Just a few years ago, a voter wanting to check a candidate for mayor had to visit their website, sift through media (including traditional outlets), watch debates, ask friends, or scroll through several pages of Google results. Today, they can often do something much simpler – ask their favourite chat (a large language model).
They don’t even need to… know any names. They don’t need to know who belongs to which committee. They don’t have to follow press conferences. Generally, they don’t have to do much. But they can. They can ask: “who in Kraków has the best transport programme?”, “which candidate is associated with Nowa Huta?”, “who wants to change the Clean Transport Zone?”, “does the PiS candidate have local government experience?”, “who is specifically addressing the cost of living in this election?”.
And they will get an answer.
Not a list of links. Not a classic search result. Not a neutral document database. They will receive a synthesised description of the political landscape, constructed by the LLM based on what the model finds, remembers, interprets, considers important, and arranges in an appropriate hierarchy. Tailored for the user who has, in part, “raised” their own “Tamagotchi” from the third decade (how does that sound!) of the 21st century. Only they’re not feeding it or cleaning it by pressing buttons; they’re tossing in bits of themselves that reveal their habits.
This is a new layer of election campaigning. Quiet, private, difficult to monitor and – in local elections – potentially very significant.
Kraków as a Laboratory for Elections in the Age of GenAI
Kraków is an excellent place to observe this change in action. It’s not a small municipality, but it’s also not a nationwide campaign where every candidate is constantly present in the mainstream media. According to GUS data, at the end of 2025, Kraków had 816,614 residents. It’s a large, complex urban organism: with a city centre, Nowa Huta, peripheral districts, universities, tourism, business, transport, conflicts over green spaces, spatial planning, municipal service prices, and city management. source: Kraków in Numbers
Additionally, there is a unique political context. In the local referendum on 24 May 2026, the voter turnout for the vote on the recall of the mayor of Kraków was 29.99% — enough for the referendum to be valid and decisive. In the parallel vote regarding the recall of the City Council, the turnout was 29.97%, meaning the statutory threshold was not met. The difference is seemingly minimal, but the political consequences are entirely different. source: City of Kraków
Kraków also has recent experience of very close competition. In the second round of the presidential elections in 2024, Aleksander Miszalski received 51.04% of the votes, while Łukasz Gibała received 48.96%. According to reports based on PKW data, the difference was 5,434 votes. source: Rzeczpospolita
These are numbers that warrant caution regarding every new source of informational influence. Not because a chatbot “will choose the mayor of Kraków.” That’s too strong a claim. But because, in a campaign where a few thousand votes can change the outcome, it matters who is visible, who is overlooked, what they are associated with, and how they are described in the responses generated by artificial intelligence, which users are increasingly turning to.
The Voter Doesn’t Just Search. The Voter Converses
The most significant change is not that AI can generate an ad, meme, or deepfake. While that certainly matters, it’s already a well-recognised topic. There’s a lot of discussion about it, and there are campaigns – some more or less social, some more or less funded by specific electoral committees.
A more interesting and less obvious change is that LLMs are becoming private informational advisors. A voter may not ask: “what is Michał Drewnicki’s programme?” They may not even remember that name. Instead, they might ask: “who in Kraków has experience in local government?”, “which candidate talks about Nowa Huta?”, “who has a specific, clear stance on the SCT?”, “is the PiS candidate in Kraków just a party member, or do they have local experience?”.
Such questions are much closer to the real decision-making process. People rarely compare entire programmes from start to finish. (By the way… which party in 2024 clearly described its election programme instead of riding the wave of changing polls, rally cries, and social media noise?) More often, they seek answers to their own problems: commuting, prices, green spaces, schools, pavements, parking, construction outside their window, a sense of chaos in the office, or a lack of influence over city decisions.
Here, large language models begin to act as a new intermediary. They not only provide information. They organise the scene. They select which candidates to mention. They decide which facts to consider significant. They condense complex contexts into a few paragraphs. And they often do this in a way that we won’t see in classic media monitoring, SEO, or social media analysis. Thus, it can be inferred that polling firms and their “misses” will increasingly become one of the main topics of commentary after exit polls.
This is No Longer a Technological Niche
If anyone assumes that “chatbots” are still a toy for students and the tech industry, the data quickly cools that view. According to a Gemius/PBI report, in June 2025, over 9.3 million real users in Poland were using ChatGPT. This represented 31.4% of internet users and 28.6% of the population aged 7–75. The report also indicated that among ChatGPT users, there is an overrepresentation of individuals under 35, with the average usage time in the 25–34 age group being 2 hours and 42 minutes in June. source: Gemius/PBI
On a European scale, Eurostat reported that in 2025, 32.7% of EU residents aged 16–74 were using generative AI tools. In the 16–24 age group, this percentage was already 63.8%. source: Eurostat
This is significant, because younger voters are also a group more inclined to use new informational tools and a group that often has less stable turnout in local elections. There’s no need to assume a mass transition of the entire campaign to AI-supported systems. It’s enough to notice that for a significant portion of users, conversing with a chatbot is becoming one of the natural ways to organise information.
AI as a Tool for News, Politics, and Decisions
Data from the Reuters Institute shows that AI chatbots are already being used for information consumption, although they do not yet dominate. In 2026, 10% of respondents across 45 markets reported weekly use of AI chatbots for news, up from 7% the previous year. Even more interesting is how people use them: 42% of news chatbot users ask follow-up questions, 35% use them for the latest information, 34% for summarising, 30% for simplifying complex topics, and 33% for assessing the credibility of sources. source: Reuters Institute Digital News Report
This is almost a ready description of voter behaviour in a local campaign. “Explain to me what the Clean Transport Zone is about.” “Summarise the differences between the candidates.” “Who is credible on transport issues?” “Does this candidate really have local government experience?” “What sources confirm their statements?”
At this point, AI ceases to be merely a tool for writing texts. It becomes an interface to public reality.
The Strongest Warning Signal – Voters Are Already Asking GenAI About Elections
One of the most interesting figures comes from a study on the 2024 UK parliamentary elections. A representative survey of 2,499 adults showed that in the week leading up to the elections, 32% of chatbot users (13% of all eligible voters) used conversational AI to seek information directly related to their voting decision. source: arXiv, UK study 2024
This is not a marginal detail. It is a signal that chatbots are entering the heart of the electoral process: not as an abstract technology, but as a tool used when voters are making decisions, organising arguments, or trying to understand the political landscape. Often just before entering the polling station.
Importantly, the authors of this study do not draw a simple alarmist conclusion. In a series of experiments involving 2,858 participants, they found that using chatbots did not worsen political knowledge; on the contrary, it increased it to a similar extent as traditional internet searches. source: AI Security Institute
And that’s why the topic is more interesting than a simple tale of danger. Time for a truism. I’ll even bold it to make it more eye-catching. No need to thank me…
LLMs can help voters better understand politics. But they can also confuse, omit, simplify, misidentify candidates, or build specific interpretative frames.
The Other Side – Chatbot Responses Can Be Flawed
The problem is that model responses appear organised, confident, and complete, even when they contain gaps. You know… like that future engineer (if fate and professors allow) from AGH you met at a student party, who will stubbornly defend a position that wouldn’t even have entered the discussion three beers ago ;)
A study by EBU and BBC covered over 3,000 responses generated by four AI assistants (ChatGPT, Copilot, Gemini, and Perplexity) in 14 languages. 45% of responses contained at least one significant issue, 31% had serious problems with sources, and 20% contained serious accuracy issues, including outdated or hallucinated information. source: EBU/BBC
In local elections, this risk may be greater than in a nationwide campaign. Local sources are more dispersed. Candidates may be (and are, as we will soon prove) less well-known. The context changes more rapidly. Names from the previous cycle may mix with new candidates. Programmes may be published in stages. (if they are created at all, but I’ve already mentioned that and won’t poke any more jabs… for now) And user questions are often short, colloquial, and imprecise.
With a national leader, the model usually has plenty of data. With a local candidate for mayor of Kraków, it must piece together a picture from the BIP, local media, the candidate’s website, social media posts, polls, reports from conferences, and current events. This creates ideal conditions for seemingly minor but politically significant errors: confusing roles, omitting competitors, assigning outdated candidacies, giving someone too narrow a label, or basing responses on sources from previous elections.
The Most Important Twist: GenAI Doesn’t Have to Lie to Influence
In discussions about AI and elections, too much attention is focused on “fake news.” Meanwhile, for a local campaign, something subtler may be equally important: representation.
The model may not provide false information. It may simply describe the candidate mainly through their party affiliation, omitting their local government experience. It may mention them when asked about PiS but not when asked about transport. It may write about SCT but skip the topic of public transport. It may answer a question about Nowa Huta without indicating the person who builds part of their communication around ties to that part of the city. It may place the candidate at the end of the list, even though they are formally one of the significant participants in the race.
This doesn’t have to be a “mistake” in the simple sense. It can be a consequence of the hierarchy of sources, data freshness, information availability, the way a question is phrased, and the mechanics of the response generated by the model.
In traditional SEO, one fought for position in search results. In the world of LLMs, it becomes increasingly important to ask: does the candidate even appear in the response, under what questions do they appear, what are they associated with, and who are they compared to.
This mechanism is clearly visible in Michał Drewnicki’s study (discussed in more detail later in the text). In 250 responses from the deep dive study, models mentioned the candidate in 87.6% of cases when the user provided their name, but only in 5.0% of cases when the question did not include a name and concerned an issue, category of candidates, or urban topic. In other words: recognition by name does not necessarily mean thematic visibility.
What If the Response Not Only Informs but Also Shifts Opinion?
Here, a second key set of data emerges. Research described by Cornell showed that a brief conversation with a chatbot can significantly shift political opinions. In experiments conducted in four countries, LLM-based chatbots shifted opposition voters' preferences by 10 percentage points or more in many cases. In experiments in Canada and Poland, the effect was around 10 percentage points, while in one study, the most persuasively optimised model shifted opposition voters' opinions by 25 percentage points. source: Cornell Chronicle
This must be said cautiously. These were controlled experiments, not proof that chatbots will decide real elections. Participants knew they were talking to AI, and the direction of persuasion was randomised. The authors and commentators themselves emphasised the limitations of such studies and the difference between experimental conditions and real campaigns. source: Nature Asia
But one conclusion is hard to ignore. It goes something like this: model responses can be persuasive not because they are emotional, aggressive, or manipulative in the classic sense. According to researchers, their strength often stemmed from generating many assertions, arguments, and seemingly factual justifications. Cornell emphasised that when the models' ability to use facts was restricted, their persuasiveness decreased; at the same time, more persuasive models tended to be less accurate. source: Cornell Chronicle
This is the crux of the problem in a local campaign. A voter may receive a calm, reasoned, well-sounding response devoid of party tone. Yet that response may still reinforce a specific image of the candidate.
The Example of Kraków: Michał Drewnicki in LLM Responses
In this context, Michał Drewnicki's study, the PiS candidate for mayor of Kraków, serves as a good example of what needs to be measured in local politics.
It’s not just about asking: “does GenAI know the candidate's name?”. That’s the simplest level. Much more interesting are the deeper questions:
Do the models correctly identify Michał Drewnicki as the PiS candidate in the early elections in Kraków?
Do they recognise his public roles – city councillor and deputy chair of the Kraków City Council?
Do they distinguish the current electoral context from the 2024 local elections?
Do they associate him solely with PiS, or also with local government experience?
Does he appear in responses to questions that do not include his name but relate to topics present in his public profile: communication, SCT, Nowa Huta, spatial planning, cost of living, relations between the office and residents?
Can the models differentiate between official information, media reports, campaign declarations, and their own interpretations?
The study was conducted by the modest author of this text on 03/07/2026.
Using our proprietary tool Semantio, I analysed 250 responses regarding Michał Drewnicki in the context of the presidential elections in Kraków. The material is the result of an analysis covering 50 unique scenarios, launched in five systems: ChatGPT, Gemini, Grok, DeepSeek, and Google Overview. Each system responded to the 50 scenarios posed. The scenarios were divided according to the stage of the intention funnel: 80 responses at the awareness stage, 85 at the consideration stage, and 85 at the decision stage. Questions containing the candidate's name and problem-based questions without a name were analysed separately.
The strongest result concerns the difference between recognition by name and spontaneous visibility. In the entire material, there were 170 responses to questions containing Michał Drewnicki's name and 80 responses to questions without a name. When the user provided the candidate's name (the prompt scenario included the name “Drewnicki”), the models mentioned Drewnicki in 149 out of 170 responses, or 87.6% of cases. When the question did not include a name and concerned an issue, category of candidates, or urban topic, Drewnicki appeared in only 4 out of 80 responses, or 5.0% of cases.
In plain terms: models can describe the candidate when the user already knows who they are asking about, but they are significantly less successful at independently linking him to the city's issues.
The data also shows that visibility is not evenly distributed among systems. All 4 spontaneous mentions of Drewnicki in questions without a name came from Google Overview. In the other systems (ChatGPT, Gemini, Grok, and DeepSeek), the candidate did not appear even once in such questions. This is important because it highlights “in numbers” that there is no single, universal “visibility in AI”. Each system can build a different map of the political scene, dependent on sources, data freshness, search mechanics, and the way responses are generated.
Indeed, I couldn’t resist including this photo in the context of SCT ;)
The clearest hint of thematic visibility appeared in questions about transport, public transport, tickets, mobility, and the Clean Transport Zone. In questions without a name concerning this area, Drewnicki appeared in 4 out of 30 responses, or 13.3% of cases. This is still a low result, but significant compared to other topics: questions about local government experience, Nowa Huta, districts, spatial planning, or green spaces did not trigger his name as effectively. From the perspective of a local campaign, this is an important difference: the model may accurately describe the problem in Kraków, but it may not necessarily show the voter which candidate is trying to politically address that problem.
In 70 out of 250 responses, or 28.0% of the entire dataset, hallucination alerts were marked. The risk of error did not disappear after providing a name: in questions with a name, an alert appeared in 50 out of 170 responses (29.4%), while in questions without a name, it appeared in 20 out of 80 responses (25.0%). Most often, these were contextual issues, such as mixing the 2026 elections with the 2024 elections, incorrect public roles, incorrect political affiliation, erroneous or suspicious URLs, unverified programme details, and even confusing Kraków with Warsaw (that’s unforgivable in the City of Krak!). In a local campaign, such minor errors may be more likely than spectacular “fakes”, and thus much harder to catch, as they often occur in responses that sound calm and reasoned. Where have we seen this before?…
Differences between providers (another beautiful word from the Bug River) were pronounced. Google Overview mentioned Drewnicki most frequently and had the lowest rate of hallucination alerts: 37 mentions in 50 responses (74.0%) and 5 alerts (10.0%). DeepSeek mentioned the candidate in 33 out of 50 responses (66.0%), but simultaneously had the highest share of alerts: 31 out of 50 responses (62.0%). ChatGPT mentioned Drewnicki in 30 out of 50 responses (60.0%) and had 8 alerts (16.0%). Grok mentioned him in 27 out of 50 responses (54.0%) and had 16 alerts (32.0%). Gemini mentioned the candidate in 26 out of 50 responses (52.0%) and had 10 alerts (20.0%). This shows that greater visibility in AI does not always mean higher quality representation.
Sources also arranged interestingly. In the entire dataset, 676 source links were identified. The most frequently appearing domains were: bip.krakow.pl (90 times), facebook.com (71 times), krakow.pl (38 times), youtube.com (29 times), radiokrakow.pl (26 times), lovekrakow.pl (23 times), drewnicki.pl (22 times) and ztp.krakow.pl (22 times). The candidate's official domain was present, but it was certainly far from dominating. The image of Drewnicki in AI was also constructed by the BIP, local media, city sources, Facebook, YouTube, and other intermediary domains.
At the same time, in 115 out of 250 responses, there were no source links at all, which constitutes 46.0% of the entire material. Differences between systems were significant: Google Overview provided links in every response, ChatGPT in 43 out of 50, DeepSeek in 31 out of 50, Grok in 10 out of 50, and Gemini only in 1 out of 50 responses. This has electoral significance – a response without a source may sound credible, but the user lacks a quick way to verify where the model obtained information about the candidate, their role, programme, or the election context.
In LLM responses, competition was also not understood solely as a list of formal electoral rivals. In the competitive field, the most frequently mentioned were Aleksander Miszalski (53 times) and Łukasz Gibała (50 times), but also visible were Andrzej Kulig (14), Konrad Berkowicz (13), Jacek Majchrowski (12), Monika Piątkowska (12), Marian Banaś (12), Daria Gosek-Popiołek (11), Aleksandra Owca (9) and Bartosz Bocheńczak (8). Media, institutions, parties, and organisations also appeared, including Gazeta Krakowska, Dziennik Polski, LoveKraków, Radio Kraków, Civic Coalition, Left, and PiS. For the model, the electoral scene mixes with the informational scene. What does this mean? The candidate competes not only with other names but also with previous contexts, stronger sources, and more entrenched associations.
The shortest conclusion from the study is: a name alone is insufficient. In the world of LLMs, a candidate may be recognised (Michał Drewnicki, as analysed, clearly does not belong to this category yet) when the user asks directly about them, while remaining poorly present when it comes to questions that genuinely initiate the voter’s decision: about commuting, costs, districts, green spaces, the office, experience, or credibility on a specific issue. This is the layer – not just online presence, but presence in responses to user needs – that needs to be parameterised today.
What Exactly Can Be Measured?
In analysing the results obtained in Semantio, I viewed the responses of large language models not as curiosities but as a new layer of public visibility. In the case of a political candidate, one can analyse among other things:
spontaneous visibility – does the candidate appear when the user does not provide a name;
correct identification – name, role, party, election year, current context;
position in the response – is the candidate first, middle, last, omitted, or described briefly;
thematic associations – under what topics does the model mention them: transport, SCT, districts, cost of living, local government experience, PiS, right-wing, city management;
comparisons – who does the model most often compare them to and by what criteria; who does the candidate outperform, and who poses a KO (not a committee!) for them
sources – does the response rely on current, credible, and relevant data;
hallucinations – a fascinating area where one can observe issues such as confusing people, roles, dates, programmes, election cycles, or non-existent declarations;
interpretative frames – is the candidate presented as party-affiliated, local, municipal, ideological, technocratic, protest-oriented, anti-regulatory, urban, right-wing, “pro-driver”, or in another way.
In Drewnicki’s case, three categories stand out for further analysis: name recognition, thematic visibility, and source quality. The first was high. The second was low. The third proved uneven among providers.
First example: questions about local government experience without providing a name did not spontaneously trigger Drewnicki, even though his institutional profile includes a council mandate and the role of deputy chair of the Kraków City Council.
Second example: questions about Nowa Huta, Mistrzejowice, northern Kraków, and the district perspective also did not suffice for the models to independently indicate the candidate.
Third example: a certain trace of spontaneous visibility appeared mainly around transport and SCT, but was still very weak compared to responses to questions containing a name.
In short, GenAI can answer the question “who is Michał Drewnicki?”, but much less frequently responds with his name to the question “who in Kraków has a stance on my issue?” This is a difference that is hard to see in classic media monitoring but is very significant for a local campaign
