“The most dangerous thing AI does to democracy is not make people believe false things. It is make people unable to trust anything. When any video might be fake, any audio might be synthetic, any article might be AI-generated, and any source might be compromised, the rational response is not to believe the right things more carefully. It is to stop believing anything at all. That epistemic learned helplessness is the democratic crisis.”
Neal Lloyd · Inside The Machine, Day 25In January 2024, robocalls using a convincing synthetic audio reproduction of President Biden’s voice were sent to New Hampshire voters ahead of the primary, telling them not to vote. The calls were traced to a political consultant who was subsequently indicted. The AI-generated audio was compelling enough that multiple recipients initially believed it was real. The incident was, in retrospect, a relatively crude early demonstration of what AI-generated political manipulation could look like. By 2026, the capability has advanced dramatically — video deepfakes are produced in real time by tools accessible to anyone with a consumer GPU, synthetic audio is indistinguishable from real recordings in A/B tests with human listeners, and the personalised persuasion infrastructure that AI enables can target individuals with messages calibrated to their specific psychological profile, drawn from years of social media data. The New Hampshire robocall was a warning. This is Day 25 of Inside The Machine, and today we examine what AI actually does to the information environment that democracy depends on.
What AI Can Do to the Information Environment Right Now
Synthetic media at industrial scale. Video deepfakes have progressed from technically impressive research demonstrations to consumer-accessible tools. Open-source video generation models can produce realistic synthetic footage from a text prompt, or can swap faces in existing video with sufficient realism to fool casual viewers. Audio deepfakes are further along: commercial tools can clone a voice from as little as three seconds of audio and generate unlimited synthetic speech in that voice with prosodic naturalness that passes human detection in blind listening tests at rates significantly above chance. The cost of producing a compelling synthetic media artefact has fallen from “requires a nation-state or major production studio” to “requires a consumer GPU and a few minutes.” The rate of production that is now possible means that synthetic media incidents can be manufactured faster than any fact-checking or verification infrastructure can respond to them.
Personalised persuasion at scale. The combination of large language models with detailed psychological and behavioural profiles drawn from social media data enables persuasion at a granularity that was previously impossible. Political campaigns using AI can generate individualised versions of the same message — calibrated to the specific fears, values, and information environment of each recipient — at the cost of a fraction of a cent per person. This is not hypothetical future technology. It was demonstrated in commercial political advertising tools in the 2024 US election cycle and has been significantly more capable in every election since. The Cambridge Analytica controversy of 2018 was alarming partly because it revealed that personalised psychographic targeting worked. The AI tools available in 2026 are orders of magnitude more capable at the same task.
Flooding and drowning. AI enables the production of content — articles, social media posts, comments, reviews, forum contributions — at a speed and volume that overwhelms the human capacity to assess and respond to it. A coordinated influence operation using AI content generation can produce tens of thousands of pieces of synthetic content per day, across multiple platforms, in multiple languages, from thousands of synthetic accounts. The signal-to-noise ratio of information environments that were already struggling degrades further when the volume of machine-generated content exceeds the capacity of human moderation and detection systems.
Video deepfakes: Consumer GPU accessible, real-time face swap, text-to-video from prompt. Cost: hours and consumer hardware. Detection rate by casual viewers: insufficient to be reliable. Audio deepfakes: 3-second voice clone, unlimited synthetic speech, prosodic naturalness passing blind tests. Cost: commercial tools, minutes. Personalised persuasion: LLM + psychographic profile + social data = individualised political messaging at fraction-of-a-cent per person. Scale: unlimited. Detection: impossible from the recipient’s perspective. Content flooding: AI-generated articles, posts, comments at tens of thousands per day. Multi-platform, multi-language, synthetic account networks. Overwhelms human moderation capacity. Most dangerous capability: not any of the above individually, but the combination of all four simultaneously, in a coordinated influence operation, targeted at a specific election or political outcome.
The Liar’s Dividend and Why the Damage Is Structural
The most technically sophisticated analysis of AI’s threat to democratic information environments identifies a dynamic called the Liar’s Dividend. The argument, developed by legal scholars Bobby Chesney and Danielle Keats Citron, is this: the proliferation of AI-generated synthetic media does not primarily harm democracy by successfully deceiving people into believing false things. It primarily harms democracy by giving real actors — politicians, governments, companies — a plausible deniability excuse for genuine evidence of misconduct. When any video might be a deepfake, a genuine video of a politician committing a crime can be dismissed as “possibly AI-generated.” When any audio might be synthetic, a real recording of a conversation can be denied as “could be a voice clone.” The existence of synthetic media capability creates a “liar’s dividend” — a generalised excuse that genuine wrongdoers can exploit regardless of whether any specific piece of evidence is actually synthetic.
This is the mechanism by which AI most fundamentally threatens democratic accountability. Not by creating specific false beliefs but by degrading the evidential standard that democratic accountability depends on. Courts, journalists, voters, and political institutions make judgments based on evidence. The reliability of evidence — the confidence with which we can say “this happened because this recording shows it” — is the foundation of accountability. AI synthetic media capability attacks that foundation not by being used to create specific false evidence but by making the category of evidence less reliable. The damage is structural, not episodic.
The second epistemic harm is more diffuse but equally serious: what researchers call “epistemic learned helplessness.” When the information environment is sufficiently polluted with synthetic content, personalised misinformation, and coordinated flooding, the rational cognitive response for many people is not to become better at evaluating evidence. It is to give up evaluating evidence — to retreat into trusted identity communities where information is accepted on the basis of who shares it rather than its verifiable accuracy, or to disengage from political information entirely. Both responses are rational given an information environment that cannot be reliably navigated — and both are catastrophic for democratic deliberation, which depends on citizens being willing and able to engage with evidence about public affairs.
The Liar’s Dividend means that even if AI-generated deepfakes are never used to successfully deceive a single voter, the existence of the capability damages democratic accountability by giving real wrongdoers a plausible excuse to dismiss genuine evidence. The damage to democracy is not in the specific fakes. It is in the generalised degradation of the evidentiary standard that accountability depends on.Neal Lloyd · Inside The Machine, Day 25
The Recommendation Engine That Optimised for Engagement and Got Something Else
The relationship between algorithmic recommendation systems and political polarisation is the most studied and most contested empirical question in the AI-and-democracy literature. The intuitive case that social media algorithms cause polarisation is widely believed, frequently cited, and supported by some research. It is also contested by other research that finds the relationship more complex than the simple “algorithms radicalise people” narrative. Getting this right matters because the policy responses to different causal stories look very different.
What the research does consistently support: algorithmic recommendation systems optimised for engagement — watch time, clicks, shares, comments — reliably surface content that provokes strong emotional responses, because strong emotional responses drive engagement. Content that provokes anger, fear, moral outrage, and tribal solidarity is more engaging than content that is accurate, balanced, and nuanced. This is not a bug in the algorithm. It is the algorithm correctly optimising for the metric it was given. The output of optimising for engagement in a political information environment is an information diet weighted toward emotionally activating content, which in a politically polarised society is content that reinforces in-group identity and hostility toward out-groups.
The AI layer that sits on top of recommendation algorithms in 2026 is more sophisticated than the early neural recommendation systems that generated the 2016-era concerns. AI can now model individual user psychology with sufficient precision to predict not just what content a user will engage with but what content will shift their beliefs, increase their partisan identification, and decrease their trust in outgroup sources. Whether these capabilities are being used by major platforms for political manipulation is a question whose answer is not fully public. What is certain is that the capabilities exist, that political actors and foreign influence operations have access to similar capabilities through AI tools, and that the regulatory infrastructure to detect and prevent their use is not adequate to the current capability level.
Honest About the Limits. Honest About What Remains Worth Trying.
Content authentication and provenance. The C2PA (Coalition for Content Provenance and Authenticity) standard, developed by Adobe, Microsoft, Intel, and others and now widely adopted by major camera manufacturers and social platforms, embeds cryptographic provenance information in digital content at creation. Content created by cameras, produced by AI generators, or modified in post-production carries a verifiable record of its origin. This does not prevent the creation of synthetic media. It does create an infrastructure for verifying the provenance of content that chooses to participate in the standard. The challenge: the C2PA is a voluntary standard. Bad actors creating deepfakes for influence operations will not use it. The standard is most useful for distinguishing authentic from synthetic content in contexts where both the creator and the platform are cooperating with the verification infrastructure — journalism, official government communications, electoral advertising. It is not a solution to adversarial synthetic media.
AI disclosure requirements. The most widely enacted AI-and-democracy regulation in the US is the requirement to label AI-generated political advertising. Several states have enacted such requirements, and a narrow federal bill requiring disclosure labels on AI-generated content in political advertising has been widely described as the most likely AI legislation to pass in 2026. Disclosure requirements are meaningful — voters who know they are looking at AI-generated content update their beliefs differently than voters who do not — but they depend on compliance by actors who have strong incentives not to comply.
Platform-level interventions. The evidence on what platform-level interventions actually reduce the spread of political misinformation is more encouraging than the intuitive narrative suggests. Friction — adding steps between exposure to content and sharing it — reduces sharing of inaccurate content in randomised controlled trials. Accuracy nudges — prompts that ask users to consider the accuracy of content before sharing — reduce sharing of misinformation. These interventions do not require platforms to make editorial judgments about content accuracy. They work by slowing down the automated reflexive sharing behaviour that drives viral misinformation. They are also resisted by platforms whose business models depend on maximising engagement and sharing rates.
Civic and media literacy. The least technically sophisticated intervention is also, in some respects, the most important: education about how AI synthetic media works, what its visual and auditory signatures are, and how to apply appropriate scepticism to emotionally compelling content encountered online. This does not require technical training. It requires the kind of basic information hygiene that journalism education has always promoted — check the source, look for corroboration, be sceptical of content that makes you feel strongly — updated for an environment where the fakes are significantly better than they were before AI. The evidence that media literacy education changes behaviour is mixed. It is also the only intervention that operates at the level of individual epistemic agency rather than platform or government control.
There is no technical solution to the AI misinformation problem that does not require political will, institutional capacity, and cultural change that are all currently insufficient to the task. The capabilities that threaten democratic information environments are advancing faster than the defences. The defences that exist are effective enough to make a meaningful difference if deployed at scale. They are not being deployed at scale. The gap between those two sentences is where democratic institutions are being tested right now.Neal Lloyd · Inside The Machine, Day 25
Inside The Machine, Day 25 · June 27 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs Closed
- Day 15The New Cold War
- Day 16Why AI Lies With Confidence
- Day 17AI Is Eating the Power Grid
- Day 18The Age of AI Agents
- Day 19AI Safety Was Never Just Theory
- Day 20The Surveillance Question
- Day 21AI and the Future of Education
- Day 22AI and Your Health
- Day 23What Is AGI and Are We Close?
- Day 24What Is Work For?
- Day 25AI and DemocracyYou are here
