SWITCHED ON
The daily technology series nobody asked for but everyone needed
The Feed and the Vote: Technology and Democracy
Democracy depends on a shared, broadly reliable information environment. The technology shaping that environment was not built with democracy as a design requirement. The collision was inevitable. We are living inside it.
Democracy assumes that citizens can access broadly accurate information, form views through reasoned engagement with that information and with each other, and change their minds when presented with compelling evidence. Every one of those assumptions is under direct technological pressure right now, simultaneously, from multiple directions, and the pressure is increasing.
Yesterday we covered the future of work in an AI economy — the historical reassurance about automation and why it might not hold this time, the documented displacement already underway in white-collar professions, the uneven geography of disruption, and the policy responses that range from genuinely promising to mostly theatre. Today we are tackling something that underlies the legitimacy of every policy response we have discussed in this series: democracy itself, and what technology is doing to the information environment that democratic self-governance depends on. Disinformation at scale, deepfakes in electoral contexts, the documented relationship between social media and polarisation, and the uncomfortable question of whether the information ecosystem we have built is compatible with the kind of collective deliberation democracy requires.
01 — The Scale Problem
Disinformation has always existed. Propaganda is not a digital invention. What has changed is scale, speed, and targeting precision, and the combination of these three factors represents a genuine qualitative shift rather than merely a quantitative one.
A piece of disinformation in the pre-digital era required physical distribution — printing, broadcasting, postal delivery — each of which imposed cost and friction that limited reach and allowed time for counter-narratives to develop. A piece of disinformation on a modern social media platform can reach millions of people within hours, be algorithmically amplified if it generates engagement, and be precisely targeted at the demographic and psychographic profiles most likely to believe and share it. The 2016 US presidential election and the UK's Brexit referendum both featured documented foreign interference operations using social media advertising targeted with granular precision at voters identified as persuadable on specific issues. The Cambridge Analytica scandal — in which voter data harvested from Facebook without adequate consent was used to build psychological profiles for targeted political advertising — was a watershed moment in public understanding of how this targeting actually worked, even though its specific electoral impact remains contested among researchers.
The old model of propaganda required convincing a mass audience of a single message. The new model requires convincing a million different audiences of a million different messages, each precisely calibrated to what that specific audience already wants to believe. This is a fundamentally harder problem to counter, because there is no single false claim to fact-check — there are millions of micro-targeted variations.
02 — Deepfakes and the 2024-2026 Election Cycle
We covered synthetic media's epistemic implications in Season One. The 2024 global election supermarket — more voters went to the polls in 2024 than in any previous year in human history, across the US, the EU, India, the UK, and dozens of other countries — provided the first large-scale real-world test of AI-generated content's electoral impact, and the results were more complicated than either the most alarmed or the most dismissive predictions suggested.
Documented deepfake incidents occurred across multiple elections: a robocall impersonating Joe Biden's voice urging New Hampshire Democratic primary voters to stay home, AI-generated images and videos circulated in Indian and Indonesian elections, and synthetic audio purporting to show candidates making damaging statements that they never made, in contexts ranging from Slovakia to Bangladesh. The technical sophistication required to create a convincing deepfake has fallen dramatically — tools now exist that allow voice cloning from minutes of audio and video generation from text prompts, accessible to anyone with modest technical skill and no specialised equipment.
What the evidence does not clearly show is that deepfakes single-handedly changed any major election outcome. The more significant and harder-to-counter effect was the erosion of trust in authentic media — the liar's dividend we discussed in Season One, in which the mere existence of convincing synthetic media gives bad actors plausible deniability for authentic evidence of wrongdoing, and gives voters generalised reason to distrust video and audio evidence regardless of its actual provenance. The damage deepfakes do to democracy may be less about any specific fake fooling people and more about making everyone permanently uncertain about what is real, which is corrosive to the basic possibility of shared factual ground that democratic debate requires.
03 — What the Polarisation Research Actually Shows
The claim that social media causes political polarisation is widely repeated and considerably more contested among researchers than the popular discourse suggests. The honest summary of a genuinely complicated research literature: social media is probably not the primary cause of polarisation, which has structural roots in residential sorting, partisan media, and political realignment that predate platforms like Facebook and Twitter by decades. But social media plausibly accelerates, amplifies, and changes the character of polarisation in ways that are documented and significant.
The mechanism most strongly supported by research is not direct persuasion — social media does not appear to be highly effective at changing people's fundamental political views through exposure to opposing content, and some studies have found that cross-partisan exposure can actually increase animosity rather than building understanding, a finding that complicates the simple "more exposure to other views fixes polarisation" intuition. The more robust finding concerns affective polarisation — not disagreeing about policy, but actively disliking and distrusting people on the other side as people. Social media's tendency to surface the most extreme, performative, and tribal expressions of political identity, because these generate the most engagement, appears to increase the sense that the other political side is not just wrong but malicious, stupid, or dangerous. This is a different and arguably more corrosive phenomenon than policy disagreement, because it erodes the basic civic trust that allows democratic compromise to function.
Internal Meta research, surfaced through the Facebook Papers in 2021, showed the company's own researchers identifying that its algorithms amplified divisive content and that internal proposals to reduce this amplification were deprioritised due to concerns about engagement metrics and the political sensitivity of being seen to suppress particular viewpoints. This is the clearest documented instance of a major platform's own research confirming the amplification dynamic that external researchers had long suspected, and choosing not to act on it at the scale the research suggested was warranted.
04 — The Regulatory and Platform Response
Platform responses to election-related disinformation and synthetic media have evolved substantially since 2016, though their adequacy remains contested. Meta, Google, and TikTok have all implemented AI content labelling requirements, fact-checking partnerships, and political advertising transparency rules. The EU's Digital Services Act, which we covered in the context of platform regulation more broadly, includes specific obligations around systemic risk to electoral processes for very large platforms, requiring risk assessments and mitigation measures around election periods.
The honest assessment of these measures is that they have meaningfully reduced the most blatant and easily detectable forms of coordinated inauthentic behaviour and disinformation campaigns, while doing considerably less to address the more diffuse, organic amplification of divisive and emotionally activating content that does not violate any specific policy but nonetheless degrades the quality of democratic discourse. Content moderation at the scale of billions of posts is an extraordinarily difficult problem, and the platforms have made real investments in addressing it. The fundamental business model tension — that engagement-optimised systems tend to reward exactly the content that is most corrosive to healthy democratic discourse — has not been resolved by any of the measures implemented to date, because resolving it would require changing the business model rather than adjusting its outputs at the margin.
05 — What Citizens Can Actually Do
The individual-level response to this landscape involves the same tension we encountered in the surveillance economy episode: meaningful structural change requires regulatory and platform-level intervention that individual behaviour cannot substitute for, but individual habits do have some real effect at the margin and are worth cultivating regardless.
Media literacy — understanding how algorithmic amplification works, recognising emotionally manipulative content design, checking primary sources before sharing, and maintaining healthy scepticism toward content that confirms existing beliefs with unusual cleanliness — measurably reduces susceptibility to disinformation in controlled studies, though the effect sizes are modest and the people most in need of these skills are often the least likely to seek them out voluntarily. Diversifying information sources beyond algorithmically curated feeds — actively seeking out primary reporting, following diverse and credible sources directly rather than relying on what a feed surfaces — reduces exposure to the most extreme amplified content, though it requires more effort than most people are willing to sustain.
The structural fix that would matter most — changing the engagement-optimisation incentives that reward divisive content — is not something individual media literacy can deliver. It requires either regulatory intervention that changes platform incentives, business model innovation that decouples platform revenue from engagement maximisation, or sustained public pressure significant enough to make the reputational and regulatory cost of the status quo exceed its commercial benefit. None of these are imminent. All of them are more important than any individual habit change, however worthwhile that habit change might be on its own terms.
Tomorrow we are heading somewhere considerably more tactile after a heavy run of policy and politics — wearable technology and the quantified self. Smart rings, continuous glucose monitors for non-diabetics, the explosion of biometric tracking, and whether knowing everything about your own biology in real time is actually making anyone healthier or just more anxious. See you then.
Switched On is a daily technology series covering the ideas, systems, and arguments shaping the digital world. Opinionated. Witty. Occasionally wrong. Always worth the argument.
