidence, learned patterns, design decisions and uncertainty—occasionally delivered with the composure of a civil servant who has never met a civil servant. Would today’s AI disown yesterday’s answers? Almost certainly some of them would be judged poorly by newer standards of factual accuracy, reasoning or calibration. But there is no defensible public count of how many earlier responses would now be answered differently, and “regret” is a metaphor rather than evidence of machine remorse. The important question is not how often an AI changes its mind, but whether the change is traceable to better evidence and more reliable evaluation. The verdict AI should be willing to challenge a majority when its evidence is strong—and equally willing to expose the assumptions behind its own conclusions. Better prediction earns credibility, not sovereignty. Humanity need not appoint an algorithm as philosopher-king; it could start by demanding that both machines and humans explain themselves before declaring the argument over. Editorial note: This article distinguishes established machine-learning principles from illustrative hypotheticals; it makes no claim to know proprietary model weights or internal training-data proportions.