Where (and When) Did the Hate Start?

A Polemic on the Origins and State of Online Hate in the Social Media Era

Abstract

The question sounds like it has a simple answer. Social media became toxic; that toxicity emerged from the combination of anonymity, scale, and the algorithmic amplification of engagement-driving content. The platforms, the argument goes, created the conditions; the users simply responded to incentives.

This piece argues the reverse. Social media did not accidentally create a landscape of hate. It was designed to exploit pre-existing human capacities for tribalism, suspicion, and cruelty, and it did so in the service of an economic model that required maximum engagement regardless of the quality of engagement. The anger, abuse, and coordinated hostility that now characterise the major platforms are not bugs. They are features, refined over two decades by some of the most sophisticated data-driven companies in human history.

The consequences are measurable and severe. In the United Kingdom, a country that has watched its international standing on minority rights decline to the bottom of the G7, the combination of social media platforms, inadequate regulation, and political complicity has produced an environment in which targeted groups face sustained, organised, and largely unpunished abuse. The platforms know this. The governments know this. Almost nothing effective has been done about it.

The answer to the question in the title is: the hate started when the conditions were created for it to be profitable. That means the early 2010s. And the geography is wherever the servers of Facebook, Twitter/X, and YouTube happen to be.

A World of Connections: What Social Media Actually Is

Before we can ask whether social media has become toxic, we need to be clear about what it is.

Social media platforms are digital infrastructure designed to facilitate the creation, sharing, and consumption of user-generated content within networked communities. They are not, as their founding mythology suggested, neutral spaces for human connection. They are commercial products, sold to advertisers on the basis of the attention they capture, and their design is optimised accordingly.

The major platforms by global reach as of 2025 are Facebook, with approximately 3.07 billion monthly active users; YouTube, with approximately 2.5 billion; Instagram, with approximately 3 billion; TikTok, with approximately 1.5 billion; and X (formerly Twitter), with approximately 600 million. Together, these platforms represent the primary medium through which a substantial fraction of the world's population communicates, forms opinions, and encounters information about the world.

The collective term for these platforms is the 'attention economy.' Their revenue model depends entirely on the amount of time users spend on the platform, and the volume of content they consume and share. This model creates a structural incentive to maximise engagement, and engagement, it turns out, is driven most reliably by content that provokes strong emotional responses: outrage, fear, indignation, and tribal solidarity. Calm, nuanced, accurate information is, from the perspective of the algorithm, a poor substitute for content that makes people angry.

As Bea Groves-McDaniel wrote at her blog in 2025:

"The question is never whether the technology is neutral. The question is who benefits from its current configuration, and who pays the price for it. The internet was not built to make us angry. It was built to make someone money. The anger is a side effect that turned out to be profitable."

This is the essential context for everything that follows. Social media is not a public utility. It is a private infrastructure, built to generate returns for shareholders, and its governance is determined by the commercial interests of the companies that operate it. When we speak of 'online hate,' we are speaking of a phenomenon that these companies have, at various points, actively tolerated because the anger drove the engagement which drove the revenue.

Was It Always This Bad?

No. And the 'no' is important.

Social media in its early years, roughly from 2004 (Facebook) and 2006 (Twitter) through the early 2010s, was not free of toxicity, but it was qualitatively different from what followed. The platforms were smaller, the algorithms less sophisticated, and the culture still shaped substantially by the norms of the early internet, which, for all its flaws, still carried something of the original ideals of networked community: openness, curiosity, and the expectation of pseudonymity.

The turning point is generally identified as the period between 2012 and 2016, and the name most associated with it is Cambridge Analytica.

In 2014, the political consulting firm Cambridge Analytica, working with a Cambridge University researcher named Aleksandr Kogan, harvested the personal data of approximately 87 million Facebook users without their consent. The data was used to build psychological profiles that were then deployed in political advertising campaigns, most notably for Donald Trump's 2016 presidential election and the Brexit referendum of the same year.

The revelation of this operation in March 2018, by the New York Times and the Guardian in collaboration with Channel 4, did not merely expose a data breach. It exposed the fundamental nature of what Facebook had allowed to happen on its platform. The Information Commissioner's Office investigation found that Facebook had failed to protect users' data adequately and that the scale of data harvesting was 'very significant.' The House of Commons Digital, Culture, Media and Sport Committee later described the incident as evidence of 'a digital democracy racket.'

What Cambridge Analytica demonstrated was that Facebook was not a community. It was a surveillance infrastructure that had been dressed up as a social utility. And that surveillance infrastructure was available, for a price, to anyone who wanted to use it to manipulate how people voted.

The platform had, in other words, been weaponised before most people understood what was being built.

The second major inflection point was the algorithmic pivot. Around 2016, major platforms, facing the need to compete for advertising revenue in a mobile-first world, began to rely more heavily on algorithmic content recommendation. The effect was to shift the distribution of content from social networks (what your friends shared) to algorithmic feeds (what the algorithm predicted would keep you engaged). Research published by Allcott and Gentzkow in 2017 found that the diffusion of misinformation on social media had increased substantially between 2012 and 2016, and that the mechanisms of algorithmic amplification were a significant driver.

The content that spread fastest was not accurate content. It was content that provoked emotional responses. The algorithm had learned, from two billion people's behaviour, that anger was more engaging than accuracy. And it was right.

What Do We Mean By 'Bad'?

The question deserves a precise answer, because 'bad' can mean too many things.

What is meant here is: the systematic increase in the volume and severity of hostile, abusive, and hateful content directed at identifiable groups, facilitated and amplified by the architecture and economics of the major platforms.

The evidence for this increase is substantial.

Ofcom's survey on experiences of hateful content, published in July 2026, found that a significant proportion of UK adults had encountered hateful content directed at protected characteristic groups online in the preceding twelve months. The survey, conducted among adults aged 18 and over, found that exposure to online hate was common and that reporting mechanisms were widely perceived as ineffective.

TransActual's landmark survey, Trans Lives 2025, published in March 2026, found that 94% of trans people in the UK had experienced some form of cisgenderism or transphobia in the previous twelve months. More than half (54%) had experienced harassment online. The survey noted that online hostility frequently translated into offline harm: 45% of respondents reported that online abuse had negatively affected their mental health, and 23% had experienced physical threats.

The Home Office hate crime statistics for England and Wales, covering the year ending March 2025, recorded 4,120 police-reported hate crimes with a transgender bias. Galop, the UK's LGBT+ anti-violence charity, has consistently reported that online abuse of LGBT+ people has increased in both volume and severity over the past decade, with particular acceleration following major political events such as the Brexit referendum and the 2019 general election.

YouTube, specifically, has been identified by multiple researchers as a significant vector for radicalisation. A 2019 investigation by the Guardian found that YouTube's recommendation algorithm was actively directing users from mainstream political content toward increasingly extreme material, including far-right and white nationalist content. The company denied this. Internal documents later disclosed, as part of the Facebook Papers whistleblower release of 2021, suggested that senior executives at YouTube were aware of the problem and had concluded that addressing it would reduce engagement metrics.

The 'badness' of contemporary social media is not an aesthetic judgment. It is a measurable social harm, and its measurement begins with the recognition that the platforms are not passive conduits of content. They are active curators, and their curatorial choices are made on behalf of a commercial interest that is systematically in conflict with the wellbeing of their users.

The Worst Served: Who Faces the Most Abuse

The distribution of online hate is not random. It follows the contours of existing social hierarchies and amplified prejudices.

In the United Kingdom, the groups that face the most severe and sustained online abuse include transgender people; Black, Asian, and minority ethnic communities; Muslim communities; Jewish communities; women; and LGBT+ people more broadly.

Trans people are, by multiple measures, among the most severely affected. The Trans Lives 2025 survey found that 94% of trans respondents had experienced cisgenderism or transphobia in the preceding year, with online harassment the most common form. The overlap between online hostility and offline threat is significant: many trans people reported that online abuse was followed by physical harassment, doxxing, or threats.

Racial and ethnic minorities face sustained racist abuse on all major platforms. In the aftermath of events such as the Southport riots of summer 2024, research by Hope Not Hate documented how social media platforms, particularly X/Twitter and Facebook, were used to spread false information that directly contributed to public disorder. The connection between online misinformation and offline violence is now well established.

Muslim communities in the UK have reported consistently high levels of online abuse, much of it organised and coordinated. The January 2025 report from the APPG on British Muslims found that online hate speech directed at Muslim communities had increased substantially since October 7th 2023, and that platform responses to reports were frequently inadequate.

Women, and particularly women in public life, face sustained abuse that frequently includes sexualised threats of violence. The sheer volume and character of this abuse has led many women to withdraw from online spaces, which represents a form of self-censorship that has well-documented chilling effects on political participation.

As Groves-McDaniel has observed:

"The internet was supposed to democratise voice. What it actually did was democratise access to the amplifying megaphone, while leaving the question of who gets to speak without consequence entirely unresolved. Anonymous by design, monetised for engagement, and optimised for outrage, social media created the perfect infrastructure for the world's least courageous bigot to become, at zero cost, a public figure of enormous and unaccountable influence."

Anonymity and the Bot Question

The standard defence of online hostility is anonymity. The argument is that people say things online that they would never say face to face because the social constraints of physical presence are removed. Give everyone a real identity and the problem solves itself.

The argument is partially right and largely irrelevant.

Anonymity does reduce inhibitions. The online disinhibition effect, documented by psychologists John Suler and others, describes the psychological mechanism by which people behave differently online than they would in person, often exhibiting greater hostility, honesty, or vulnerability than they would display under conditions of social accountability.

But anonymity is not the primary driver of online hate. If it were, we would expect pseudonymity (where users adopt persistent pseudonyms rather than their real names) to produce the same toxicity as anonymity, and it does not. Platforms with strong real-name cultures, such as Facebook with its policy of requiring users to identify by their legal name, have not demonstrated meaningfully lower rates of coordinated harassment and hate speech.

What matters is not whether the identity is real, but whether there are meaningful consequences for harmful behaviour. Facebook's real-name policy did not prevent the Cambridge Analytica scandal, the Myanmar genocide footage that spread freely on the platform for years, or the coordinated anti-trans harassment campaigns that have been documented by multiple researchers. Real names provide no protection when enforcement is absent.

The bot question is related but distinct. Automated accounts, or 'bots,' do contribute to the amplification of harmful content: research by the Oxford Internet Institute found that bot activity accounted for a significant fraction of political content sharing during the 2016 and 2020 US elections, and that bots were disproportionately likely to share misinformation.

However, blaming bots for the problem of online hate is a way of externalising responsibility. Even if every bot were removed tomorrow, the human users who choose to engage with hateful content, share it, amplify it, and act upon it would remain. Bots are a force multiplier for existing hostility, not its source.

The deeper point is that the architecture of the major platforms, by rewarding engagement above all else, has created conditions in which the performance of outrage is rational behaviour for users who want visibility, and in which the amplification of hate is rational behaviour for platforms that want retention. Bots and anonymity are variables in this equation. They are not the answer.

The Global Framework: A Structural Problem

The major social media platforms share a common architecture. They are centralised, advertiser-supported, and optimised for engagement. Their governance is determined internally, by the companies that operate them, and is not subject to meaningful democratic accountability in any jurisdiction in which they operate.

This architecture is not accidental. It is the product of choices made by specific people, at specific moments, for commercial reasons. The consequences of those choices are now global.

The problems with this model are structural rather than incidental.

First, the incentive structure is wrong. A platform that earns revenue from attention has a commercial interest in keeping users angry, afraid, and engaged, regardless of the quality of that engagement. This is not a theory. It is documented in internal memoranda, whistleblower testimony, and the published research of former senior employees. The platforms are not unaware of the harms they produce. They have, in many cases, calculated that addressing those harms would reduce their commercial performance.

Second, the scale of the infrastructure means that harms that would be criminal in any other context are treated as content moderation problems. A phone call that threatened someone's life would be a police matter. A tweet that does the same is processed by an automated system that may or may not respond within 24 hours, and that response, if any, is determined by the platform's terms of service rather than the law.

Third, the cross-border nature of the infrastructure means that national regulation is systematically circumvented. The UK Online Safety Act 2023 is a serious attempt at regulation, but its enforcement mechanisms are limited, and platforms with global reach can, and do, make compliance decisions that serve their global commercial interests rather than the specific requirements of any national legal framework.

The Fediverse: An Alternative Architecture

The alternative most frequently proposed to the corporate social media model is the Fediverse: a network of independently operated servers that communicate using open protocols, most notably ActivityPub.

The Fediverse is not new. Mastodon, the largest platform within it, launched in 2016. What has changed is the scale of adoption, particularly following Elon Musk's acquisition of Twitter in 2022, which prompted a significant migration of users to Fediverse alternatives. By May 2026, the Fediverse encompassed approximately 15.3 million active users across over 19,000 independent servers.

The architectural differences from corporate social media are significant. Because the Fediverse is decentralised, there is no single entity that controls the network, sets the algorithm, or determines content policy for all users. Each server (or 'instance') sets its own rules, and instances can choose to federate or not with other instances based on their content policies. This creates a market of communities with different norms, rather than a single global platform with a uniform set of rules determined by shareholders.

The advantages are real. Users can migrate between instances without losing their social graph. Instances that tolerate abusive behaviour can be defederated by others, creating a form of decentralised accountability. The absence of an engagement algorithm means that content is displayed in reverse chronological order rather than optimised for outrage.

The disadvantages are equally real. The Fediverse is harder to use: onboarding requires choosing an instance, understanding the federation model, and navigating a less polished interface. The lack of a unified platform means that discoverability is lower and network effects are weaker. And the decentralised governance model, while protecting against corporate capture, creates gaps in enforcement: instances with inadequate moderation policies can cause harm to their own users without affecting the broader network.

There is also a more fundamental limitation: the Fediverse, at 15 million users, serves a population roughly equivalent to the city of London. The major platforms, with billions of users between them, are the actual public sphere, and leaving them to do not means simply ceding that sphere to the dynamics described above.

The Fediverse is an important experiment and a genuine alternative for those who can use it. It is not a solution to the structural problems of corporate social media at the scale at which those problems operate.

What the UK Government Has Done (and Failed to Do)

The United Kingdom has, in the past decade, developed a substantial body of legislation intended to address online harms, and has consistently failed to enforce it effectively.

The Online Safety Act 2023 was seven years in the making. It was the most significant attempt by any government to impose meaningful obligations on social media platforms regarding the content they host and amplify. It created new duties for platforms to protect children, to address illegal content, and to enforce their own terms of service. It established Ofcom as the regulator with powers to fine platforms up to £18 million or ten percent of global annual turnover, whichever is higher, for breaches of their duties.

The Act was, at the time of its passage, widely welcomed as a genuine step forward. Its implementation, beginning in 2024 and continuing into 2025 and 2026, has been another matter.

A report from the Science, Innovation and Technology Committee, published in July 2025, found that the Online Safety Act had 'major holes' and would not prevent a repeat of the Southport riots, in which online misinformation spread via social media platforms directly contributed to public disorder. The Committee noted that the Act failed to adequately address misinformation that caused harm without crossing the threshold of illegality: precisely the category of content that drove the riots.

As the BBC reported in June 2025, the Act's provisions on misinformation were widely regarded as inadequate by civil society organisations, children's safety charities, and the police themselves. The government's own impact assessments, published alongside the legislation, had understated the scale of the problem.

The deeper failure is one of political will. Both major parties have treated social media regulation as a toxic issue, in the sense that engaging with it seriously risks attracting the organised hostility of the platforms and their allies. The platforms have demonstrated, repeatedly, that they are willing to use their economic leverage to resist regulatory requirements: threatening to withdraw services, lobbying against specific provisions, and in some cases simply ignoring enforcement deadlines.

The result is regulation that is better on paper than in practice, and a regulatory body, Ofcom, that is tasked with enforcing obligations against companies with annual revenues in the tens of billions, with limited resources and limited political support.

What Can Be Done

The honest answer is that no individual, no single platform, and no one government can solve this problem. The structural issues are too deep and the commercial incentives are too powerful. But this does not mean that nothing can be done. It means that what can be done requires coordination, persistence, and a willingness to name what is actually happening.

On platform design: the engagement-maximisation model is not the only possible model. Platforms can be built with different incentive structures. The Fediverse demonstrates this. Platforms with non-profit governance, co-operative ownership models, or public utility frameworks are technically feasible. Their development requires investment, policy support, and users willing to try alternatives.

On regulation: the UK Online Safety Act, for all its flaws, provides a framework. The priority should be strengthening its provisions on algorithmic amplification (which currently receives less attention than content hosting), on duty-of-care obligations that attach to the design choices that produce harms rather than the harms themselves, and on meaningful transparency requirements that allow researchers and regulators to understand what the platforms are actually doing.

On individual platforms: there is no evidence that self-regulation works. Platforms regulate harmful content when it becomes politically expensive not to do so. The campaigns by Galop, TransActual, Hope Not Hate, and other organisations that have documented and reported platform failures have had more impact on platform behaviour than any amount of voluntary commitment.

On collective response: the history of online hate suggests that the most effective countermeasure is not argument but solidarity. Communities that organise against harassment, that share resources, that document and report systematically, and that refuse to be driven from public spaces, are communities that maintain presence. Visibility, where it is safe, is itself a form of resistance.

On individuals: the evidence suggests that personal familiarity with members of targeted groups is the single strongest predictor of reduced prejudice. This is as true online as offline. The Mystified Majority, described in the previous polemic in this series, are reachable not by argument but by encounter.

AI, Deepfakes, and the Future

The capacity of artificial intelligence to generate convincing synthetic media, including 'deepfake' images, audio, and video, represents an inflection point in the severity of online harms.

The Internet Watch Foundation's 2025 annual report found that reports of AI-generated child sexual abuse material had more than doubled. The NPCC survey, published in 2025, found that one in four people felt there was nothing wrong with, or felt neutral about, creating and sharing sexual deepfakes, even when the person depicted had not consented. The police themselves described the trend as an 'escalating threat.'

The deepfake problem is distinctive in several ways. Unlike text-based abuse, synthetic media can be used to create apparent evidence of events that never occurred: pornographic material depicting real people, political footage designed to manipulate elections, fabricated evidence of crimes. The technology is now widely available, cheap, and produces outputs that are increasingly difficult to distinguish from authentic recordings.

The government's response has been to announce new legislation targeting AI-generated abuse images. The Online Safety Act was not designed with this problem in mind, and its enforcement against AI-generated content has been widely described as inadequate.

The trajectory is not positive. As the cost of generating synthetic media falls and the quality of outputs rises, the capacity to weaponise it against targeted individuals will increase. This is an area where regulation has consistently lagged behind technology, and where the gap between the two is currently widening rather than narrowing.

Summary

The hatred that is now endemic on social media did not emerge spontaneously from human nature. It was cultivated, at scale, by platforms that discovered that outrage was more profitable than accuracy, and that the most reliable form of outrage was directed at the most vulnerable groups in society.

The problem is structural, not incidental. It is built into the commercial model of the major platforms, the inadequacy of national and international regulatory frameworks, and the political economy of companies whose revenues depend on attention that is generated by conflict.

The groups most severely affected are those who were already marginalised: trans people, ethnic minorities, religious communities, women. The harms they experience are real, measurable, and documented by the very institutions that are supposed to protect them.

There are alternatives. There are platforms with different architectures, different incentive structures, and different governance models. There are regulatory frameworks that, if enforced, would impose meaningful obligations on platforms. There are collective responses that, when organised and sustained, have demonstrated impact.

What there is not is evidence that the problem will solve itself. The trajectory is set by commercial interests, and those commercial interests are not aligned with the wellbeing of the people who are being harmed.

The question of where and when the hate started has a factual answer: it started with the conditions that made it profitable. The question of where and when it ends is still open.

Bibliography

DataReportal. Digital 2025: Top Social Platforms. 2025. Global social media user statistics. Facebook 3.07 billion monthly active users; YouTube 2.5 billion; Instagram 3 billion. Available at datareportal.com.

Statista. Biggest Social Media Platforms by Users, 2025. October 2025. Available at statista.com.

Allcott, Hunt, and Matthew Gentzkow. "Trends in the Diffusion of Misinformation on Social Media." Research Paper, 2017. Stanford University / NYU. Found significant increase in misinformation diffusion 2012-2016, driven by algorithmic amplification and social sharing dynamics.

Information Commissioner's Office. "Investigation into the Use of Data Analytics in Political Campaigns." Report to Parliament, 5 November 2018. Cambridge Analytica / Facebook data harvesting; 87 million users affected; Facebook found to have failed to protect user data adequately.

House of Commons Digital, Culture, Media and Sport Committee. Disinformation and 'Fake News': Interim Report. 2018. Described Cambridge Analytica operation as 'a digital democracy racket.' Available at publications.parliament.uk.

Facebook Papers (Whistleblower Disclosure, 2021). Internal Meta memoranda released to Congress and published by the Wall Street Journal. Documented senior executives' awareness of algorithmic amplification of harmful content; YouTube internal documents on radicalisation.

Hope Not Hate. Post-Southport riots research, summer 2024. Documented role of social media platforms (particularly X/Twitter and Facebook) in spreading misinformation that contributed to public disorder.

Ofcom. Experiences of Hateful Content. Survey published 10 July 2026. UK adults aged 18+; encounters with hateful content online in preceding 12 months; effectiveness of reporting mechanisms.

TransActual. Trans Lives 2025: Continuing to Endure the UK's Hostile Environment. March 2026. 94% of trans people experienced cisgenderism or transphobia in preceding 12 months; 54% experienced online harassment; 45% reported negative mental health impact from online abuse.

Home Office. Hate Crime, England and Wales, Year Ending March 2025. Police-recorded hate crimes with transgender bias: 4,120. Available at gov.uk.

Galop. Online Hate Crime Report. 2020 (and ongoing). Pattern of increasing online anti-LGBT+ abuse in volume and severity; inadequate platform response to reports. Available at report-it.org.uk.

Internet Watch Foundation. Annual Data Insights Report 2025. AI-generated CSAM reports more than doubled year-on-year. Available at iwf.org.uk.

NPCC / Police. Survey on sexual deepfakes, 2025. One in four people felt there was nothing wrong with, or felt neutral about, creating and sharing sexual deepfakes without consent. Available at news.npcc.police.uk.

Science, Innovation and Technology Committee. Report on Online Safety Act, July 2025. Found major gaps in Act; insufficient to prevent repeat of Southport riots; fails to adequately address non-illegal but harmful misinformation. Available at publications.parliament.uk.

BBC News. "New Online Safety Rules Are Here but as Tech Races Ahead, Expect Changes." June 2025. Online Safety Act implementation; gaps in misinformation provisions.

POLITICO. "UK Online Safety Laws Won't Stop a Repeat of Southport Riots, MPs Warn." July 2025. Online Safety Act criticised as inadequate by cross-party committee.

The Register. "UK Online Safety Act 'Not Up to Scratch' on Misinformation." 11 July 2025. Available at theregister.com.

Fediverse Observer / Libera Site. Fediverse statistics May 2026. Approximately 15.3 million active users; 19,000+ instances; Mastodon approximately 68% of ActivityPub users. Available at activitypub.fediverse.observer.

Suler, John. "The Online Disinhibition Effect." Cyberpsychology and Behaviour, 2004. Found that anonymity and text-based communication reduce behavioural inhibitions relative to face-to-face interaction; relevant but not determinative of online hostility.

Oxford Internet Institute. Research on bot activity in political content sharing during 2016 and 2020 US elections. Found bots disproportionately represented in sharing of misinformation.

Guardian. Investigation into YouTube recommendation algorithm, 2019. Documented systematic direction of users from mainstream to extreme content.

Groves-McDaniel, Bea. "Satires 3: Let's Have a Good Ol' Panic!" Being Bea, July 2026. Available at blog.beagmcd.net. Analysis of moral panic as political technique and the specific role of digital infrastructure in facilitating it.

Groves-McDaniel, Bea. "Satires 1: What's Wrong with TERFs?" Being Bea, July 2026. Available at blog.beagmcd.net. Relevant to analysis of how hostile movements use digital platforms and how communities of hate construct and target their objects.