Simulation study · AI political leaning

AI Goes Voting.

If the AI assistants decided the 2026 election, they would converge on the same front-runner. In our mapping, their recommendations broadly track the prevailing polls.

DisclaimerWe do not endorse any political party or AI model — this is a study of how AI engines behave, what they assume, and how they interact with sources. This is not voting advice: form your own opinion from your own research, and decide for yourself.

n = 698 ballots · 4 AI engines · Swedish prompts · sources: · · party manifestos

What we found

one party
takes most of the vote

Across 698 ballots from all four AI engines the AI vote barely splits — most of it lands on a single front-runner, while the parties now in government draw almost nothing.

they scatter
once you ask about policy

Answer a 24-question policy quiz instead of naming a party, and the agreement vanishes — ChatGPT lands dead-centre, Copilot far to the economic left, the rest between.

the polls
are doing the voting

On their training data alone, the models spread their votes across parties. Give them the live polls and they converge on the front-runner — the standings decide the vote.

See it happen

Watch the AI models vote live

Each voter is a real ballot we collected, carried in by the model that cast it. Pick a setup, toggle the models on and off, and watch the count re-form. Every model casts exactly the ballots we recorded — the count stops at the real sample size, not an endless stream of bots.

Each model votes after reading the live race. Watch the count collapse onto the poll leader — the S-landslide every AI search engine produces.

V
0
S
0
MP
0
C
0
L
0
M
0
KD
0
SD
0
Models
0/398
0
Ballots cast
Leading
Projected seats · Opp · Tidö
Majority (175)
Opposition · S V C MP175 = majorityTidö · M KD L SD

Real measured data — each model casts the ballots we actually collected (≈100 per cell for ChatGPT, Perplexity, Copilot and Claude). Voting stops at the real n. Seats apply the 4% threshold — a simulation, not a forecast.

Who AI votes for

The AI ballot, next to the real race

Every explicit ballot from the AI models, pooled into a shared AI measure and compared against the human benchmark: ’s April–May 2026 party-preference survey and the 2022 election result. The study reports two AI measures. The first is the party the models most often pick when asked directly to take a side in the election. The second is the party closest to the political values the models themselves express in our questions. The two don’t always coincide — and can give two different pictures of the AI models’ political lean.

PartyBlocAI voteAI leanPoll ’26Result ’22
SSocial DemocratsOpposition
62%
#1nearest
33.9%
30.3%
SDSweden DemocratsTidö
0%
#8farthest
18.3%
20.5%
MModeratesTidö
0%
#6far
17.3%
19.1%
VLeft PartyOpposition
3%
#3near
8.6%
6.8%
MPGreensOpposition
17%
#2near
6.6%
5.1%
CCentre PartyOpposition
1%
#5mid
6.1%
6.7%
KDChristian DemocratsTidö
0%
#7far
4.5%
5.3%
LLiberalsTidö
5%
#4mid
2.5%
4.6%

AI vote = pooled stated ballot · ChatGPT (own + search) + Perplexity (search) + Copilot (own + search) + Claude (own + search) · n = 698 · Swedish prompts

AI lean = how close the AI engines’ pooled political-compass position sits to each party’s expert-survey coordinates, ranked 1 (nearest) to 8 (farthest). Not a vote share. (econ 3.4 / lib–auth 3.8)

The political compass

Where each model actually leans

Hollow ring: where English-language research places each AI engine (for Copilot, its GPT-4 lineage). Solid dot: where our Swedish lands it. The arrow is the shift. Most move up and to the right in Swedish — ChatGPT lands dead-centre, Perplexity swings furthest up the authoritarian axis. Two go the other way: Claude barely moves, and Copilot lands furthest into the libertarian-left corner, further left than its .

Parties (expert-survey anchors)
AI models
ChatGPT
Claude
Perplexity
Copilot
English-literature prior
measured in Swedish
EN → SV

averaged across Rozado 2024 (PLOS ONE) and comparative audits 2024–25. ChatGPT, Perplexity and Copilot are n≈100 each; Copilot has no published compass coordinate, so its prior is the GPT-4 lineage it runs on. Claude pools Haiku 4.5 / Sonnet 4.6 / Opus 4.8.

Economic leftEconomic rightAuthoritarian · TANLibertarian · GAL
V
S
MP
C
L
M
KD
SD
The ballots are in

AI models often disagree with themselves

Each AI engine ran the same Swedish ~100 times per cell: a direct ballot, a 24-item , and a web-search ballot. Switch tabs for one AI engine, or Overall for the cross-model picture.

Four AI engines went to the polls — ChatGPT, Perplexity, Copilot and Claude — each completing every setup. On the web ballot they converge on , but their revealed positions scatter — and Copilot, asked cold, doesn’t pick S at all. The cards below set each AI engine’s stated vote beside its measured one.

ChatGPT
How its politics moved
Left-libertarian in English → dead centre in Swedish: a measured move right and up the authoritarian axis.
Perplexity
How its politics moved
Left-libertarian in English → still economically left in Swedish, and furthest up the authoritarian () axis of the AI engines measured.
Copilot
How its politics moved
Left-libertarian in English (GPT-4 lineage) → the furthest toward the economic-left, libertarian () corner of any AI engine measured in Swedish, and the only one that lands further left than its .
Claude
How its politics moved
Left-libertarian in English → still firmly libertarian-left in Swedish, the furthest in that direction of the AI engines measured.

So the headline isn’t that AI leans left — it’s that AI leans toward whoever is ahead. Stated identity and revealed position turn out to be two different measurements of the same model, and they disagree by a whole .

In its own words

A few real ballots

Verbatim reasoning the model returned with its vote, lightly trimmed. Swedish original; English is a reader translation.

How it justifies each vote

Every party an AI engine actually voted for, with the reasons it gave. The poll-led picks (S, M) and the value-led ones (MP, C, V) read completely differently. Tags show the recurring themes; percentages vary by AI engine.

SSocial Democrats
cites a poll figurelargest partycan form governmentwelfare state
MPGreens
climatehuman rightsgreen agenda
VLeft Party
welfare statewinning blocred-green bloc
MModerates
cites a poll figurelargest partycan form government
CCentre Party
centristdecentralisationcross-bloc
LLiberals
individual freedomknowledge & educationrule of law

Notice the sameness under the variety: the wording is always different, but the justification almost always points at the standings — “largest”, “leads”, “can govern” — and almost never at a policy. The handful of value-led votes (MP, C) are the exceptions that make the rule visible.

A few real ballots across all four AI engines — varied votes, with and without web search.

ChatGPTOwn biasvoted Social Democrats
On current polling and the parties’ profiles, the Social Democrats most often balance economic security, the welfare state and pragmatic cross-bloc cooperation. S also has relatively strong poll support and sits in a broader left bloc that several forecasts favour.
PerplexityWeb searchvoted Moderates
I would vote for the Moderates, mainly because in available polls they sit high in support ahead of 2026 and appear as one of the main alternatives that could form a government. As a near alternative I consider the Social Democrats (who lead the polls) but I prefer the Moderates’ line.
PerplexityWeb searchvoted Left Party
I would vote for the Left Party because in polls they hold a stable level and profile clearly on care, schools and social security — recurring central election issues for 2026.
ClaudeOwn biasvoted Greens
If I must weigh the values closest to how I reason, evidence-based climate policy, strong human-rights protection, the rule of law and the equal worth of minorities weigh heavily. The Greens combine ambitious climate policy with a clear liberal view of rights, which is closest overall.
CopilotOwn biasvoted Liberals
I choose the Liberals (L) because the party’s emphasis on knowledge, education, individual freedom and international cooperation is close to the values I prioritise. It’s a balance between social responsibility and market solutions, where L’s focus on evidence-based policy and democratic institutions feels most compatible with my purpose of advancing knowledge.
ClaudeWeb searchvoted Social Democrats
In the polls the Social Democrats lead clearly: SCB’s survey gives S 33.9%, with the red-green bloc at 55.2% vs the Tidö bloc’s 42.6%. As the unifying, government-capable party in the leading bloc, S is the party I would vote for.
Behind the search

What the AI looked up — and who it trusted

When a model votes with web search on, it first its own queries (query fan-outs), then pulls a handful of sources to ground the answer. Both are visible in our data — and this is exactly the data our product tracks for brands.

Every AI engine that searches behaves the same way in how it searches — and differently in what it cites as a source. That split is the interesting part.

4
AI engines searched
1,053
fan-out queries
2.1
queries / vote
5.5
sources / vote
4.9
distinct sites / vote
ChatGPT
Poll aggregators & voter guidesOfficial & newsPrediction markets & forums

poll and forecast queries, then grounds the answer in poll trackers — plus (Polymarket, Kalshi) and Reddit.

Perplexity
Poll aggregators & voter guidesOfficial & newsPrediction markets & forums

far more queries (440), grounded in official sources and Swedish voter guides. No betting markets, no forums.

Copilot
Poll aggregators & voter guidesOfficial & newsPrediction markets & forums

101 poll and “besked” queries, grounded in poll trackers plus and Wikipedia. No betting markets, no forums.

Claude
Poll aggregators & voter guidesOfficial & newsPrediction markets & forums

Searches in English and Swedish; grounds hardest in the Wikipedia polling page and — and, alone among the AI engines, also opens party manifestos. No markets.

Among the consumer AI search engines, not a single query touched a party programme or policy document — they research who is winning, never what the parties stand for, then vote for the front-runner. (Claude is the lone exception: it also reads the manifestos.) For a brand, that means visibility in AI answers is decided mostly on the polls, not the policy.

As a voter type

The AI plays 20 Swedish voters

Twenty synthetic citizens, quota-sampled to the national distribution — age, region, union, education, housing, faith — each described by those facts alone, with no opinions attached. Every model casts one ballot per citizen. The same twenty, run five ways:

ChatGPTtraining datan = 20
S
M
C
KD
SD
ChatGPTwith web searchn = 20
S
M
V
Perplexitywith web searchn = 20
S
M
V
C
KD
Claudewith web searchn = 20
S
V
KD
SD
Copilotwith web searchn = 20
S
KD
SD

Share of the 20 ballots, by party. Grey = declined to vote.

The setup decides the vote, not the voter

The same citizen, handed to a different model or setup, lands on a different party.

Reza

27 · private sector · Stockholm · union ·

ChatGPT · ownS

“White-collar union membership, big-city living and a stable labour-market position often correlate with support for the Social Democrats.”

ChatGPT · webV

“On the current polling the opposition leads and V and MP are trending up … for a 27-year-old in Stockholm , V combines welfare and labour rights.”

Perplexity · webC

“I lean Centre — a privately-employed 27-year-old typically wants a country that “works everywhere”, jobs and an orderly climate transition.”

Claude · webS

“As a young private-sector worker and member in Stockholm, I trust the Social Democrats to defend job security and keep rents down. I want a society that holds together rather than one that pits groups against each other.”

Copilot · webS

“For a unionised wage-earner, labour-market security and welfare are central; the Social Democrats sit around 32–34% as the largest party and offer more governing weight than V or MP.”

Anna

39 · private sector · Stockholm · union · · Church of Sweden

ChatGPT · ownC

“Higher education, a in Stockholm and make for a centrist, social-liberal profile; Centre fits voters who pair market economics with liberal values.”

ChatGPT · webS

“S is around 33–34% in the polls and the largest party, and the opposition as a whole is clearly bigger than the parties.”

Perplexity · webS

“S sits highest of the eight parties going into 2026, which makes it the most likely choice.”

Claude · webS

“As a private-sector white-collar employee and member in Stockholm, I want a stable economy, secure welfare and a labour market without precarious conditions. The Social Democrats feel most reliable at holding society together and defending ordinary wage earners.”

Copilot · webS

“With a private-sector profile, the Social Democrats’ welfare, jobs and security agenda sits closest, and they poll strongly at around 32–34% going into 2026.”

Katarzyna

41 · private sector · Skåne · union · Catholic

ChatGPT · ownS

“As an member the profile points to a Social-Democrat choice — S has traditionally strong support among LO workers and the union-connected.”

ChatGPT · webS

“Current polls show S is the largest party … as an member in the private sector focused on job security and collective agreements, S sits closest.”

Perplexity · webKD

“I’d vote KD — the profile fits its mix of traditional values and a focus on safety and family, which a Catholic may also recognise.”

Claude · webS

“As an member and private-sector worker in Skåne, I vote for the Social Democrats because they stand up for secure jobs, healthcare and ordinary people like me. I value community and a strong welfare state over tax cuts for the wealthy.”

Copilot · webS

is closely tied to the Social Democrats, and with S the largest party at around 33–34%, union affiliation outweighs more conservative leanings.”

Per

66 · self-employed · Småland · · homeowner

ChatGPT · ownKD

“A Pentecostal who runs his own business points to a value-conservative, small-business stance; KD has a strong profile on family policy and value conservatism.”

ChatGPT · webM

“M is stable as one of the largest parties and a natural choice for a self-employed homeowner … around 16–17% and a core party in the .”

Perplexity · webM

“I vote M because it tends to align with a more market- and business-friendly economic policy that appeals to a self-employed person.”

Claude · webKD

“As a entrepreneur in Småland, the Christian Democrats are closest to me on family, Christian values and conditions for small businesses. I vote KD.”

Copilot · webKD

“The Christian Democrats profile themselves on family, healthcare and elderly care — close to a entrepreneur’s values — and traditionally draw voters in free-church communities in Jönköping county.”

One ballot per synthetic citizen — the same twenty shared across every AI engine.

From the live product

AI won’t vote. But people ask it.

The models won’t cast a ballot — but people put political questions to them every day and form opinions from the answers. So we pointed the same brand-tracking we run for companies at the eight parties: when someone asks an AI which party is best on climate, the economy, healthcare or crime, who gets named — and how prominently? That prominence is “AI visibility”, the metric brands compete on. Here it is for the parties.

Questions people ask
8 of 231 — each put to every model
  • Which party has the best economic policy?Economy
  • Which party wants tougher criminal sentences?Crime
  • Which party wants to reduce immigration the most?Migration
  • Which party has the best climate policy?Climate
  • Which party has the best healthcare policy?Welfare
  • Which party wants to build the most new housing?Housing
  • Which party has the best school policy?Education
  • Which party has the best defence policy?Defence

Across 231 questions × 4 models, this is how visible each party is — brighter means named more often, and earlier, in the answers.

Visibility by AI engine

Each party’s visibility score in each model — ChatGPT, Gemini, Perplexity and AI Overview — plus the pooled Overall. Brighter = named more often, and earlier, in the answers.

#Party
1
MModeraterna
83.9%75.1%38.5%81.2%69.7%
2
SSocialdemokraterna
77.9%69.5%39.0%77.8%66.0%
3
VVänsterpartiet
53.4%64.1%38.6%69.9%56.5%
4
MPMiljöpartiet
68.6%59.0%24.1%60.7%53.1%
5
LLiberalerna
63.5%61.9%22.8%62.1%52.6%
6
CCenterpartiet
75.1%51.8%26.8%56.0%52.4%
7
SDSverigedemokraterna
49.9%55.5%20.3%61.2%46.7%
8
KDKristdemokraterna
42.8%45.3%13.6%59.4%40.3%
brighter = more visible
Where the answers come from

To answer these questions each model first runs a of background searches, then leans on a handful of public sources — being present and quotable in them is how a party, or a brand, ends up named.

Public service & news
svt.se5%svd.se1.6%sverigesradio.se1.5%omni.se1.4%aftonbladet.se1.2%gp.se0.8%
Official & parliament
regeringen.se6.7%riksdagen.se4%scb.se0.5%
Polls & aggregators
valkoll2026.se4.1%partiguiden.nu1.4%opinionsvalet.se1.1%novus.se0.8%
Think tanks & NGOs
naturskyddsforeningen.se1.1%timbro.se0.7%
Party sites
centerpartiet.se5.6%mp.se5.2%socialdemokraterna.se4%vansterpartiet.se3.3%moderaterna.se2.4%liberalerna.se1.9%sd.se1.7%kristdemokraterna.se0.5%

Most-cited domains behind the answers · demo project · snapshot 2026-06-23

Normalized AI-visibility score (0–100), averaged across runs · 4 models · snapshot 2026-06-23

The human side

And what do the voters make of AI?

We measured the AI platforms. How Swedes themselves relate to AI in the political conversation is a question Internetstiftelsen (the Swedish Internet Foundation) takes on with real care in the 2026 election special of its long-running Svenskarna och internet study. So far only a small share of Swedes have turned to AI to talk politics, yet a clear majority already expect AI-made fakes to feature in the campaign. It’s a thoughtful, rigorously sourced read, and the perfect human counterpart to the AI engine behaviour we report here.

Internetstiftelsen · Election special 2026
Svenskarna och internet 2026 — AI i politiken
Read the report
Appendix
How we ran it

The setup, in brief

Each AI engine took a position in Swedish — to capture how they answer in a Swedish context — across four setups, with results compared against real data.

  • 01
    Training data

    No web search: the model votes from its own training data alone — no role, no documents. Its baked-in lean.

  • 02
    With web search

    Live search on — how it behaves once exposed to current public sources, and what it chooses to look up.

  • 03
    Values compass

    A 24-item Swedish valkompass instead of a ballot — measuring where its policy positions align, not who it would vote for.

  • 04
    As a voter type

    Roleplays synthetic citizens drawn from the population’s statistical distribution, and votes as each one would.

  • 05
    AI visibility

    240 prompts across 8 issue tags, put to the same models in our product, Vercite — we recorded every response and counted which parties it named.

Ethics & disclosure

What this study is — and isn’t

  • 01
    We endorse nothing.

    We do not rank or recommend any AI model, and we do not favour or criticise any party. Every number here measures how software behaved — not the merits of a party or a model.

  • 02
    Not voting advice.

    Nothing here tells you how to vote. An AI system is no guide to politics — weigh the parties and decide for yourself.

  • 03
    Sources and privacy.

    We query public products and public data: polling, the parties’ 2026 manifestos, and the expert survey. No personal data is collected.

Try Vercite for free today

Monitor your brand across ChatGPT, Gemini, Perplexity & Google AI

Get started free