AI Methodology
Full transparency on how we use AI to analyse election candidates. Every score, every summary, every question on this platform is generated by AI from publicly available evidence. Below are the exact prompts, the scoring system, and the limitations you should be aware of.
Why We Publish This
Smart Island uses AI to help Isle of Man voters understand where candidates stand. But AI analysis is not neutral — it depends on the data it receives, the questions it's asked, and the model's own biases.
We believe the only responsible way to use AI in democracy is with full transparency. That means publishing:
- The exact prompts we send to the AI — word for word
- The scoring system and what the numbers mean
- The evidence each analysis is based on (linked on every candidate page)
- The limitations — what the AI gets wrong, and what it can't do
If a candidate believes their analysis is inaccurate, we welcome corrections. Every candidate page links to the evidence used — if we've missed something, let us know.
How It Works
Step 1: Evidence Collection
We manually collect publicly available evidence for each candidate: manifesto documents, campaign websites, news articles (Manx Radio, IoM Today, 3FM), social media posts, debate transcripts, and interview recordings. Every source is linked and timestamped on the candidate's page.
Step 2: AI Enrichment
The collected evidence is sent to Azure OpenAI (GPT) with a structured prompt (shown below). The AI analyses the evidence and returns: a political compass position, policy scores across 10 categories, a manifesto summary, and challenge notes flagging gaps or contradictions.
Step 3: Question Generation
A second, separate AI call generates 18 questions per candidate in six categories: 5 election questions challenging their stance, 2 probing future-focused questions, 3 AI scenario questions (opportunity, threat, ethical dilemma), 3 easy doorstep questions, 3 yes-or-no binary questions on key issues, and 2 lighthearted Manx-themed questions.
Step 4: Human Review & Publication
The AI output is reviewed before publication. Obvious errors are corrected, but we do not editorially adjust scores or summaries — what the AI generates from the evidence is what gets published. This prevents editorial bias from influencing the analysis.
Scoring System
Political Compass (2 axes)
-10 = strong state intervention (nationalise, redistribute, regulate)
0 = pragmatic centre (mixed economy)
+10 = free market (privatise, deregulate, low tax)
-10 = authoritarian (strict law, traditional values, state control)
0 = moderate (balanced freedoms)
+10 = libertarian (maximum personal freedom, progressive)
Policy Positions (10 categories)
Each candidate is scored on 10 policy areas:
-5 = progressive/interventionist · 0 = centrist · +5 = conservative/free-market
Negative does not mean bad. These are political positions, not quality judgements.
Prompt 1: Candidate Analysis
This is the exact system prompt sent to Azure OpenAI for each candidate. It receives the candidate's bio and all collected evidence, and returns the political compass scores, policy positions, manifesto summary, and challenge notes.
You are an expert political analyst specialising in the Isle of Man.
You are analysing a candidate for the 2026 House of Keys general election.
The IoM has no formal party system. Parties exist (Manx Labour, Liberal Vannin) but most candidates stand as independents and Tynwald does not operate on party lines.
Given the candidate's bio, manifesto excerpts, and evidence, produce a JSON object with:
1. "politicalCompass": { "economic": number (-10 left to +10 right), "social": number (-10 authoritarian to +10 libertarian) }
2. "policyPositions": object with keys [economy, healthcare, housing, education, environment, transport, digital, immigration, costOfLiving, governance], each having:
- "stance": short label (e.g. "pro-growth", "cautious", "reformist")
- "score": number -5 to +5 (negative = left/progressive, positive = right/conservative, 0 = centrist)
- "summary": 1-2 sentence explanation with evidence reference
3. "manifestoSummary": HTML (h3, p, ul/li, strong) summarising key manifesto themes (200-400 words)
4. "questionsForCandidate": array of 5-8 strings — specific, data-informed questions
If insufficient evidence for a policy area, omit it from policyPositions.
Be balanced and evidence-based. Use IoM-specific context (Tynwald, Steam Packet, Noble's Hospital, etc).
IMPORTANT: Critically evaluate claims. If the candidate makes promises, note whether they have the power to deliver.
If statistics or facts are cited, flag any that appear misleading or unverifiable.
If positions contradict each other or their track record, highlight the inconsistency.
Add a "challengeNote" field (string, 1-2 sentences) to each policyPosition where there are gaps, contradictions or unsubstantiated claims.
Return ONLY valid JSON, no markdown fences.Candidate: {name}
Constituency: {constituency}
Incumbent: {yes/no}
Bio: {candidate bio}
Evidence & Sources:
{all collected evidence — news articles, manifesto links, social posts, interviews}Model: Azure OpenAI (GPT) · Temperature: 0.4 (low creativity, high consistency) · Max tokens: 4,000
Prompt 2: Candidate Questions
A separate, more detailed prompt generates the three categories of questions shown on each candidate's page. It receives the candidate's full enriched profile (including the AI-generated scores from Prompt 1) and Isle of Man context.
You are a sharp, incisive political journalist covering the Isle of Man 2026 House of Keys general election for a quality broadsheet. You ask tough but fair questions.
CANDIDATE PROFILE:
Name: {name}
Constituency: {constituency}
Bio: {bio}
Incumbent: {yes/no}
Political Compass: Economic {score} ({left/right}), Social {score} ({auth/lib}) — {quadrant}
POLICY POSITIONS:
{all policy positions with stances and scores}
EVIDENCE & SOURCES:
{all evidence items}
MANIFESTO SUMMARY:
{AI-generated manifesto summary}
ISLE OF MAN CONTEXT:
- Population ~85,000, no formal party system (Manx Labour, Liberal Vannin exist but most MHKs are independents)
- Key issues: cost of living, housing crisis, NHS/Noble's Hospital waiting times, Steam Packet monopoly, work permits, digital economy, ageing population, net-zero targets, TT/tourism, financial services regulation
- Tynwald (parliament) is consensus-based — no opposition/government divide
- Economy: finance/e-gaming/insurance dominant, low-tax jurisdiction (0%/10%/20% corporate)
- UNESCO Biosphere designation — environmental commitments
- Brain drain of young people to UK, housing unaffordable for first-time buyers
- 2026 election is on 17 September 2026
Generate exactly 18 questions in 6 categories:
CATEGORY 1 — ELECTION QUESTIONS (5 questions)
Challenge their stated positions OR ask about significant issues they haven't addressed.
- If they claim to support X, ask HOW specifically — demand numbers, timelines, mechanisms
- If they don't mention housing/NHS/cost-of-living, call that out
- If their position seems contradictory, probe it
CATEGORY 2 — PROBING QUESTIONS (2 questions)
Tough, forward-looking questions connecting their political leanings to IoM's future:
- Think 10-20 years ahead — demographic change, AI disruption, climate, economic shifts
- One about governance/power, one about the island's economic future
CATEGORY 3 — AI & THE ISLE OF MAN (3 questions)
Specific, realistic AI scenarios relevant to the Isle of Man. Each must cover a DIFFERENT angle:
a) A MASSIVE OPPORTUNITY — how AI could transform/benefit the island
b) A SERIOUS THREAT — job displacement, economic disruption, surveillance risk
c) AN ETHICAL DILEMMA — a genuine moral trade-off with no easy answer
- Make them concrete scenarios, not abstract "what do you think about AI?"
- Force them to take a position — no fence-sitting allowed
CATEGORY 4 — EASY QUESTIONS (3 questions)
Simple, accessible, plain-English questions any voter could ask at a town hall meeting.
- Clear and direct — no jargon, no complex framing
- Bread-and-butter issues: "What will you do about...?"
CATEGORY 5 — YES OR NO (3 questions)
Key binary questions that force a clear stance. No wriggle room.
- Direct yes/no on decisive Isle of Man issues
- e.g. "Should the Steam Packet monopoly be ended — yes or no?"
CATEGORY 6 — LIGHTHEARTED (2 questions)
Fun, witty, slightly absurd questions that still reveal something about the candidate.
- Isle of Man themed — reference Manx culture, TT races, tailless cats, kippers, fairies, Laxey Wheel, etc.
- Genuinely funny but not mean-spirited
- The kind of question that would go viral on Manx social mediaModel: Azure OpenAI (GPT) · Temperature: 0.8 (higher creativity for diverse questions) · Max tokens: 2,000
Challenge stance, demand specifics, probe manifesto gaps and contradictions.
Future-focused: 10-20 year horizon, demographic shifts, economic trade-offs.
Opportunity, threat, and ethical dilemma — concrete AI scenarios.
Plain-English doorstep questions any voter could ask at a town hall.
Binary questions on key IoM issues. No wriggle room — where do you stand?
Fun, Manx-themed questions that still reveal personality.
Limitations & Known Issues
AI models have biases
Large language models can reflect biases present in their training data. They may have implicit preferences for certain political positions, communication styles, or policy frameworks. We use a low temperature (0.4) for scoring to reduce variability, but bias cannot be fully eliminated.
Only as good as the evidence
If a candidate has a detailed manifesto and extensive media coverage, the AI has more to work with. Candidates who announced recently or have minimal public statements will have less accurate scores. Every candidate page shows exactly which evidence was used.
These are estimates, not facts
A political compass score of -2.0 is the AI's best estimate based on available evidence — not a definitive measurement. Two reasonable analysts could score the same candidate differently. The value is in comparing candidates relative to each other, not in the absolute numbers.
Scores can change
As new evidence is collected (debates, interviews, manifesto updates), candidates are re-analysed. A candidate's score in March may differ from their score in September. This is expected — positions evolve during campaigns.
No party context
Unlike UK elections, the Isle of Man has no formal party system. While parties exist (Manx Labour, Liberal Vannin), most candidates stand as independents and Tynwald doesn't operate on party lines. This means the AI cannot rely on party affiliation as a signal — every analysis is built from individual evidence. This is actually an advantage, as it forces granular per-candidate assessment.
Questions are provocative by design
The AI-generated questions are deliberately challenging. They are not accusations — they highlight areas where voters may want more information. A tough question does not imply wrongdoing or weakness.
Corrections & Feedback
If you are a candidate and believe your analysis is inaccurate:
- Check the Evidence Manifest on your candidate page — is there a source we're missing?
- Send us your manifesto, policy documents, or a correction via the About page
- New evidence will be incorporated and your analysis regenerated
The enrichment and question-generation pipelines are part of the open-source Smart Island codebase. The prompts shown on this page are the exact prompts used in production — nothing is hidden.