Algorithmic amplification

A conceptual diagram illustrating how an engagement algorithm takes content beyond a user's organic social graph and amplifies it to out-of-network users

Algorithmic amplification is the process by which automated ranking and recommendation systems on digital platforms increase the visibility of certain content beyond its initial audience. Major platforms, including Facebook, YouTube, TikTok, and X (formerly Twitter), use such systems to determine what appears in users' feeds and search results. The term is used in research on social media and digital media regulation to describe how platform design choices influence the distribution of online information.

Unlike chronological feeds, algorithmic systems evaluate content using signals such as engagement rates, viewing duration, and predicted relevance to individual users. Content that performs strongly on these metrics may be promoted to progressively larger audiences through feeds, suggested videos, and search results, and early engagement can feed back into the model that ranks what users are shown next. The process is distinct from content moderation, which involves removing, labelling, or restricting content under platform rules, although the two may interact in practice. Recommendation systems also support content discovery and public-interest communication, and their ranking decisions shape the visibility and income of creators and news organisations.

Algorithmic amplification has been linked to the spread of misinformation, the circulation of extremist and other harmful material, and harms affecting young users' mental health. It also features in debates about filter bubbles and echo chambers, and about political polarisation, where findings vary across platforms and methods. Studies of X have found uneven amplification of political content. On Facebook and Instagram, large experiments found limited effects on users' attitudes. A counterfactual-bot study of YouTube found that user preferences played a larger role than recommendations in determining partisan consumption. Researchers have also examined how state actors and automated accounts use recommendation systems to shape what is seen. The scale and direction of these effects remain debated, in part because independent researchers have limited access to platform systems. Internal Meta documents disclosed by whistleblowers indicated that engagement-based ranking rewarded divisive content. A 2026 BBC investigation reported that competitive pressure between Meta and TikTok led to safety trade-offs in content recommendation. Meta denied the whistleblowers' claims, while TikTok disputed the account of its case-review process.

Governments in the European Union, United Kingdom, United States, and China have taken differing approaches to regulating recommendation systems. The EU's Digital Services Act requires the largest platforms to assess and mitigate systemic risks associated with recommendation systems, and to offer users at least one recommendation option not based on profiling. The UK's Online Safety Act 2023 requires services to assess risks arising from algorithms and, under Ofcom's child-safety codes, requires some providers to filter harmful content from children's feeds. A House of Commons committee concluded in 2025 that the Act did not adequately address the amplification of legal but harmful content. No comprehensive federal law has been enacted in the United States, where courts have considered whether platforms are liable for what their algorithms recommend and whether ranking constitutes protected speech. China's provisions on algorithmic recommendations, in force since March 2022, require providers to let users turn off personalised recommendations, and some providers to register their algorithms with the state; the media scholar Jian Xu describes China as the first country to enact and apply laws regulating algorithms and generative artificial intelligence.

Terminology

The term algorithmic amplification is used in media studies, platform governance scholarship and regulatory literature to describe how automated systems influence the distribution of content. The communications scholar Tarleton Gillespie argues that content-sharing platforms present themselves as hosts of user content while playing down how they intervene, including by sorting content, highlighting some posts over others, and filtering what users see.

The phrase also appears in regulatory and legislative discussions of recommendation systems. The European Union's Digital Services Act (DSA) identifies recommendation systems as a potential source of systemic risk, and the term appears frequently in academic and policy commentary on regulation. In the United States, proposals including the Filter Bubble Transparency Act and the Kids Online Safety Act (KOSA) have used it to frame requirements around recommendation system transparency. In the United Kingdom, the House of Commons Science, Innovation and Technology Committee used the term in a 2025 report on how recommendation algorithms contributed to the spread of misinformation during the 2024 Southport riots. A Joint Declaration on AI and Freedom of Expression was adopted in October 2025 by four international freedom of expression mandate holders, including the UN Special Rapporteur on Freedom of Opinion and Expression and the OSCE Representative on Freedom of the Media. The declaration stated that recommender systems and other AI-powered curation tools exert "a large hidden influence and gatekeeper role" over what information people access and consume.

Development of recommendation systems

Modern recommendation systems predate social media. A 2021 overview in AI Magazine by Dietmar Jannach and colleagues traced the origins of modern recommendation systems to the early 1990s, when they were first used experimentally for personal email and information filtering. The 1992 Tapestry mail system and the 1994 GroupLens news filtering system were early milestones before recommendation systems spread into e-commerce and other online services. As large platforms developed during the 2000s, they increasingly sorted, filtered and highlighted material rather than merely hosting it.

Facebook introduced its News Feed in 2006, which gradually shifted from chronological presentation to algorithmically ranked content. YouTube altered its suggested-videos and search ranking algorithms in 2012 to prioritise watch time over clicks, saying it wanted to "better surface the videos that viewers actually watch, over those that they click on and then abandon". TikTok, launched internationally in 2018, adopted a model in which its primary content surface, the For You feed, is driven almost entirely by algorithmic recommendation rather than by a user's social graph. An internal document obtained by The New York Times in 2021 showed that the platform's algorithm optimised for retention and time spent, using signals such as watch duration, replays, likes, and comments to score and rank videos.

Algorithmic recommendation also became central to platforms outside social media. Spotify's personalised features, including Discover Weekly, Release Radar, and Home recommendations, use behavioural signals and inferred "taste profiles" to surface tracks and artists beyond a listener's existing library. An ethnographic study of music curators at streaming platforms described this blend of algorithmic and human editorial selection as an "algo-torial" model of gatekeeping, through which playlist placement can substantially increase a track's reach. Amazon adopted item-based collaborative filtering for product recommendations in 1998, and its recommendation engine has been described as one of the earliest large-scale deployments of recommendation technology in e-commerce, shaping product visibility for many shoppers.

Comparable recommendation dynamics have also been identified on adult content platforms. Law professor Amy Adler has argued that from 2007 onwards the pornography industry migrated to algorithm-driven streaming platforms, most of which are controlled by a single near-monopoly company, Aylo (formerly MindGeek). These platforms use algorithmic search engines, suggestions, rigid categorisation, and AI-driven search term optimisation in ways that produce the same distorting effects found on mainstream speech platforms, including filter bubbles, feed loops, and the tendency of algorithmic recommendations to alter individual preferences.

Mechanisms

A simplified recommendation pipeline showing how content can be filtered and ranked into a personalised feed

Recommendation systems commonly combine collaborative filtering, which predicts a user's preferences from the behaviour of similar users, with machine-learning models that predict which content a user is likely to engage with based on their prior activity. Such systems predict the utility of each item for a given user and filter a large catalogue down to a short ranked list of items with the highest predicted relevance.

Engagement-based ranking systems score content by predicting whether a user will interact with it, using signals such as clicks, shares, likes, replies, video watching, and time spent on a post. Users' revealed preferences, expressed through this behaviour, do not always align with the preferences they state when surveyed.

The feedback loop in a recommender system: user interactions feed back into the model that ranks what they are shown next

Popularity signals can create feedback dynamics in which early engagement increases the likelihood that content will be shown to additional users. Experimental research on online cultural markets has demonstrated how such feedback processes can produce unequal visibility outcomes even when initial differences in content quality are small. As users consume recommendations and their responses are fed back into the model, recommender systems can enter a feedback loop. Simulation studies have found that such loops can amplify popularity bias, reduce the diversity of recommended items, and homogenise users' preferences over time.

Beneficial and public-interest uses

Recommendation systems can help users navigate large volumes of content by surfacing material predicted to match users' interests or needs, which can improve discoverability on platforms with large content libraries. In public health communication, platforms can help health authorities distribute timely information at scale, though the same systems also risk amplifying misinformation alongside official guidance.

Sociologist Zeynep Tufekci has argued that networked public spheres became increasingly centred on large commercial platforms, where visibility was often shaped by opaque algorithms and platform policies. In the case of the Egyptian uprising of 2011, she wrote that many Egyptians joined Facebook for social reasons, which also drew them into a networked public sphere where exposure to activists' political material was a side effect of Facebook use.

Social media platforms have also been used during emergencies to distribute situational updates and coordinate response. These platforms have become a significant channel for public participation and back-channel communication in crises, though the same platform infrastructure that accelerates the spread of useful information can also amplify rumours and misinformation during fast-moving events.

Algorithmic amplification also affects cultural visibility on streaming platforms. A 2023 UK government report on music streaming described recommendation systems as acting as a cultural intermediary between listeners and music, while noting persistent concern among creators and industry stakeholders that such systems may unfairly advantage some artists or genres over others.

Musicologist Georgina Born and computer scientist Fernando Diaz have argued that recommendation algorithms for cultural content should promote diversity and a commonality of experience rather than optimising solely for individual engagement, drawing on the programming traditions of public service broadcasting organisations such as the BBC.

Effects on information ecosystems

Research on algorithmic amplification has examined how it shapes the wider information environment. Much of this work concerns the spread of misinformation and harmful content, and the visibility and revenues of creators and news outlets. Other studies address filter bubbles, political polarisation, the effects on young users' mental health, and the use of recommendation systems by state actors.

Misinformation and harmful content

False news stories spread faster and more broadly than accurate stories on Twitter (now X), according to a 2018 study, although the authors attributed this primarily to human sharing behaviour rather than to platform algorithms. Concerns about the recommendation of borderline or conspiratorial material led YouTube to announce changes in January 2019 aimed at reducing recommendations of videos that approached but did not violate the platform's rules.

In 2021, former Facebook product manager Frances Haugen disclosed thousands of internal company documents to the U.S. Securities and Exchange Commission and the Wall Street Journal, which published them as the Facebook Files. The documents included internal research showing that a 2018 change to Facebook's news feed algorithm, designed to prioritise what the company called "meaningful social interactions", had the effect of amplifying divisive and emotionally charged content because such material generated higher engagement. In an interview with CBS News ahead of her testimony to the United States Senate in October 2021, Haugen argued that Facebook's engagement-based ranking systematically promoted content that provoked anger and that the company's leadership was aware of this effect but had not acted to address it. Facebook, which later rebranded as Meta Platforms, said it had invested $13 billion in safety and security and employed 40,000 people working on the issue.

A 2024 study using experimental "counterfactual bots" to isolate the causal role of YouTube's recommender found that, on average, the algorithm pushed users towards more moderate content rather than more extreme material. This moderating effect was strongest for heavy consumers of partisan content, and the authors concluded that individual user preferences played a larger role than algorithmic recommendations in determining consumption patterns.

A 2025 algorithmic audit of X found that the platform's engagement-based ranking algorithm amplified content that was emotionally charged and hostile to members of opposing political groups, compared to a reverse-chronological baseline. Users did not prefer the political content selected by the engagement-based algorithm when asked to evaluate it directly, suggesting a gap between what drives engagement and what users report valuing.

Social bots, automated accounts that mimic human behaviour on social media platforms, can also function as amplifiers of algorithmic visibility. A 2022 analysis of 1.6 million COVID-19-related tweets alongside 50,000 news stories found that bot accounts, which constituted approximately 9 per cent of the dataset, selectively promoted certain pandemic-related topics. The topics that bots amplified predicted subsequent coverage by partisan news outlets, and the relationship was bi-directional: news coverage also predicted subsequent bot activity on the same topics.

Human rights investigations have also linked algorithmic amplification to mass violence. Amnesty International argued that Facebook's news feed, groups, and recommendation features actively amplified anti-Rohingya hatred in Myanmar in the years preceding the 2017 atrocities, helping to intensify the circulation of divisive and inflammatory content.

The 2019 Christchurch mosque attacks in New Zealand, in which the attacker live-streamed the shooting, drew attention to the role of recommendation systems in redistributing terrorist content. The Christchurch Call, a multilateral commitment adopted by governments and technology companies in May 2019, identified algorithmic amplification as a factor in the spread of terrorist and violent extremist content and committed signatories to review how recommendation algorithms direct users towards such material. In 2022, the Call launched the Christchurch Call Initiative on Algorithmic Outcomes (CCIAO) to develop privacy-preserving tools for independent researchers to audit recommendation systems for radicalisation pathways. The legal question of whether platforms bear liability for algorithmically recommending terrorist content reached the Supreme Court of the United States in Gonzalez v. Google LLC (2023), in which the family of a victim of the November 2015 Paris attacks argued that YouTube's recommendation algorithm had directed users towards ISIS recruitment videos. The Court declined to rule on the Section 230 question, disposing of the case on other grounds. The Third Circuit reached a different conclusion in Anderson v. TikTok (2024), in which the mother of a ten-year-old girl who died attempting the "blackout challenge" alleged that TikTok's algorithm had promoted the video to her daughter. The court held that the recommendations a platform's algorithm makes constitute the platform's own "expressive activity" rather than third-party content, and so are not shielded from liability by Section 230. The panel relied in part on the Supreme Court's reasoning in Moody v. NetChoice (2024), which described platforms' editorial choices about ranking and presentation as protected expression. The Third Circuit denied TikTok's petition for rehearing en banc in October 2024. Whether recommendation algorithms actively drive users towards extremist content remains disputed. A 2021 peer-reviewed study found that while extremist and fringe content did appear in platform recommendations, policymakers had yet to grasp the difficulty of "de-amplifying" legal but borderline material, and that the conceptual distinction between users' own choices and algorithmic effects was often unclear in both academic and policy discussion.

A 2026 BBC investigation based on testimony from more than a dozen whistleblowers and insiders at Meta and TikTok reported that competitive pressure between the two companies led to safety trade-offs in content recommendation. Matt Motyl, a former senior researcher at Meta, shared internal research on Instagram Reels, launched in 2020 in response to TikTok. One research paper indicated that posts on Reels drew a higher prevalence of harmful comments than posts on the main Instagram feed: 75 per cent higher for bullying and harassment, 19 per cent higher for hate speech, and 7 per cent higher for violence and incitement. A former Meta engineer said that senior management had told him to allow more borderline harmful content in users' feeds to compete with TikTok, and that they had attributed the decision to the company's falling share price. Internal documents shared by Motyl described how Facebook's engagement-based algorithm tended to reward negativity, and said that the financial incentives its algorithms created did not appear to be aligned with the company's mission. Meta denied the whistleblowers' claims, saying it had strict policies to protect users and had made significant investments in safety and security over the previous decade.

Creator visibility and economic effects

Social media platforms maximise the value of their audiences by increasing audience size, time spent on the platform and user interaction, since these interactions provide data for attracting audiences and targeting messages to advertisers. A small number of dominant platforms concentrate the distribution of algorithmically directed audience attention, and their potential to act as gatekeepers over the content they disseminate means that the impact of any flaws or biases in their algorithmic curation systems can be considerable.

A study of 37 German legacy news outlets' Facebook and Twitter activity between 2013 and 2017 examined how outlets adjusted their use of clickbait headlines over time. On Facebook, outlets converged towards an industry-wide clickbait standard, though the same convergence was not found on Twitter. Across both platforms, past user interaction with clickbait predicted higher subsequent clickbait supply. The relationship between clickbait and user interaction followed an inverted U-shape: moderate levels generated the most interaction, while higher levels led to declining returns. Introducing algorithmic curation was not shown to directly increase clickbait supply. Facebook's anti-clickbait algorithm changes did, however, disperse the previously convergent behaviour of outlets, reducing industry-wide homogeneity.

The rapid adoption of generative artificial intelligence tools from 2023 onwards has lowered the cost of content production, increasing the volume of material available for recommendation systems to rank. Game-theoretic modelling by the marketing scholars Tianxin Zou, Zijun Shi, and Yue Wu examined the welfare effects. Where such tools raise the quality of low-quality content only modestly, they found, the influx of that content crowds out the production of high-quality content, reducing consumer welfare and creators' total profit. The negative effect is more likely where platforms screen content quality effectively, because screening amplifies the visibility that the improved low-quality material gains.

Share of Google searches in which users clicked a traditional search result, with and without an AI Overview. Clicking a result was one of five outcomes measured; most other visits continued searching or left Google (Pew Research Center, 2025)

The integration of large language models into search has introduced a further layer of algorithmic curation that operates differently from earlier feed-ranking and search-ranking systems. Google's AI Overviews feature, launched in May 2024, displays a generated summary above the conventional list of search results. A Pew Research Center analysis published in July 2025, based on the browsing activity of 900 US adults during March 2025, found that users clicked on a traditional search result in 8 per cent of visits when an AI Overview was present, compared with 15 per cent when no summary appeared. Users clicked on links within the AI summary itself in about 1 per cent of visits, and were more likely to end their browsing session after seeing a summary. Communications scholar Dan Valeriu Voinea has argued that conversational answer engines, alongside news recommender systems, increasingly mediate which sources are visible to users, with platform companies and AI providers acting as gatekeepers whose optimisation goals may diverge from journalistic news values.

Filter bubbles and echo chambers

The idea that algorithmic personalisation can narrow users' exposure to information is often expressed through two related concepts. The internet activist Eli Pariser coined the term filter bubble for personalised filters that create a distinct information environment for each user based on predictions about what they will do and want next. The related echo chamber concept, associated with the legal scholar Cass Sunstein, concerns users' growing ability to filter what they see and providers' growing ability to filter information for them, reducing unplanned exposure to views they would not have chosen in advance.

Empirical research has provided limited support for the strong form of the filter bubble hypothesis. A 2015 study of 10.1 million United States Facebook users found that algorithmic ranking reduced exposure to ideologically cross-cutting content by about 15 per cent. Users' own choices about what to click limited such exposure more than the ranking algorithm did. A study of the web-browsing histories of 50,000 United States news readers found that search engines and social networks were associated with a greater average ideological distance between individuals. They were also associated with greater exposure to material from a reader's less preferred side of the political spectrum. The authors judged the overall effects to be modest, noting that most online news consumption came from direct visits to mainstream outlets. A 2018 survey-based study concluded that only a small proportion of people exist within echo chambers, because most use a range of media sources. It found that political interest and diverse media use further reduced the likelihood of ideological enclosure.

Political content and polarisation

Research on whether algorithmic recommendation amplifies political content in a particular ideological direction has produced varying findings across platforms, methodologies, and time periods. Studies of X have identified directional effects, while an experiment on Facebook found limited measurable effects on political attitudes, and a counterfactual-bot study of YouTube found that user preferences played a larger role than recommendations in partisan consumption.

A large-scale study drew on a long-running randomised experiment whose control group included nearly two million daily active Twitter accounts. At the aggregate party level, the ranked timeline amplified the mainstream right more than the mainstream left in six of seven countries. The study also found greater aggregate amplification of right-leaning US news sources, although results varied with the media-bias classification used. It did not identify the cause of the right-leaning asymmetry. At the individual level, amplification was not significantly associated with politicians' party affiliation, and far-left and far-right parties were generally amplified less than centrist parties.

A 2025 sock-puppet audit of X during the 2024 United States presidential election found a different pattern. The audit deployed 120 monitoring accounts and found that both left- and right-leaning accounts received amplified exposure to ideologically aligned content and reduced exposure to opposing viewpoints. Newly created neutral accounts, which followed no one, received a default right-leaning bias in their recommended content. X's algorithm also amplified political commentators and influencers alongside traditional media and political figures, a shift from the patterns observed in earlier studies.

Large-scale experimental studies of Facebook and Instagram during the 2020 United States presidential election found that algorithmic ranking altered the mix of political content users encountered but produced limited measurable effects on political attitudes or polarisation over the study period. A 2025 audit found that engagement-based ranking altered the political content shown and users' immediate evaluations of political groups. The Meta experiments assessed political attitudes, whereas Huszár and colleagues measured political-content amplification.

Research on search engines has examined whether ranking algorithms produce comparable effects. Controlled experiments involving 4,556 undecided voters in the United States and India found that biased search rankings could shift voting preferences by 20 per cent or more, with most participants unaware of the manipulation; the researchers termed this the search engine manipulation effect. An audit of Google's Top Stories box by Daniel Trielli and Nicholas Diakopoulos found that impressions were concentrated among a small number of mainstream outlets and skewed towards left-leaning news sources, which accounted for 62.4 per cent of impressions against 11.3 per cent from right-leaning sources; the authors attributed the imbalance principally to the greater volume of news material published by left-leaning outlets rather than to bias in the algorithm itself. However, a study that tracked real users' Google Search activity during the 2018 and 2020 US elections found that partisan identification had a small and inconsistent relationship with the news sources Google's algorithm showed users, but a larger and more consistent relationship with the sources users chose to click on and engage with. The authors concluded that user choice, rather than algorithmic curation, was the primary driver of exposure to partisan and unreliable news through search.

Mental health and minors

A chronological feed (left) compared with an algorithmically ranked feed (right). Content with high engagement scores is moved to the top; less engaging content is deprioritised

The effects of algorithmic recommendation on young users' mental health have become a subject of policy debate in multiple jurisdictions. A Wall Street Journal investigation found that TikTok's algorithm could narrow recommendations towards material related to self-harm, eating disorders, or drug use within hours of a user showing interest in adjacent content. A 2023 Amnesty International report reached similar conclusions about TikTok's For You feed, arguing that targeted recommendations could rapidly intensify exposure to depressive and self-harm-related material among vulnerable young users.

A member of TikTok's trust and safety team, who spoke to the BBC during 2025, said the company's internal system rated relatively trivial cases involving politicians as a higher priority for review than cases involving harm to teenagers. In one example shown to the BBC, a political figure who had been mocked online was prioritised over a 16-year-old in Iraq who reported that sexualised images purporting to be of her were being shared on the platform. The employee said the company prioritised political cases to maintain relationships with politicians and governments and avoid regulation or bans, rather than because of the severity of the harm reported. TikTok rejected this characterisation, saying that specialist workflows for certain issues did not result in the deprioritisation of child safety cases, which were handled by dedicated teams within parallel review structures.

Social psychologist Shoshana Zuboff situated these engagement-driven systems within a wider critique of platform business models, arguing that user behaviour is extracted as "behavioral surplus" and processed into prediction products, with automated systems increasingly designed to shape behaviour towards guaranteed commercial outcomes.

These concerns have informed legislative activity. The Kids Online Safety Act (KOSA), introduced as S.1409 in 2023, would require covered platforms to allow minors to opt out of personalised recommendation systems and would impose a duty of care concerning harms linked to platform design. The Senate passed KOSA in July 2024. A new KOSA bill, S.1748, was introduced in May 2025. New York's Stop Addictive Feeds Exploitation (SAFE) For Kids Act, signed into law in June 2024, restricts algorithmically personalised feeds for users under 18 unless parental consent is granted. Without consent, minors are to be shown content only from accounts they follow or otherwise select, in a set sequence such as chronological order. The Act takes effect 180 days after the state attorney general finalises implementing rules, which were issued in proposed form in September 2025. In the United Kingdom, the communications regulator Ofcom finalised its Protection of Children Codes of Practice under the Online Safety Act 2023 on 24 April 2025. From 25 July 2025, any provider operating a recommender system that its risk assessment rates as medium or high risk for harmful content must configure the system to filter such content from children's feeds.

State use and control

Research has examined how state actors interact with platform visibility systems, both by producing content designed for algorithmic distribution and by deploying automated accounts to shape what is seen.

A 2025 study by Yingdan Lu and colleagues identified more than 18,000 regime-affiliated accounts on Douyin (the Chinese sister app of TikTok) that posted over five million videos, which the authors characterised as a decentralised state propaganda system. Rather than relying on control of the platform's ranking and recommendation algorithms, which the authors argue cannot on its own reach fragmented audiences, the model mobilised large numbers of government-linked accounts producing high volumes of varied content that flowed between central and local accounts and could then be promoted algorithmically. The authors concluded that this decentralised approach could reach fragmented audiences more effectively than a centralised, top-down system.

In authoritarian contexts, automated accounts can function alongside platform algorithms to shape information visibility. A study of Persian-language Twitter during the first wave of the COVID-19 pandemic found that pro-regime clusters contained a high proportion of bot accounts, with one cluster consisting of 76 per cent automated users. These bots used similar framing strategies to human regime supporters but operated in a coordinated manner to amplify pro-government narratives and suppress dissenting content. Anti-regime communities also contained automated accounts, though their clusters were primarily directed by non-bot users. Research has also examined how users respond to perceived algorithmic suppression of political content. A 2025 study of Palestinian social media users during the May 2021 Sheikh Jarrah events found that activists reported experiencing shadow banning and content demotion on platforms including Instagram and Facebook. In response, they developed circumvention strategies such as altering language, using coded terms, and coordinating posting schedules to maintain visibility within algorithmic ranking systems.

Methods of study

Research on algorithmic amplification is constrained by limited independent access to the internal workings of platform recommendation systems. A 2021 report by the Ada Lovelace Institute identified six technical methods available for auditing such systems: code audits, user surveys, scraping audits, API audits, sock puppet audits, and crowdsourced audits, each with its own limitations and challenges.

Evidence has also come from randomised experiments conducted by the platforms themselves. The Huszár et al. study of Twitter's recommendation algorithm used a long-running randomised experiment maintained by the platform, in which a control group of nearly two million daily active accounts received a reverse-chronological feed rather than an algorithmically ranked one. Meta ran similar experiments in 2020, in which it deactivated algorithmic ranking for randomly selected Facebook and Instagram users during the US presidential election. Platform-run experiments can provide large samples and controlled comparisons, but external scrutiny depends on companies granting access to their systems and data.

A sock-puppet audit: automated accounts with controlled profiles use a platform, and the recommendations returned to each account are recorded and compared

Where platform cooperation is unavailable, researchers have used external methods. Sock-puppet audits create artificial accounts with controlled characteristics to observe what content is recommended to them. The Ye, Luceri, and Ferrara audit deployed 120 sock-puppet accounts across four political orientations on X during the 2024 US presidential election, collecting over nine million recommended posts over six weeks. The method allows researchers to isolate algorithmic behaviour from individual user choices, but the artificial accounts do not interact with content as real users do, raising questions about ecological validity. A 2024 simulation study by Paul Bouchaud and Pedro Ramaciotti found that design choices in sock-puppet audits, including the number of accounts followed and the length of simulated browsing sessions, could alter the conclusions drawn about the same platform.

Data donation studies take a different approach, recruiting real users to install browser extensions that record the recommendations they receive. The approach captures genuine user experience but introduces self-selection bias, since participants may not be representative of the wider user population.

These methodological constraints have informed regulatory responses. The DSA requires very large online platforms, defined as those with more than 45 million monthly active users in the European Union, to provide data access to approved researchers. Academics from the University of Sussex noted that the Online Safety Act allows Ofcom to seek information from categorised services about algorithms affecting the display, promotion, restriction or recommendation of content.

Regulation

Governments have taken markedly different approaches to regulating recommendation systems. The European Union and United Kingdom impose transparency and risk-assessment duties on large platforms. In the United States, no comprehensive federal law has been enacted, though several bills have been introduced. China requires providers to register their algorithms with regulators and to let users disable personalised recommendations.

European Union

The DSA, which became fully applicable to all platforms by 17 February 2024, requires very large online platforms to assess and mitigate systemic risks associated with recommendation systems, including risks to public discourse, fundamental rights, and the mental health of minors. Platforms must offer users at least one recommendation option not based on profiling. Article 27 of the DSA requires transparency about how recommendations are generated, while Articles 34 and 35 impose additional obligations on very large online platforms and search engines. In October 2024, the European Commission issued requests for information to YouTube, Snapchat, and TikTok about the design of their recommender systems and their role in amplifying risks related to elections, civic discourse, and child safety.

United Kingdom

The Online Safety Act 2023 requires providers to assess how algorithms may increase users' exposure to illegal content and children's exposure to harmful content. Providers must then mitigate identified risks through their systems, design and algorithms. Illegal-content duties became enforceable on 17 March 2025. Services likely to be accessed by children had to complete child-risk assessments by 24 July 2025.

In July 2025, the House of Commons Science, Innovation and Technology Committee concluded that the Online Safety Act did not adequately address the algorithmic amplification of legal but harmful content. The committee cited the 2024 Southport riots and recommended that platforms deprioritise content found misleading by fact-checkers. It also said that government lacked accurate, current information about recommendation algorithms and recommended independent research to inform future standards and duties. University of Sussex scholars giving evidence to the committee argued that the Act lacks safety duties focused specifically on the development and adoption of recommendation algorithms. They acknowledged that its broader risk duties can apply to algorithmic systems.

The Online Safety Act's categorisation framework also raised questions about how regulation designed for platforms that use algorithmic recommendation applies to those that do not. In 2025, the Wikimedia Foundation filed a judicial review challenging the categorisation regulations, which it said could place Wikipedia under the Act's strictest tier of obligations. The foundation argued that Category 1 duties, among them user identity verification, would undermine the privacy and safety of its volunteer contributors and expose the encyclopaedia to manipulation and vandalism. The High Court dismissed the challenge in August 2025 but stated that the ruling did not give Ofcom or the government "a green light to implement a regime that would significantly impede Wikipedia's operations", and that the foundation could bring a further challenge if Ofcom wrongly concluded that Wikipedia fell within Category 1.

Ofcom published the register on 30 June 2026. Category 1 designation applies to services that exceed a user-number threshold and use a content recommender system, and eleven services were placed in that category, among them Facebook, Instagram, TikTok, X, and YouTube. Wikipedia was not designated, but was added to a separate list of emerging Category 1 services, which carries no additional duties under the Act. The Wikimedia Foundation said it was relieved by the decision, while noting that Wikipedia could be reassessed at any time.

United States

No federal legislation specifically regulating algorithmic amplification had been enacted as of July 2026. The Filter Bubble Transparency Act, introduced in multiple congressional sessions since 2019, sought to require platforms to offer alternatives to algorithmically ranked feeds. The Senate passed KOSA in July 2024, and a new KOSA bill, S.1748, was introduced in May 2025. In June 2026, the House of Representatives passed the Kids Internet and Digital Safety (KIDS) Act by 267 votes to 117. The bill would require platforms to offer children ways to limit addictive features and to adopt policies addressing harms including sexual exploitation. The Senate had supported stronger standards including a duty of care, setting up a potential conflict between the chambers. At the state level, New York's SAFE For Kids Act (2024) restricts algorithmically personalised feeds for users under 18 without parental consent, and takes effect once the state attorney general finalises implementing rules.

China

The media scholar Jian Xu describes China as the first country in the world to enact and apply laws regulating algorithms and generative artificial intelligence. The Provisions on the Management of Algorithmic Recommendations in Internet Information Service, jointly issued by the Cyberspace Administration of China and three other agencies, took effect on 1 March 2022 and were China's first regulation of algorithms to carry legal force. They require internet platforms to allow users to disable personalised recommendations, and prohibit the use of algorithms to spread illegal or harmful content, to induce addiction to the internet among minors, or to discriminate in pricing against returning customers. Providers whose services carry public-opinion attributes or a capacity for social mobilisation must register their algorithms with the Cyberspace Administration, including details of their data, models, and risk prevention mechanisms. The administration published the first batch of 30 registrations in August 2022, covering most of the major Chinese technology companies, among them Alibaba, Tencent, ByteDance, and Baidu. By April 2023, 262 providers had registered.

Xu has argued that the ideological and political implications of algorithmic applications are the primary concern of Chinese regulators, and that the Cyberspace Administration's lead role reflects this priority. The regulatory framework was developed in three phases: initial post-event penalties against technology companies, followed by ethics guidelines and industry self-discipline pacts, and then binding legislation. Xu noted that the transparency requirements apply only to algorithms used by commercial platforms, not to those used for government decision-making or public administration.

Debate

Academic and policy debate about algorithmic amplification has centred on whether engagement-driven recommendation represents a structural problem in platform design, and on how large its effects are relative to other factors shaping information consumption.

Zuboff has characterised engagement-driven recommendation as part of a broader economic logic in which user attention and behavioural data are extracted and commodified by platform companies. Born and Diaz have argued from a cultural theory perspective that personalisation in recommender systems weakens the shared experiences on which cultural citizenship depends.

The 2024 YouTube counterfactual-bot study found that user preferences were a more significant driver of partisan consumption than the algorithm itself. In separate experiments on Meta's platforms during the 2020 election cycle, algorithmic ranking altered the content mix but had limited measurable effects on attitudes.

The political debate about algorithmic amplification in the United States has been shaped by competing claims. Republican politicians, including Republicans on the House Judiciary Committee, have alleged that platforms suppress conservative viewpoints through content moderation and ranking. Florida and Texas enacted laws in 2021 restricting platforms' content-moderation choices, including on viewpoint grounds. Other researchers have reported that ranked feeds can favour right-leaning political content. The 2021 NYU Stern report said that no trustworthy large-scale study had shown conservative content being removed for ideological reasons. It cited Facebook engagement rankings in which right-leaning pages often performed strongly.

The question of whether governments can regulate algorithmic curation without infringing platforms' speech rights has been tested in the courts. In Moody v. NetChoice (2024), the Supreme Court considered First Amendment challenges to the Florida and Texas laws. The majority discussed Facebook's News Feed and YouTube's homepage on the record before it. It said that selecting, ordering, labelling and removing third-party posts in those feeds involved protected editorial choices. The Court also said that Texas could not justify altering the feeds merely to change the balance of viewpoints. It nevertheless vacated both appellate judgments and remanded the cases because the lower courts had not assessed the full range of the laws' applications required for facial challenges.

A separate strand of the debate concerns whether reducing an item's algorithmic distribution, rather than removing it, offers a way to limit harm without restricting speech. The disinformation researcher Renée DiResta argued that there is no right to algorithmic amplification and that "free speech is not the same as free reach". She framed reduced distribution, or de-amplification, as distinct from removing content.

Proposed technical alternatives have sought to address some of these concerns. An exploratory analysis in the 2025 audit found that reranking collected tweets by users' stated preferences reduced anger, partisanship and out-group animosity, but could increase exposure to ideologically aligned content.