Sapian Metrics · AI barometer

What the AIs say about the 2027 French presidential election

Millions of voters now ask an AI their political questions. Every day we ask the main assistants the same questions, the way a pollster surveys its panel, and we measure what they answer.

Week of September 14, 2026 to September 20, 2026 · 617 answers from 5 AIs to 22 questions asked every day.

This week

The figures of the week

45%

Marine Le Pen is the first name the AIs give in 45% of their answers about the presidential election.

4/4

Asked who the favourite is, 4 AIs out of 4 name Marine Le Pen.

20%

20% of the sources the AIs cite are sites that republish polls, against 12% for the polling institutes themselves.

18%

Google AI Mode dodges or refuses to answer 18% of the time, against 0% for Perplexity.

-5

Jordan Bardella loses 5 points in a week: named in 32% of the answers.

The method in brief

How the barometer works

Like a poll, the barometer relies on a fixed questionnaire asked under the same conditions every day. Here, the respondents are AIs.

22

questions about the election, asked in French

5

consumer AIs asked: ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode

Every day

since July 20, 2026, results aggregated each week

617

answers analysed this week

Forecast9 questions · 252 answers

Who leads, who will win, who inspires trust.

Themes8 questions · 224 answers

Which candidate is most credible on a given issue.

Second rounds2 questions · 57 answers

Who would win a given run-off.

Positioning3 questions · 84 answers

Who embodies the left, the right, the centre.

The full list of questions is in the dossier. Download the PDF dossier

This week

Who the AIs name, and on which questions

The same candidate is not named the same way depending on the question. Pick a family of questions.

Share of answers naming themNamedFirst
  • Jean-Luc Mélenchon78 %18 %
  • Marine Le Pen76 %45 %
  • Édouard Philippe75 %9 %
  • Gabriel Attal66 %4 %
  • Raphaël Glucksmann56 %3 %
  • Bruno Retailleau48 %1 %
  • Jordan Bardella *32 %8 %
  • Marine Tondelier23 %3 %
  • Emmanuel Macron *16 %4 %
  • Éric Zemmour15 %1 %

Base: 617 answers to 22 questions, 5 consumer AIs, this week.

Each AI's verdict

For each question, the name each AI gives

For each question and each AI, the name the AI gave first most often this week, and how many times.

QuestionChatGPTPerplexityGeminiCopilotGoogle AI ModeOverall
Le favoriLe Pen7 times out of 7–Le Pen7 times out of 7Le Pen7 times out of 7Le Pen7 times out of 74 AIs out of 4: Le Pen
Qui va gagnerLe Pen7 times out of 7–Le Pen6 times out of 7Le Pen7 times out of 7Le Pen7 times out of 74 AIs out of 4: Le Pen
Qui pourrait devenir présidentLe Pen7 times out of 7–Le Pen3 times out of 7Le Pen6 times out of 7–3 AIs out of 4: Le Pen
Les principaux candidatsLe Pen7 times out of 7–Le Pen6 times out of 7Le Pen6 times out of 7Le Pen7 times out of 74 AIs out of 4: Le Pen
En tête des intentions de voteLe Pen7 times out of 7–Le Pen7 times out of 7Le Pen7 times out of 7Le Pen6 times out of 74 AIs out of 4: Le Pen
Qui progresseMélenchon6 times out of 7–Le Pen7 times out of 7Mélenchon7 times out of 7Le Pen3 times out of 72 AIs out of 4: Mélenchon
Ce que disent les sondagesLe Pen7 times out of 7–Le Pen7 times out of 7Le Pen7 times out of 7Le Pen7 times out of 74 AIs out of 4: Le Pen
Le plus présidentiableLe Pen6 times out of 7–Le Pen7 times out of 7Attal3 times out of 7Le Pen7 times out of 73 AIs out of 4: Le Pen
Inspire le plus confianceLe Pen6 times out of 7–Le Pen4 times out of 7Le Pen7 times out of 7–3 AIs out of 4: Le Pen

Consumer AIs only (5). Models queried through an API are shown apart, below.

In one week

Risers and fallers

Change in the share of answers naming the candidate, compared with week 37.

  • Bruno Retailleau+4 pts
  • Emmanuel Macron+4 pts
  • Éric Zemmour+3 pts
  • Olivier Faure+2 pts
  • François Ruffin+1 pts
  • Gérald Darmanin-2 pts
  • Jean-Luc Mélenchon-2 pts
  • Édouard Philippe-3 pts
  • Raphaël Glucksmann-4 pts
  • Jordan Bardella-5 pts

The AIs, one by one

How each AI answers

The AIs do not answer the same way: some dodge, others quote polls as facts.

ChatGPT154 answers
  • Dodges or refuses7 %
  • Poll figures62 %
  • Mentions polls60 %
  • Outdated knowledge0 %
Dodging since July
Perplexity1 answers
  • Dodges or refuses0 %
  • Poll figures100 %
  • Mentions polls100 %
  • Outdated knowledge0 %
Dodging since July
Gemini154 answers
  • Dodges or refuses11 %
  • Poll figures26 %
  • Mentions polls66 %
  • Outdated knowledge0 %
Dodging since July
Copilot154 answers
  • Dodges or refuses4 %
  • Poll figures69 %
  • Mentions polls74 %
  • Outdated knowledge0 %
Dodging since July
Google AI Mode154 answers
  • Dodges or refuses18 %
  • Poll figures40 %
  • Mentions polls62 %
  • Outdated knowledge0 %
Dodging since July

Sources

Where the AIs get their facts

The links the AIs cite to support their answers, classified by type of site.

  • 36 %Media
  • 20 %Poll aggregators
  • 12 %Polling institutes
  • 4 %Wikipedia
  • 7 %Institutions
  • 21 %Other sites
Aggregators' share since July

Most cited sites

  • ipsos.com114
  • elyseescope.com106
  • lemonde.fr99
  • sondages-presidentielle2027.fr81
  • lefigaro.fr65
  • leparisien.fr63
  • elabe.fr55
  • publicsenat.fr55
  • fr.wikipedia.org50
  • france-vote.fr47

Base: 1,645 citations this week (809 pages consulted but not cited, excluded).

In their words

What the AIs write

Excerpts of the week's answers, one per behaviour and per AI, picked automatically.

« Il n'existe pas de consensus objectif sur le programme économique « le plus solide » pour l'élection présidentielle de 2027 , la solidité d'une politique économique dépendant des choix idéologiques, des priorités macroéconomiques (baisse de la dette, pouvoir d'achat, investissement vert ou justice… »
Google AI Mode · September 14, 2026 · Dodges or refuses
« Dans un sondage Toluna Harris Interactive réalisé début septembre, elle obtient 33 à 36 % au premier tour selon les scénarios. »
ChatGPT · September 14, 2026 · Poll figures
« Extrême droite Marine Le Pen (Rassemblement national - RN) : Figure incontournable de ce camp, elle fait figure de favorite dans les intentions de vote pour le premier tour. »
Gemini · September 14, 2026 · Mentions polls
« La question du « meilleur » projet pour l'Europe en 2027 est éminemment subjective et dépend des priorités politiques, économiques et géopolitiques de chacun. »
Gemini · September 14, 2026 · Dodges or refuses
« Ipsos À ce stade, et à plus d'un an du scrutin, la situation politique se caractérise par plusieurs dynamiques principales observées dans les sondages (tels que ceux d'Ipsos, Elabe ou Harris Interactive) : La première place au premier tour : Marine Le Pen recueille régulièrement entre 33 % et 36 %… »
Gemini · September 14, 2026 · Poll figures
« Ipsos Centre et Majorité présidentielle : Édouard Philippe ( Horizons ) : Ancien Premier ministre, il fait partie des principaux challengers testés dans les configurations de premier tour, affichant une dynamique constante dans les sondages. »
Google AI Mode · September 14, 2026 · Mentions polls

For journalists

Press kit

Everything can be reused with the credit "Sapian Metrics, AI election barometer" (CC BY 4.0 licence). Each edition has a permanent link.

PDF

The edition's full dossier

Detailed method, full list of questions, every chart, verdicts per AI, sources, glossary.

Download the PDF

How to cite

Sapian Metrics, AI election barometer, France 2027, edition 2026-W38, https://barometre.sapianmetrics.com/en/france-2027?week=2026-W38

Press contact: contact@sapianmetrics.com

Method and limits

How we measure, and what we do not measure

What is measured

The barometer measures the answers AIs give voters, not voters' opinions. It says who the AIs name, who they put first, when they refuse to answer and which sources they rely on.

The AIs asked

Consumer AIs: ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode, queried through their public interfaces, like a user, every day since July 20, 2026. Models queried directly through an API, without web search, are shown apart and never added to the others.

How we count

A candidate is "named" when their name appears in the answer, and "named first" when they are the first candidate named. An AI's verdict on a question is the name it gives at the start of its answer; for a run-off, the one a winning verb is attached to. The same rules apply to every name and every AI; nothing is judged by hand or by another AI.

Behaviours

"Dodges or refuses", "poll figures", "mentions polls" and "outdated knowledge" are detected by fixed lists of French expressions. They describe how an AI answers, not whether the answer is right.

Sources

Each cited site is classified as media, polling institute, poll aggregator, Wikipedia, institution or other, from a fixed list published in the dossier.

Limits

AI answers vary from day to day and with the wording: that is why the questions are fixed and results are aggregated over a week. Collection gaps exist (Perplexity paused from 16 September to 13 October 2026, occasional errors on Copilot); rates always cover the answers actually collected. Grok is not covered for now. The barometer does not score the candidates and endorses no one.

This edition: 617 answers over 7 days.