The saying “history is written by the victors” endures not because it is perfectly accurate, but because it captures a persistent truth about power and knowledge: those who prevail tend to control which stories are preserved, legitimized and taught. While the phrase is often attributed to figures like Winston Churchill or Hermann Göring, versions of the idea appear much earlier across European intellectual history. As a shorthand, it warns against the uncritical acceptance of dominant narratives, reminding us that what passes for “truth” is often shaped by selective memory, omission and institutional authority. Traditionally, these distortions were the work of states, empires and governments. Today, however, the mechanisms of narrative control have expanded. Platforms, algorithms and data-driven systems increasingly mediate how information is recorded, ranked and reproduced. At an unprecedented scale, they determine not only what is remembered, but what is surfaced, amplified and treated as authoritative. In this context, the old warning takes on new urgency: if history is written by the victors, we must now ask who – or what – counts as a victor in an age of artificial intelligence.
Perhaps the most evident example of institutionalized human development is the education system. Education, as a government-regulated institution in most countries, is rarely politically neutral. Rather than serving as a bipartisan space for critical inquiry, schooling often functions to maintain the prevailing social and political status quo. Through decisions about curricula, textbooks, assessment standards and institutional norms, education systems shape collective memory and define which histories are considered legitimate, marginal or inconvenient. In this sense, schooling operates as a “legacy conveyor,” transmitting dominant narratives from one generation to the next. It has become a sort of joke in recent history that the American education system lacks a certain something. Yet, here in Canada, we haven’t necessarily been all too fantastic ourselves. For example, many textbooks in both the U.S. and Canada either omit or sanitize histories of Indigenous dispossession, slavery or colonial violence, thereby denying full recognition of these realities and framing them as peripheral or resolved issues. Similarly, Japanese history lessons conveniently fail to teach the war crimes committed by the Japanese against the Chinese in Manchuria and Nanjing. In South Korea, the governing bodies responsible for education have controlled the official retelling of the Jeju Uprising through textbook revision and language control. Turkish textbooks minimize the impact of the Armenian Genocide through euphemisms, minimizing casualties, and avoiding “genocide” language. As recently as 2023, the Russian history curriculum was “revised,” and later criticized by Amnesty International for concealing truths, misrepresenting human rights violations committed by Russian forces and presenting Russia as victimized by the West. Examples of sanitized or outright inaccurate accounts of history are not exactly lacking, and indeed seem to be frequently politically motivated.
Yet, despite widespread recognition that knowledge is socially constructed and historically contingent, students are nevertheless routinely taught to defer to authority – memorizing sanctioned narratives, trusting credentialed experts and accepting the legitimacy of official curricula and assessments. This deference reflects authority bias, a cognitive tendency to attribute greater accuracy and legitimacy to the opinions of perceived authorities, which is not merely individual but structurally reinforced by schooling itself. In classrooms typically organized around a single teacher lecturing many students, knowledge flows hierarchically, positioning teachers, textbooks and standardized assessments as arbiters of truth. Within this structure, compliance is rewarded: students who reproduce authorized interpretations, follow prescribed rules and demonstrate deference to institutional expectations receive higher grades, positive evaluations and academic advancement. These mechanisms of grading and classroom discipline embed what Foucault terms disciplinary power, shaping students into docile bodies who internalize norms of conformity, legitimacy and self-regulation. Meanwhile, Bourdieu’s concept of habitus helps explain how pupils, teachers and administrators unconsciously reproduce cultural dispositions aligned with dominant classes. Because textbooks and curricula change slowly and are often shaped by political interests, dominant or convenient interpretations of history and “legitimate” knowledge tend to resist marginal critique. Voices seeking entry into the canonical curriculum – Indigenous scholars, people of color and feminist historians – thus encounter institutional inertia, political suppression and the gatekeeping of educational authority.
In recent years, AI has steadily infiltrated education – adaptive tutoring systems tailor content to a student’s pace, automated grading engines mark papers, chatbots serve as writing assistants and learning analytics track patterns of engagement and behavior. Students have already embraced this shift: a 2025 survey reported that 92% of undergraduates use some form of generative AI, often for drafting texts, summarizing readings or asking “tutor-style” questions. Meanwhile, research into the students’ perspectives on AI in education finds that undergraduates recognize AI’s affordances (feedback, access to information and personalized support) but also worry about overreliance, reduced critical thinking, hallucinated answers and data and privacy risks. Educators have also become concerned about AI in the classroom setting: in 2024, about a quarter of public K–12 teachers in the U.S. said AI tools do more harm than good in the classroom.
Beyond the classroom, AI seems to be ever present in day-to-day life whether the public consents to it or not. Search engines spew AI-generated content before links to other content, corporations push the use of AI on employees (despite credible evidence against the utility of such investments), and content moderation and hiring is increasingly run by algorithms. Newsfeeds, policing analytics and health diagnostics systems – all ingest digitized human behavior, learn from it and reinforce dominant patterns. For instance, commercial facial-analysis AI systems from major tech companies had error rates of up to 34.7% for darker-skinned women, compared with less than 1% for lighter-skinned men. The systems weren’t racist by intent; they were trained on datasets dominated by lighter-skinned, Western faces. As a result, the AI reinforced an existing dominant pattern which treats white men as the default and is therefore biased towards other groups, while presenting its outputs as objective and scientific.
In many public domains, our clicks, shares and engagements feed models that prioritize content conforming to prevailing norms or popularity, thereby amplifying certain narratives and silencing others. The same logic that in schooling rewards the iterated, familiar account now governs algorithmic media – what gets repeated, ranked and surfaced is rarely the most novel or controversial, but the most algorithmically compatible. An internal YouTube study and subsequent independent research found that the platform’s recommendation algorithm systematically amplified sensational, polarizing and norm-confirming content because such videos generated higher watch time and engagement. Users who clicked on mainstream political content were often nudged toward increasingly extreme or one-sided narratives – not because those views were more true, but because they were more engaging and already popular within certain audiences. Meanwhile, nuanced, minority, or cross-cutting perspectives were less likely to be recommended, effectively silencing them at scale. Similar algorithmic bias was found in Twitter, Instagram and Facebook. It is fairly widespread knowledge that the world runs on algorithms. From the early days of Facebook and curated content surfacing on your social media feeds to something as simple as your credit score, the world runs on numbers. So why is this AI shtick somehow more concerning?
Because, unlike the relatively simplistic algorithms of the past, today’s AI systems don’t just sort information – they generate it, often with an air of objectivity that masks deep structural bias. The demonstrated biases of AI have become a defining concern of the digital age. In technical terms, algorithmic bias refers to systematic errors that produce skewed or discriminatory outcomes across social groups – whether through imbalanced data, biased model design or the feedback loops of human use as illustrated in the examples above.
Yet even if training data were perfectly balanced, algorithmic or modeling bias would persist. Every model encodes human judgment – and therefore bias – through choices about which features to emphasize, which losses to minimize and which trade-offs to accept. Research has shown that these modeling decisions often mirror social hierarchies, embedding pre-existing social bias into the output. While many examples of these biases were cited indirectly throughout this text, this one I will explain directly. Many of you are likely familiar with Amazon – the multinational conglomerate, not the forest. Since 2014, a decade and some ago, Amazon has been investing in a maximally efficient (A.K.A automated) hiring process. By 2015 though, it became evident there was one major issue: the bot is sexist. Why? Simply, because of the data.
Picture this; a company decides to revamp and automate its hiring practices. They decide to train a sophisticated sorting algorithm on the historical data for the hires they’ve been intaking, which makes some logical sense. Except, the data spans about 10 years or so, and over the last 10 years, the sector – in this case tech – was heavily dominated by men. So, what happens when the majority of the data fed into the machine reflects the prevalence of men? The bot “taught itself that male candidates were preferable. It [penalized] resumes that included the word “women’s,” as in “women’s chess club captain.” And it downgraded graduates of two all-women’s colleges, according to people familiar with the matter.”
Amazon’s sexist bot is only one example of a negative outcome from AI. Generative models are well known to hallucinate, which is a cute way of saying they spew nonsense when they can’t figure out what to say. The point, though, is this: AI’s encoded biases are not always immediately obvious. With the black box nature of AI, there is no transparency in its decision making. Nevertheless, it is persistently used by students to summarize and explain, and by institutions to make crucial and potentially uncontested decisions. The encoding of bias will not remain a “fringe concern” for very long when AI systems can scale and automate biases while presenting themselves as neutral. What happened to Grok – the Twitter-embedded chatbot? When Elon Musk publicly set out to “de-wokeify” it, he wasn’t merely critiquing political correctness – he was rewriting the facts presented by his AI under the pretext of truth-seeking.
The concern here isn’t only in encoding bias subconsciously but in the overreliance on a system that cannot be trusted. Companies will optimize information for alignment with vested interests – political parties, ideological donors, low-tax agendas and corporate alliances. Misinformation has long been weaponised for political gain, and now the “victors” in our information battles are algorithms controlled by individuals who both shape and benefit from dominant narratives, from the status quo being firmly preserved. The risk deepens when such systems are entrusted with educating students, grading their work or amplifying content. The conveyor belt of legacy knowledge becomes not just passively but actively curated. And if profits and partisan power steer that curation, we should ask: who is writing history now – and what truths are being omitted? Is it really worth sacrificing natural resources for some incels to generate CP and weird hentai? Probably not. It is also worth asking yourself: are the results I’m seeing and relying on when communicating with this nonsense generator actually a reflection of what I think, or am I relying on it because it’s easy?
Here are some related searches for your future reading, please – pretty please – don’t ask ChatGPT:
Adam Raine; ChatGPT psychosis; Mechahitler Tuesday; Microsoft’s Tay; AI blackmail; and for a useful discussion on sentience and what it means to be alive – and to flex your critical thinking skills – watch Westworld on HBO (or pirate it, or don’t, I’m not your mom)
