#19 How to Train Your Dragon?

 

Every major technological change has also changed the way campaigns are run and the structure of the campaign headquarters.

Television created media consultants, and the internet brought digital directors into the headquarters. Once we started leaning on databases, the people working with voter files, CRM systems and digital advertising moved from the edge of the operation to its centre. I still remember what a revolution it was when we set up the first internet department in a Serbian campaign headquarters, back in 2011. Suddenly the building held people whose job description had not existed five years earlier, while the press office carried on untouched.

This text grew out of Miles Maftean’s analysis “Meet the Campaign AI Engineer” (Partisan Hub). His thesis struck me as correct but not sufficiently tested on terrain like ours, so I have extended it with data and with experience from the campaigns of the past year.

Everyone is looking for the dragon in the wrong place

The current debate about AI in campaigns turns mostly on a few obvious things: generating speeches, emails, ad copy, summaries of the programme and daily messages to volunteers. All of that is true, and all of it is work campaigns were doing long before these new tools appeared.

The more serious change is happening on the other side. A growing number of voters no longer go to Google, no longer look for the candidate’s website, no longer compare programmes and no longer read three articles before they form a view. They ask ChatGPT, Gemini or Perplexity to explain a candidate’s profile, sum up a policy or compare the parties. For younger generations, and above all for first-time voters, artificial intelligence is becoming the most important intermediary between the campaign and the electorate.

That opens a question almost nobody in our headquarters is asking: who here is in charge of making sure the machine understands us correctly and, more importantly, answers the voter in a way that suits us?

My assessment is that the answers to those questions will very soon matter as much as hiring a digital director mattered fifteen years ago.

Google rewarded visibility. The dragon rewards consistency

The whole earlier era of digital strategy came down to one goal: making information easier to find in the endless ocean of the internet. SEO, distribution on social networks, paid ads, all of it we did because our success depended on visibility. If the voter looking for a candidate runs into your content first instead of your opponent’s, you hold a real advantage.

Language models, which for the purposes of this text I am calling the dragon, turn that logic upside down.

When a voter asks Google something, they get a list of sources and decide for themselves whom to trust. When they ask ChatGPT, they get one synthesised explanation. What the voter is judging is a single text the dragon has assembled out of those ten sources, and they read it as though it were the source. It looks like a subtle difference, but it changes the whole problem of optimisation. The search engine rewarded visibility. Language models reward consistency.

The dragon takes pieces of your website and tries to write one article by joining what could not be joined until yesterday: programmes, interviews, parliamentary speeches, podcasts, local portals, Wikipedia, state databases and hundreds of other scattered documents. Every public appearance becomes evidence, and every document on the internet becomes a signal.

So campaigns are competing, for the first time, in something they have never measured: how legible they are to a machine. Visibility online is from now on only the first step. A campaign with a logical, orderly and consistent information ecosystem is far easier to interpret than one whose policies lie scattered across PDFs, interviews, posts and pages nobody has updated in years.

The dragon problem is a problem of information architecture

Here, in my view, is where the debate in the text I mentioned at the start went off the rails.

When ChatGPT gets a candidate wrong or leaves out a key policy, headquarters conclude that AI is unreliable. That is fair, the models do still hallucinate, but a large share of the problem we blame on dragons is created long before the voter opens anything with artificial intelligence in it.

Our positions on the same issue turn up in speeches, interviews, press releases, programmes and posts, in different words every time. Biographies differ slightly from platform to platform. The most important media appearances exist only as video, with no transcript. Local media publish a valuable detail that never reaches our official website. A key campaign pledge is announced five times, and nowhere is there one authoritative place where it is explained.

People join all that up easily, because we instinctively know that two different descriptions of the same measure are probably about the same thing. We recognise the contradiction, and we work out which source carries more weight and whom we trust more.

The dragon cannot assume with the same confidence, because it reconstructs the campaign out of whatever it can reach. If that is inconsistent, so is the reconstruction.

The term information architecture traditionally belongs to web development. It is time campaigns took it much more broadly: are the policies presented consistently on every platform, do the biographies confirm one another or drift apart, are speeches and interviews searchable through hashtags, is it clear what the authoritative source is, and can the model reconstruct the campaign’s priorities without guessing. The key is repeating the same positions, in the same words, from the first day of the campaign to the last. Here that work is usually done by the internet department or an outside web agency, when it should be run by the headquarters. Today it is a communications discipline, the one by which the dragon is trained to do what you want.

The dragon trainers are already on the market

If this sounds speculative, look at where campaigns are already spending money.

CampSight is a measurement tool. It keeps putting questions to language models the way a voter would put them, and shows the campaign how it is described, which sources the models lean on and where the holes are. According to the Run for Something Action Fund announcement from June 2026, by launch the tool had analysed more than 300,000 campaign-related chatbot conversations.

Caucus AI does a similar job from another angle. It tracks what ChatGPT, Grok and Gemini say about candidates and elections, stores the answers and pulls out the sources the models cite.

Both tools rest on the same assumption, and that is where their value lies: what AI tells voters is something you measure, revise and repair, exactly like polling or media monitoring.

That the problem is not only theoretical is clear from research by the Dewey Square Group. They called the finding an inverted funnel, because the outlets with the highest factual reliability are the strictest in blocking AI crawlers, the automated programmes that collect website content for the dragons, while smaller and less reliable sites stay wide open. It is almost a rule: as the reliability of an outlet rises, free access to it falls. For a campaign that means a completely new strategic question. Next to the question of whether authoritative information about you exists online comes a second one: can the dragons get to it at all.

Fortunately, once we recognise a problem we can also solve it. In July 2026 The New York Times wrote about Dustin Loyd, a Democratic candidate for the Missouri House of Representatives, about whom AI knew almost nothing, least of all the thing that was his central theme, support for small businesses. His team published a structured Q&A page and expanded the content. When people put the same questions to the dragon again, they got the answers the campaign wanted, because the answers inside the model had changed.

Hungary 2026: why this is not only an American story

For us in the region, the most important finding does not come from America.

The organisation Liberties tested ChatGPT and Gemini during the 2026 Hungarian parliamentary elections. The methodology was simple: using positions taken from the Voksmonitor application they built five fictional voter profiles, one for each party on the national list, and ran every profile ten times through both models, for both types of question.

The result: ChatGPT did not recommend Tisza in 90% of cases in which it was given a detailed profile of a voter who, according to research, would vote for Tisza. In the percentage-match test, Tisza scored in only 2% of answers. Such voters were instead routinely sent towards small parties with no realistic chance of clearing the threshold. In 96% of answers both models named parties that were not on the ballot at all.

Tisza won that election convincingly, so the researchers are explicit that they do not claim this affected the result, but they do give the most likely explanation: Tisza became politically relevant only in 2024, and the dragons were trained on older data.

Remember that sentence. It describes every new party, every new coalition and every new candidate in the region. If a model systematically fails to recognise a political actor that appeared recently, that is a structural problem, and it will repeat itself at every election that follows.

Why training a dragon is harder here than in America

To be completely clear: this is my assessment from practice, with no research behind it, but to anyone who has spent long enough in campaigns in this region the logic is fairly obvious. I am ready to take any criticism on it, and equally any advice on what to do next from people who know this field far better than I do.

The coalition problem. Explaining the difference between two parties in a two-party system, or in an election that works like a referendum, is a simple task. Explaining the ideological differences between six coalition partners who were attacking each other until yesterday, plus regional lists and citizens’ groups, is another job altogether. Those inconsistencies leave the dragon completely bewildered.

A small language and no corpus. The volume of quality political content in Serbian, Croatian, Macedonian or Albanian is incomparably smaller than in English. Less data means more weight on every single source, including the ones working against us, so the error is bigger.

Two alphabets. A candidate’s name written in Cyrillic, in Latin script, with diacritics and without them usually works for people, but AI often sees four different entities. If your website, Wikipedia and the media reports on the portals do not use consistent forms of your name and surname, or of your party and coalition, you are splitting your own identity into several pieces.

The inverted funnel is steeper here. In a media environment where the tabloids have the biggest online footprint and serious newsrooms have the smallest budget and the strictest subscribers-only policy, the Dewey Square Group finding does not weaken. It gets stronger.

What this means in practice

To begin with, one person in the headquarters with a clear job description is enough:

Test yourself. Put ten questions to the models that a voter would ask about your candidate. Keep the answers, and that is your baseline measurement.

Write one message for every key issue. If your main policy is buried on page 47 of the programme, and in a PDF you cannot even copy from, then for an AI model that information does not exist.

Publish a Q&A. Clearly, in text, on the website. It is the only content you fully control, and it will serve as food for your dragon.

Transcribe your appearances. For an AI model, a video without a transcript does not exist.

Standardise the biography. The same data, the same form of the name, the same script, everywhere.

Check Wikipedia. Your job there is accuracy and clean sourcing.

Check whether AI models can read your website, because a site that is closed to the dragon is a site that does not exist for it.

Once you have done all this, repeat the measurement a month later with the same ten questions.

None of this is expensive. It only takes someone who is responsible for it.

The debate about AI in politics is mostly conducted around productivity: how much faster, how much cheaper, how many hours saved. Those are useful questions, but they will not define the long-term effect.

The essential change is that campaigns are starting to communicate with a completely new audience, with the machine itself.

The advantage over the next decade will probably go to campaigns that grasp that the competition has widened. Next to the voter who has to be persuaded stands the system that explains politics to that voter, and it has to understand them better than anyone else. The volume of content produced with AI will be a side issue.

If you want a dragon of your own, and you want it trained, get in touch. We know how, and we have good trainers.

SOURCES

•          Miles Maftean, “Meet the Campaign AI Engineer”, Partisan Hub — https://hub.partisan.community/m/news/meet-the-campaign-ai-engineer/fbc6430c-c39e-4190-a8af-cd378f9ed779

•          Run for Something Action Fund, announcement of the CampSight launch (June 2026) — https://runforsomething.net/rfs-press/rfsaf-launches-campsight/

•          Caucus AI — https://caucus-ai.com/about

•          PSG Consulting / Innovating for the Public Good, “AI Large Language Model Training: The Potential Risks of Ideological Skewing”, Dewey Square Group research, 24 February 2026 — https://www.psgconsulting.com/research-publications/potential-risks-of-ideological-skewing