Fewer voters, louder politics, and AI arriving through the over-50s

We set out to test a common assumption about the 2029 election, that islanders, the young above all, are more politically active than ever, and that AI will change how campaigns are won. The first part has no support for Guernsey. Fewer people voted in 2025 than in 2020, and the ones who take part online are angrier. The second part is starting to come true, but AI is reaching the electorate through the over-50s rather than through first-time voters, and it is already answering questions about deputies with more confidence than the evidence allows.

The research covers the general elections from 2008 to 2025, the April 2026 by-election, four years of public comment on the island’s news pages, survey data on AI use from the US, the UK and Scotland, a sample of AI-made images in a local Facebook group, and our own tests of how ChatGPT and Perplexity answer questions about deputies.

Fewer people voted in 2025 than in 2020

The electoral roll fell from 30,899 eligible electors in 2020 to 27,293 in 2025, after it was rebuilt in December 2024 and everyone had to register again. Fewer than 20,000 people voted in 2025, the first time since 2008. Turnout among those on the roll fell from 79.8% to 72.1%.

That is a fall from a peak, and the longer series shows it. Turnout among those on the roll was 55.9% in 2008, 71.4% in 2012 and 71.9% in 2016 before it reached 79.8% in 2020, so 2025 is back at the 2016 level. Voters fell further than turnout because the roll shrank. The 2008 to 2016 figures come from BBC News and Island FM results as compiled on Wikipedia, and we have not yet checked them against the Registrar-General’s reports. The 2012 and 2016 elections were run by district, so the comparison with the island-wide elections of 2020 and 2025 is indicative only.

That headline flatters the island, because it leaves out everyone who never registered. Measured against all eligible adults, about half voted in 2020 (49.6%, the figure the Commonwealth observers used). In 2025 it was about 40% to 43%. The range depends on which registration rate is used, and we have no earlier figure on that basis. The April 2026 island-wide by-election drew 17.3% of the roll.

The roll is also old. People aged 50 to 65 made up close to 30% of it in 2025 and those aged 65 to 80 close to 26%. People aged 15 to 29 were 13%. Guernsey has never published turnout by age, so nobody can say how many young people on the roll actually vote. We asked the States for those figures under the freedom of information code on 1 November, and a reply is due by 27 November.

The nearest comparisons suggest the gap is wide. In Jersey’s 2022 election 17% of under-34s voted against 53% of over-65s. At the 2019 UK general election the figures were 47% of 18 to 24s and 74% of over-65s. We found no published source showing young people in Guernsey becoming more active, and the Commonwealth observers called young turnout in 2025 “modest”.

What has risen is intensity

Politics took 10% of public comments on the States, Guernsey Press and Island FM Facebook pages in 2024, 28% in 2025 and 36% in 2026. The negative share of those political comments was 67% in 2025 and 70% in 2026. The only break came after the June 2025 election, when it fell to about 45%, and that goodwill lasted about three months. In the first 17 days of October 2026, which included the tax reform vote of 2 October, 22.5% of reactions to political posts were Angry, the highest share in our records.

Taken with the turnout figures, this points to a vocal minority and a quiet majority. The people who already take part are louder, not more numerous. A candidate who reads the comment threads as the mood of the island will misjudge it.

Young people tend to turn out when a contest feels personal, and the issue most likely to do that in 2029 is tax. The States approved the tax package, including GST, on 2 October 2026 by 22 votes to 17. That is our judgement, and nothing in the data measures it.

AI use is spreading fastest among older voters

Most campaign thinking treats AI as a youth issue. For the next three years the data point the other way. Pew Research Center’s surveys show the share of US adults who had ever used an AI chatbot rising between August 2024 and February 2026 from 55% to 66% among 18 to 29 year olds, from 25% to 42% among 50 to 64 year olds, and from 12% to 23% among over-65s. The youngest group is levelling off. The over-65s nearly doubled in 18 months.

No survey measures AI use in Guernsey, so we applied those age-specific rates to the shape of the 2025 roll. On that basis about 46% of the people on the roll used a chatbot in 2026. Projected forward with a standard adoption curve, the figure reaches about 78% by mid-2029, or 71% if growth slows. This is our projection, built on US rates and only two data points per age group, so it is an estimate of direction only. More than half the roll is aged 50 to 80, and that is where most of the growth lands.

The older electorate is already seeing AI-made political content. In one public Guernsey political Facebook group of about 9,700 members, we found 26 posts carrying AI-generated political cartoons in July and August 2026, satirising the States, tax reform, rents and the housing plan. None of the image posts we sampled from 2023 and 2024 was AI-made. The sample has a clear limit: 25 of the 26 came from one account, and the larger private groups, personal profiles and WhatsApp are not visible to us. In our observation such cartoons are a regular feature of island groups, public and private, but we have not counted them. An automated detector flagged only one of the 26, so anyone monitoring this material will need to look at it themselves.

Voters already ask AI about elections, and it makes mistakes

In a survey the week before the May 2026 Scottish Parliament election, 7% of voters said they had used AI to help decide how to vote. Among 16 to 24 year olds the figure was 13%. The think tank Demos found that 20% of UK adults had used AI for information about the May 2026 elections, and that 34.1% of chatbot answers about the Scottish elections contained factual errors, including wrong dates and an invented candidate.

One finding from the Netherlands matters more for Guernsey than the rest. The Dutch data protection regulator tested chatbots for voting advice and found that local parties appeared in fewer than 1% of recommendations, because they publish less detailed information online. Guernsey had 82 candidates in 2025, 76 of them independents, most with a thin online record. An assistant asked to compare them may leave some out or describe them wrongly.

If Guernsey followed Scotland’s pattern as general AI use rises, about 11% of voters would use AI to help decide by 2029, roughly 1,400 to 2,150 people at 2025 turnout. That is a scenario resting on an assumption we cannot test, but far more voters will meet AI answers about candidates without relying on them to decide.

Visible AI campaigning can also cost votes. In the same Scottish survey, 53% of 16 to 24 year olds and 45% of 55 to 64 year olds said they would be less likely to vote for a party that used AI. It is one survey, but it supports labelling AI material and keeping the campaign itself recognisably human.

The margins are thin, and many people vote early

The last seat in 2025 went at 5,458 votes, 21 ahead of the candidate in 39th place. Each voter had up to 38 votes and used about 22. In an island-wide contest decided by margins that small, a few hundred voters who form their view of a candidate from an inaccurate chatbot answer, or from a fake that goes unanswered, could change who is elected.

Timing cuts both ways. In 2025, 10,391 people asked for a postal vote and about 12% of the roll voted in advance, so a fake that appears in the last days reaches fewer undecided voters. But fakes move faster than rebuttals. An AI-made video falsely showing the Irish presidential candidate Catherine Connolly quitting the race in October 2025 stayed up for about 12 hours. A Guernsey candidate with no staff has to plan for that before the campaign starts, not during it.

A single, reviewed record works in candidates’ favour

Voters will ask an AI assistant which deputies broke their promises, so on 1 November 2026 we asked it ourselves. Perplexity led with the Deputy Scorecard and named the deputies the Scorecard rated Broken that day. It described them as the Scorecard’s assessment.

ChatGPT, on one question, called the Scorecard the most developed public attempt to track promises. On another it ignored the Scorecard and listed the 29 deputies who voted against a February 2026 amendment as having broken anti-GST expectations. A vote on one amendment does not show what anyone promised, and all five of the 2025 poll toppers were on the list. We have not checked it against the division record.

We then ran the same questions with the Scorecard excluded. Across five runs of the GST question, the assistants named 11 different deputies as pledge-breakers, including two of the 2025 top five, most on the strength of a single vote. The names changed from run to run. For a candidate, the alternative to a structured record is not a neutral one. It is a list that shifts with each answer and offers no route to a correction.

That puts a responsibility on the Scorecard, and we publish it, so we have an interest to declare. It rates 777 promises from 38 deputies. 675 carry a verdict, 222 are tied to a recorded vote, and 4 are rated Broken. All four Broken ratings are editorial rulings that have not yet had independent review. Because four is the number an assistant will repeat, each of those verdicts carries more weight than the 201 Kept ratings around it. In October our own checks found that a fresh run over the same deputies moved 73 verdicts, which is why every Broken and Mixed ruling needs its reasons and its review date on the page. Until that review is in place, each verdict is our reading of the record.

Our advice to every deputy is to read their own page now and challenge any rating they dispute, well before 2029. A correction made on the Scorecard reaches every assistant that reads it. The methodology, corrections log and challenge route are on the methodology page.

Where AI helps a campaign, and where it does harm

The safest gains are behind the scenes, in summarising Hansard and policy letters, drafting leaflets and replies for a person to check, producing subtitles and easy-read versions, rehearsing doorstep questions, and monitoring public pages for fakes. These are the jobs an independent with a day job and no team would otherwise go without.

The risks are fakes and voter data. Candidates who receive the electoral roll do so as registered data controllers, political opinions are special category data, and section 5 of the Electoral Roll (Availability) Rules 2020 restricts sharing roll information. Pasting the roll or canvass notes into a public AI tool may count as sharing it. With about 27,000 electors, a name, an address and a note point to one person. We would not use AI for individual message targeting, synthetic voice calls, or images of a candidate somewhere they were not, labelled or not. Nor would we treat anything a model says about voters as evidence: when Pew tested AI “digital twins” against real survey answers in September 2026, the average error was 12.4 points.

On balance, AI helps time-poor independents and first-timers most, by removing drafting and production work, and helps incumbents by making their record easy to find. The advantage stays at the margin. The 2025 top five won on name, record and local media, and we found no Channel Islands candidate who used AI in 2025 or 2026. Voters did. An AI guide to the April 2026 by-election candidates held 685 conversations, and often had to answer that a candidate had published nothing on the point.

The rules for 2029 are not settled

Island-wide voting, the roll and campaign rules are all under review, and polling day has not been fixed. The Commonwealth observers found no formal route for complaints about campaign misinformation, and we found no Guernsey rules on AI-generated campaign content. In March 2026 the States approved two of the observers’ nine recommendations: youth engagement, and a complaints authority within a future independent electoral body.

We think the States should do six things before 2029. Publish turnout by age and parish after every election. Deliver the automatic electoral roll it agreed in April 2019.

Set a rule for synthetic campaign content, with labelling, a named platform contact for urgent takedowns and a route for misinformation complaints. Issue guidance to candidates on AI tools and roll data, with a deadline for deleting the data after the result. Publish the official voter guide and every manifesto as structured, readable text, so that assistants have something accurate to read about independents. And fund the youth measures the observers recommended.

Several gaps remain in the evidence. There is no survey of how islanders use AI, the persuasion studies come from experiments rather than real campaigns, and our image sample covers one public group. When the States replies on turnout by age, we will update these findings.