The Information Environment
How algorithmic feeds, the attention economy, and AI-generated content are quietly reshaping what populations believe is true - and what to do about it.
(roughly two-thirds of all internet users)
(higher in the US, UK, Brazil, Philippines; lower in Japan and Germany)
(up from about 18% in 2013)
A note on framing. The information environment has changed more in the last fifteen years than in the previous fifty, and the change has happened faster than most institutions, regulations, or cultural habits could adjust to. The page below tries to walk through what has actually shifted, why, and what the consequences have been - without either the techno-utopian framing of the early internet era or the panic-driven framing that has become more common since. The reality is somewhere in between, and the specifics matter.
The shift since 2010
The information environment that most adults under 50 grew up in - newspapers, broadcast television, magazines, radio, with editors making explicit decisions about what to publish - has been replaced over fifteen years by something fundamentally different. A small number of platform companies now mediate most of what most people see. The platforms describe themselves as ranking content algorithmically rather than editing it, and US Section 230 law treats them this way for liability purposes; in practice they do also moderate (remove, demote, label, restrict) substantial volumes of content under their own policies, and the line between "ranking" and "editing" has been a major legal and political dispute. The algorithmic ranking layer specifically optimises for engagement metrics that turn out to favour content that is emotionally arousing, partisan, novel, or outrage-inducing. The shift was not engineered as a project to reshape public discourse; it was the cumulative result of business decisions about how to make advertising-supported platforms profitable. The consequences for public discourse have been substantial and are still being worked out.
What specifically changed. The rise of Facebook, YouTube, Twitter (now X), Instagram, TikTok, and WhatsApp as primary information sources for billions of people. The collapse of local newspapers - around 2,500 US local papers have closed since 2005, with similar patterns in most rich countries. The rise of partisan-curated news outlets that look like traditional journalism but operate without the editorial constraints that traditional journalism developed over the 20th century. The fragmentation of national audiences into thousands of niche communities each with their own information ecosystems. The emergence of always-on smartphone access that changes the cadence of how people encounter information from a few daily moments into a continuous stream.
What has not changed despite the rhetoric. People still have similar baseline interests in politics, weather, sports, family, and entertainment. The total amount of time spent consuming media has risen modestly but not dramatically. Most people continue to encounter views that broadly align with what they already believe, much as they did with newspaper subscriptions in 1985 - but the algorithmic reinforcement of this is sharper than the editorial-section choice was. Mainstream news organisations still produce most of the original journalism, but the distribution has shifted to platforms that have more leverage over audience reach than the publishers do.
Algorithmic curation and the engagement economy
The single most consequential shift has been the move from chronological feeds (showing posts in the order they appeared) to algorithmically-ranked feeds (showing posts in the order most likely to keep the user engaged). The change happened across most major platforms between roughly 2009 and 2018. The optimisation target - "engagement," typically measured as clicks, likes, comments, shares, and time-on-platform - turns out to systematically favour content with specific characteristics that are not the same as content that is true, important, or constructive.
What systematically wins on engagement-ranked platforms. Content that produces strong emotional response. Content that is novel or surprising. Content that sorts viewers into in-groups and out-groups. Content that promises secret information or hidden truths. Content that triggers outrage at a perceived enemy. Content that produces parasocial connection (the feeling of personally knowing a creator despite never having met them). Content that exploits curiosity gaps. Content that arrives at the right moment of viewer attention or boredom. Each of these is an entirely understandable optimisation target for a platform trying to maximise time-on-app; together they describe a media environment systematically biased toward emotional intensity over careful reasoning.
What this produces at population scale. Substantially more time spent on partisan and outrage-driven content than the underlying population's interests would predict. Faster and wider spread of misinformation than well-sourced information (because well-sourced information is rarely as emotionally arousing as the rumour version). Fragmentation of audiences into communities that develop their own internal vocabularies, frames, and assumed-shared facts that do not survive contact with outside communities. Polarisation at the political level partly because the platforms reward polarisation more than they reward common ground. Each of these effects has been documented in careful research; the magnitudes are contested, but the directions are clear.
The economics of attention
The underlying business model of most platform-driven media is advertising. Advertising is sold against attention. Attention is finite. The platforms therefore compete with each other and with everything else in users' lives for a share of finite attention. The products designed to win this competition tend to be the products best engineered for compulsive use - variable rewards, social validation, fear-of-missing-out triggers, frictionless engagement loops. The business model and the addiction-inducing design are not separate features that platforms have unfortunately combined; they are the same feature.
What this means for public discourse. The most-engaging content gets distributed; the most-distributed content shapes what populations encounter; what populations encounter shapes what they believe and how they feel. The system has a structural tilt toward making people more anxious, more outraged, more attached to identity-based political alignments, and less able to encounter views from outside their networks. The political conversation about "polarisation" is downstream of this; the mental-health story (covered separately on this site) is downstream of this; the trust-in-institutions decline is partly downstream of this.
What is sometimes missed. The platforms are not neutral broadcasters that happen to surface what users prefer. They are systems with specific design choices, made by specific people for specific business reasons, that produce specific information outcomes. The choice to make the algorithm engagement-optimised rather than, say, accuracy-optimised or wellbeing-optimised is itself a choice. Different choices were technically possible. The current system is the result of competition among advertising-supported platforms, which favoured engagement; a different competitive structure (subscription-based, or with regulatory engagement requirements) would have produced different outcomes.
Misinformation, disinformation, and the line between them
Two terms that get used interchangeably and should not. Misinformation is false information shared without intent to deceive - rumours, mistaken claims, viral but inaccurate posts. Disinformation is false information shared deliberately, typically by state actors or organised political operations, to influence opinion. Both spread through the same algorithmic systems; the distinction matters mainly for what to do about them.
What the research shows. Most viral misinformation is shared by ordinary users who genuinely believe what they are sharing. Verified false stories spread further and faster on Twitter than verified true stories did, in the most-cited 2018 MIT study and in subsequent replications. The pattern is partly because false stories tend to be more novel or emotionally striking; partly because the social-network structure rewards emotional content; partly because corrections do not catch up with the original false claims through the same algorithmic distribution.
Disinformation specifically. Russian information operations have been documented in elections in the US, France, the UK, Germany, and elsewhere since at least 2014, with varying levels of effect. Chinese information operations have been documented in Taiwan, Hong Kong, and increasingly elsewhere. Iranian, Saudi, Israeli, and Indian state actors have all been identified running organised influence operations. Western governments also conduct such operations - US Central Command-linked covert accounts targeting Middle East audiences were documented in 2022 disclosures; the UK 77th Brigade and earlier Integrity Initiative pursued comparable work; the State Department's Global Engagement Center coordinated narrative-shaping efforts before its 2024 wind-down. Domestic political actors increasingly use the same techniques against their own populations. The combination is harder to track and harder to push back against than any single actor's operations would be.
What works against misinformation. Source labelling, friction (a "have you read this article" prompt before sharing), prebunking (warning users in advance about manipulation techniques they will encounter), and platform downranking of repeatedly-flagged content have all shown measurable effects in research. Fact-checking and post-hoc corrections have weaker effects on the people who shared the misinformation. The asymmetry - false information spreads cheaply, true information requires effortful verification - is the central problem; reducing the asymmetry, even modestly, has measurable effects on what populations end up believing.
AI-generated content and synthetic media
The rise of generative AI since roughly 2022 has produced a new variable in the information environment that is still being absorbed. AI-generated text, images, audio, and video are now cheap to produce, increasingly difficult to distinguish from human-created content, and entering social media feeds at substantial scale. The implications for what populations encounter and trust are still being worked out.
What is currently visible. Estimates of the share of new content that is AI-generated vary widely by platform, methodology, and content type - independent academic studies (NewsGuard's AI-content-farm tracker, Stanford Internet Observatory work before its 2024 wind-down, similar Oxford Internet Institute studies) put the share of AI-generated text on social media in the high single digits to low double digits and rising; estimates for image and video are lower but rising faster. The platforms' own transparency reports are partial and not directly comparable across companies. Spam farms, content-mill websites, and political-manipulation operations have substantially scaled up since 2023 by using AI to generate the volume of plausible-looking content they previously had to produce by hand. Specific election cycles have seen AI-generated audio of candidates saying things they did not say (including a documented case in the 2024 US primaries where a fake Biden robocall was used in New Hampshire). Image-based manipulation has become trivially easy. The arms race between detection and generation favours generation.
What is harder to assess. How much of this AI-generated content is actually shifting what users believe versus producing noise that users learn to discount. Some careful research suggests that most AI-generated content currently underperforms human-generated content on engagement, partly because users can sometimes identify it and partly because the platforms downrank it when detected. The longer-run question is whether detection keeps up or whether the cost-collapse in producing convincing fake content reaches a point where it overwhelms the verification systems. The trajectory is concerning; the current state is mixed.
What is being tried. Content provenance standards (C2PA) that cryptographically attest to where content came from. Watermarking of AI-generated content (which has been quickly defeated in most attempts). Platform policies requiring disclosure of AI-generated content. EU Digital Services Act requirements for transparency. Several national-level laws on deepfakes, particularly relating to elections and to non-consensual sexual content. None of these is fully sufficient yet; collectively they describe a system trying to catch up to a technology that moves faster than regulation.
How country approaches actually compare
Countries have taken substantially different approaches to managing the information environment, and the variation provides useful data on what specific interventions work. The numbers below are rough.
The takeaway: the choice of regulatory regime substantially shapes what populations encounter. The countries with the strongest regulatory frameworks (EU, UK, Australia) are testing whether platform-level rules can produce measurably better information environments. The countries with the most state-level control (China, Russia, increasingly several others) are testing a different proposition - that state-managed information delivers social stability and resilience to outside manipulation, at costs in free expression that their governments judge acceptable and critics outside the system judge prohibitive. The US, with its Section 230 framework, has been the natural experiment in what minimal regulation produces. The next decade will reveal which approaches actually deliver the outcomes their architects hope for.
The paths from here
The information environment is changing fast enough that any specific path projection is uncertain. Each path below is one realistic shape the next decade could take.
Continued algorithmic dominance with marginal regulation
The platforms continue to dominate distribution. Engagement-optimised algorithms continue to shape what most people see. Regulation produces specific behavioural changes (transparency reports, content-removal procedures) without fundamentally restructuring the system. Polarisation continues at current levels. AI-generated content increases as a share of total content but is partially absorbed by detection systems.
Will it happen? This is the base case. The structural conditions (advertising-supported business models, network-effect-driven platform consolidation, slow regulatory cycles) point this direction. The countries with stronger regulation will diverge from the US, but the underlying dynamics persist.
A meaningful regulatory restructuring
The EU's Digital Services Act and similar frameworks expand globally. Platform liability for algorithmic distribution effects increases. Targeted advertising restrictions tighten. Recommendation systems are required to offer alternatives to engagement-optimised defaults. Algorithmic transparency becomes mandatory at scale. The information environment becomes measurably calmer, at the cost of platform business models that have to be substantially restructured.
Will it happen? Possible. The European trajectory points this direction. US federal regulation remains the binding constraint; without US movement, global regulation produces a fractured environment. Whether the political appetite for such regulation grows depends partly on whether the harms become more visible than they currently are.
AI-generated content overwhelms verification
The cost-collapse of producing convincing fake content outpaces the development of detection and provenance systems. The information environment becomes dominated by content of uncertain origin. Trust in any specific piece of media falls. Populations adapt by trusting only direct interpersonal sources, with substantial implications for political participation, journalism, science communication, and commerce.
Will it happen? A real risk. The technology trajectory favours generation over detection. Provenance standards offer some protection but require near-universal adoption to be effective, which is hard to achieve. The most likely version is partial - some content is reliably attestable, some is not, and the distinction becomes a major axis of how information is consumed.
A shift to subscription and direct-creator economies
Audiences pay creators directly through subscriptions, Substack-like platforms, Patreon, paid newsletters, and similar models. The advertising-supported social media model declines as the dominant distribution channel. Quality long-form content becomes more economically viable. The fragmented ecosystem produces less polarisation but also less shared informational ground.
Will it happen? Already partly happening. Substack, Patreon, paid podcasts, and direct creator economies have grown substantially. Whether they reach the scale needed to materially shift the dominant pattern depends on how many people are willing to pay for information rather than consume it for free with engagement-driven costs. The current pattern is an emerging niche rather than the new dominant model.
Information-literacy education succeeds at population scale
Schools and broader culture develop more effective approaches to teaching critical thinking about information sources. Specific techniques (lateral reading, source verification, identifying manipulation patterns) become as standard a part of education as basic literacy is. Populations develop measurably better defences against manipulation.
Will it happen? Some movement. Several US states have introduced media-literacy education requirements. Finland's curriculum includes specific information-literacy content and has produced measurable improvements in resistance to disinformation. Whether this scales beyond a few high-investment cases depends on educational-system priorities that are themselves contested.
National-level information segregation deepens
The pattern visible since 2014 - countries deciding which platforms operate within their borders, with what rules - intensifies, and the "splinternet" arrives in earnest in multiple directions. China and Russia have built largely separate digital ecosystems with domestic platforms and limited foreign access. The US has restricted Chinese platforms (TikTok force-sale orders, Huawei equipment bans, WeChat executive orders, US Cyber Command operations against foreign-influence infrastructure). The EU's Digital Services Act creates a third regulatory zone with its own rules on data, content, and platform liability that diverge from both US and Chinese frameworks. India has pushed digital-sovereignty policies (IT Rules 2021, UPI as a non-US payment infrastructure, occasional bans on specific Chinese apps). Iran maintains its own internet controls; Vietnam, Indonesia, Saudi Arabia, the UAE, Turkey, and several others assert their own platform rules. The result is multiple regional internet zones operating under different rules rather than a single "global" internet - and cross-border information flow becomes a contested geopolitical variable in every direction.
Will it happen? Already happening. The trend is well-established and is unlikely to reverse. The depth of segregation is the open question - Russia and China have moved furthest toward separate ecosystems; the US-EU-India-rest-of-world picture is more about layered regulatory differences than wholesale separation; but in all cases the assumption of a single open internet that was common in 1995-2010 is gone.
A high-trust information environment emerges through new architecture
Cryptographic content provenance, federated social media protocols (ActivityPub, AT Protocol), peer-verified reputation systems, and other emerging technologies allow building information systems that are structurally less manipulable than the current dominant ones. By 2035 some communities use these systems for serious information consumption while leaving the engagement-optimised systems for entertainment.
Will it happen? Possible but not the base case. Mastodon, Bluesky, Threads, and similar federated alternatives have grown but remain small relative to the dominant platforms. The technology exists; the user adoption requires either sustained network-effect dynamics that are very hard to engineer, or platform regulations that effectively force migration. Neither has yet appeared at the necessary scale.
The realistic forecast is, again, a mix. The base case is continued algorithmic dominance with growing regional regulatory divergence (paths 1 and 6). AI-generated content (path 3) is probably going to be a larger problem than current systems handle well. Subscription-based and direct-creator models (path 4) grow as a parallel ecosystem. Information-literacy education (path 5) helps at the margin. Major regulatory restructuring (path 2) is unlikely without political will that has not yet appeared at scale.
Where serious analysts disagree
The information environment is one of the topics where careful research has produced more disagreement than consensus. Each reading below is held by named scholars worth engaging directly.
The platforms have substantially damaged democratic discourse
The shift to engagement-optimised algorithmic distribution has measurably increased polarisation, undermined trust in institutions, accelerated the spread of misinformation, and weakened the shared informational ground that democratic deliberation requires. The harms are not just side effects - they are predictable consequences of the business model. Significant regulatory intervention is needed and is overdue.
Held by: Yochai Benkler (Harvard), Renée DiResta (Stanford), Frances Haugen (the Facebook whistleblower), and a substantial fraction of platform-governance scholars. Their data on specific harms is robust; the open question is which interventions actually work without producing worse problems.
The harms are real but smaller and more nuanced than the public conversation claims
Careful empirical research has often found smaller effects of platforms on polarisation, voting behaviour, and misinformation belief than the popular conversation assumes. Some of what looks like platform-driven polarisation is reflecting deeper trends that predate the platforms. The "filter bubble" hypothesis in its strong form has been mostly contradicted by empirical work showing more cross-cutting exposure than the framing implied. Getting the policy right requires getting the diagnosis right, and the diagnosis is more nuanced than the headlines suggest.
Held by: a substantial body of careful empirical research from places like Levi Boxell, Matthew Gentzkow, and Jesse Shapiro on polarisation; from Solomon Messing and others at Meta's Social Science One initiative; from various election-effects studies. The research is technically solid; the political conversation has been slow to update on it.
The mental health effects on adolescents are the largest under-discussed harm
Beyond the political and informational concerns, the platforms' effects on adolescent mental health (covered separately on this site) are substantial and under-acted-upon. The smartphones-and-social-media generation has measurably worse mental health on multiple dimensions, and the timing of the decline tracks adoption of the platforms. Whatever else is or is not true about platform effects, this specific subset is robust enough to act on.
Held by: Jonathan Haidt ("The Anxious Generation"), Jean Twenge, and a body of developmental-psychology research. The case has produced specific policy actions (Australia's social-media-for-minors restrictions, US state-level laws). It is contested in its strong form by some critics but the underlying observation is widely accepted.
AI-generated content will reshape the environment more than current platforms have
The generative-AI cost collapse for producing plausible content is a more consequential shift than the algorithmic-curation shift was. Once anyone can produce convincing fake content of anyone for almost no cost, the verification systems that the current information environment depends on may not survive intact. The next ten years of information-environment policy will be substantially about responding to this rather than to the algorithmic-curation issues that have dominated the last ten years.
Held by: Hany Farid (Berkeley, on deepfakes), Sam Gregory (WITNESS), and a community of scholars working on synthetic-media implications. The case is forward-looking and necessarily speculative; the underlying technology trajectory supports it, while the social-adaptation response is harder to project.
The fundamental problem is the business model, not the algorithms
Engagement-optimised algorithms are a consequence of advertising-supported business models. Reform that focuses on algorithm choices without addressing the business model produces marginal effects. The deeper change requires either subscription-based alternatives at scale, public-interest media institutions that do not depend on advertising, or regulatory frameworks that change what platforms can sell. Without that, the structural pull toward engagement maximisation reasserts itself.
Held by: Tim Wu (Columbia, "The Attention Merchants"), Shoshana Zuboff (Harvard, "Surveillance Capitalism"), and a school of political-economy-focused critics. Their analysis is structural rather than tactical; the policy implications are large and politically difficult.
None of these readings is fully right or wrong. What can be said from the available evidence: platforms have produced measurable harms, particularly to adolescent mental health and trust in institutions; the magnitudes of some specific harms are smaller than the public conversation claims; AI-generated content is the next frontier and may dwarf current concerns; the underlying business model is the deepest variable. Useful policy combines specific tactical interventions (transparency, friction, age-appropriate restrictions) with longer-run structural reforms that have not yet been seriously attempted.
What this means for you
The information environment touches every reader's life directly through what you see, what you believe, and what your time and attention go to. A few practical observations:
If you want a calmer information environment
The single highest-leverage move is curating your own sources rather than letting algorithms do it. Three to five news sources you trust, read deliberately, are worth more than the volume of information that algorithmic feeds provide. Subscribe to a few publications across the political spectrum. Read them as you would read a newspaper - on a schedule, not continuously. Disable algorithmic recommendation feeds where possible. The cost of doing this is some inconvenience; the benefit is materially better-informed thinking and substantially less ambient anxiety.
If you have children
The evidence on smartphones and adolescent mental health is strong enough to act on. Delay smartphone access until later in adolescence (roughly age 14 is a defensible threshold given the research). Phone-free time during the day. No phones in bedrooms at night. Parents agreeing collectively to delay smartphones works dramatically better than individual decisions because of the social-network effects. None of this is unusual or anti-technology; it is taking the developmental research seriously.
If you are concerned about AI-generated content
Develop the habit of asking "could this be fake" for emotionally arousing content from unfamiliar sources. Reverse-image search remains a strong tool for visual content. Verifying through multiple independent sources before believing or sharing remains the most reliable defence. Be especially sceptical of content that claims a specific person said or did something dramatic and is shared from accounts you do not directly know. The arms race between detection and generation is uncertain; personal habits of verification are within your control.
If you want to engage politically
The platforms have specific design features that reward emotional intensity over careful argument. Engaging political conversations on the platforms tends to produce more anger and less persuasion than engaging them in person, in print, or through more deliberate channels. The same hour spent at a local civic meeting, a community organisation, or a long thoughtful piece of writing typically produces more political effect than the same hour spent on social media debate. The platforms are designed to make you feel like engaging on them is consequential; the data suggests otherwise.
If you are a citizen voting on platform regulation
The specifics matter more than the slogans. Transparency requirements (about how algorithms work, what content gets amplified, what advertising is targeted to whom) have the strongest evidence base. Liability rules for specific harms (defamation, election manipulation, child safety) have specific track records. Comprehensive "do something about Big Tech" framings often produce regulation that the largest platforms can absorb while smaller competitors cannot, which entrenches the incumbents. Engaging with specific policy proposals on the merits is more useful than the generic political conversation about platforms.
The mechanics behind this
The information-environment story sits on top of three deeper mechanisms covered elsewhere on this site. If the analysis above depends on ideas you want to understand first, these fundamentals make the conversation more legible:
