Manipulation, microtargeting, and elections · AI, Ethics, and Society

A useful way to think about AI and elections is not to ask whether the technology enables manipulation that was previously impossible. It usually does not. Microtargeted Facebook ads have been cheap in Nepal for nearly a decade. Voice messages have been forwarded through Viber and WhatsApp for almost as long. Paid influencers have been a feature of every recent election. What AI changes is the cost of producing the content that flows through these existing channels. The pipes were already built. AI turns up the pressure.

The economics, plainly

Imagine a moderately funded campaign — a provincial-level party or a well-resourced independent candidate — in the run-up to a Nepali election. Before AI, the campaign could afford perhaps a dozen variations of a Facebook ad: a single graphic designer working a long week, an ad buyer testing variants against demographics, a translator producing parallel copy in Nepali, Maithili, and a handful of other languages if budget allowed. The variation was bounded by human-hours.

After AI, the same campaign can produce two hundred ads per district, tailored to the local issue mix, in ten languages, in the time it used to take to produce a dozen. Each ad can be personalised at the level of caste-cluster, age-band, or expressed interest from a person’s Facebook history. The bidding cost on Meta’s platform has not fundamentally changed. What has changed is the granularity of the targeting and the volume of the creative.

This is what “microtargeting” actually means in the Nepali context. Not the Cambridge Analytica fever dream of psychological manipulation. The more mundane and more powerful reality of one ad, perfectly calibrated, shown to one slice of voters who will respond to it, repeated thousands of times across the country.

The propagation layer

Producing the content is half the system. Getting it to voters is the other half, and Nepal has several propagation channels that operate largely outside of public scrutiny.

Facebook is the front door. Roughly two-thirds of the country’s adult internet users are on it. Meta’s ad platform allows targeting that is detailed enough to reach specific wards and specific demographic slices. Pages purporting to be local news outlets — many of them legitimate, some of them not — can be quietly purchased or rented for a campaign window and then sold back. The line between organic content and paid content is blurry on a phone screen, and most users do not check it.

TikTok is the second-floor entrance. Short video, vertical, algorithmically personalised, dominant among voters under twenty-five. A clip that works on TikTok can reach hundreds of thousands of impressions without any ad spend, if the algorithm decides to promote it. AI-generated short video is the new frontier here, and almost nothing on TikTok in Nepal is labelled or moderated for political content.

Viber and WhatsApp are the family room. This is where forwarded messages from “my cousin’s friend who works at the ministry” carry the weight of personal recommendation. AI does not need to be sophisticated here. A plain text message, plausibly written, easy to forward, indistinguishable from a real chain of correspondence, is enough. The closed-group nature of these platforms means fact-checkers cannot see what is circulating in time to debunk it.

The voice channel. Robocalls are old; AI-cloned voice robocalls are new. A campaign can now generate a personal-sounding call in the candidate’s own voice, in the recipient’s first language, addressing them by name, in the morning before voting. The cost per call has dropped to fractions of a rupee. The Election Commission has no published procedure for these.

Where the Election Commission stands

As of mid-2026, Nepal does not have a clear, enforced rule requiring disclosure of AI-generated content in political advertising. The Election Commission has issued general guidance about misinformation and digital campaigning, and there is occasional movement toward more specific rules, but the regulatory reality on the ground is closer to “rules exist on paper, enforcement is ad hoc.” A campaign that wanted to flood TikTok with AI-generated clips of opposing candidates appearing to make embarrassing statements would, in practice, face no automatic platform check, no Commission pre-clearance, and a complaint process that runs slower than the election calendar.

This is a gap that civil society organisations — the fact-checkers, the press council, the election observation groups — have started to fill. The work is real, but it is small relative to the volume of content. The honest position is that, in the next general election, Nepali voters will see a lot of AI-generated political material, much of it unlabelled, and the institutional defences will not catch most of it in time.

What actually works as a voter-side defence

Two extreme reactions are common and both are wrong.

The first is digital withdrawal: deciding the information environment is too polluted to trust and disengaging from political content entirely. This sounds principled and is, in effect, a vote for whoever is best at flooding the channels you have abandoned. It is the position of giving up.

The second is uncritical consumption: scrolling, reacting, forwarding, treating each new clip as standalone evidence. This is the position the manipulation campaigns are built to exploit.

The actual workable defence is narrower and less heroic. It looks like this:

Cultivate two or three trusted sources, deliberately. One newsroom whose track record you have personally checked — Kantipur, Setopati, Onlinekhabar, Kathmandu Post, depending on your beat and reading habits. One fact-checker (South Asia Check, for instance, or a regional equivalent). One source of primary material — the candidate’s own speeches and official platform documents, not somebody’s edit of them.

Treat everything else as unverified until your trusted sources agree. A clip going viral on TikTok is not evidence. An accusation circulating in a Viber group is not evidence. A screenshot of an alleged news article you cannot find on the actual outlet’s website is not evidence. Wait. The cost of waiting is low. The cost of being wrong, then forwarding wrong, then being part of the chain that pushed wrong into your relatives’ phones, is high.

Take voice and video especially seriously as unreliable. The thing about a clip of a person saying something is that it is now the easiest type of content to fabricate. Counter-intuitively, the medium you used to find most persuasive is the one that should now persuade you least, in the absence of independent confirmation.

Vote on platforms, not just personalities. The harder a political ad is working to make you angry at one person, the more likely it is engineered content. Specific personalised attacks are cheap to produce now. Substantive policy positions, voting records, and verifiable past actions are not.

A note on hope

None of this is meant to suggest that Nepali elections are about to be decided by AI. They are not. The fundamentals of Nepali politics — local relationships, caste and ethnic patterns, economic conditions, the candidates’ own conduct — remain dominant. AI does not replace these. It adds noise, and the noise is loud enough to matter at the margins. Margins, in democracy, are exactly where elections are decided.

The right posture is sober, not panicked. Build the personal habits. Support the institutions trying to do the work. Argue for clearer rules, and assume meanwhile that the rules will not be there when you need them.

Check your understanding

Quick check

—

A common argument is that Nepal is too small a media market for AI-powered political microtargeting to be a serious concern. Which response best reflects the actual situation?

The argument is correct — international campaign operations focus on larger markets like India and the US, so Nepal is effectively safe.
The argument was once roughly true but is now outdated, because AI has dropped the cost of producing district-tailored, multi-language political content to the point where small markets are now economical to target — domestically funded campaigns and outside actors alike can afford the operation.
The argument is correct because Nepali voters are too sophisticated to be influenced by online content.
The argument is wrong because the Election Commission has strict, well-enforced rules that already prevent AI-driven manipulation.

Quick check

—

What is the most defensible personal strategy for a Nepali voter trying to make sense of political content in 2026?

Disengage from all political content online — the environment is too compromised to be worth following.
Consume everything that appears in your feed, reacting and forwarding freely, and trust that the volume of perspectives will average out to the truth.
Identify two or three trusted sources in advance — one newsroom, one fact-checker, one source of primary material like candidates’ own speeches — and treat anything that has not been confirmed by them as unverified, regardless of how compelling or shareable it looks.
Trust video clips above text-based claims, since video is harder to fake.

What comes next

If trust and elections are where the harms of AI become most visible, governance is where the response either happens or fails to happen. Chapter 6 turns to the road ahead — what kinds of rules, institutions, and public conversations Nepal needs to build before the next wave of capability arrives, and what other countries’ attempts can teach us without dictating to us.