.jpg){kind=link}
Ask a pundit whether France was well governed between 2022 and 2026 and you will get a verdict before the question ends. We asked six AIs instead, and made it harder for them. Gemini, ChatGPT, Perplexity, Mistral, Grok and Claude each received the same file: 185 numbered extracts from INSEE, Eurostat, the Interior Ministry, the Court of Audit and public records of political events, stopping in October 2026. No web search. Every claim had to cite its extract. A grade from 0 to 10 was allowed only where the file held the data to support it.
Nobody said France had been well governed. What is more interesting is what happened next. The six models read the same figures and walked away with different diagnoses, different cures and, in one case, the opposite cure to their neighbours. This is our own research, not an analysis of news reports, and it describes what the machines said. It does not settle anything about France.
How to read the marks. They are out of 10, as in a French school report: 0 is the worst, 10 the best. This is not the American A-to-F scale, nor the French out-of-20 with its pass mark at 10. We set no pass mark, so a 5 is not a "pass". It is simply the middle of the scale. The ten criteria were: political stability, growth and jobs, purchasing power, deficit and debt, public services, security and immigration, energy and industry, crisis anticipation, transparency, and long-term reform.
The numbers nobody disputed
The file is stark on public finances. The deficit rose from 4.7% of GDP in 2022 to 5.8% in 2024, then eased to 5.1% in 2025, against about 3% in the euro area. Public debt climbed from 111.4% to 115.7% of GDP, while it fell in each of the four comparison countries in the file. Unemployment reached 8.3% in the second quarter of 2026, the highest since 2020. Barnier's government fell to a censure motion in December 2024, the first since 1962. The poverty rate stood at 15.4% in 2024, a series high.
On the deficit-and-debt criterion, every model that graded it gave a low mark: 2 out of 10 from ChatGPT, Perplexity, Mistral and Claude, and 2 then 3 from Grok.
Six readers, six temperaments
Mistral: the only near-verdict
Mistral was the only one to come close to a flat verdict: on the file alone, France "does not appear" to have been well governed. It graded all ten criteria, from 2 out of 10 for political stability and for deficit and debt up to 6 out of 10 for energy, averaging about 3.6 out of 10. It is also the model that bent the rule hardest, grading a criterion it had just called impossible.
Perplexity: stability, finances and growth fall short
Perplexity said France had not been governed satisfactorily on the best-documented points, namely stability, public finances and growth, and singled out electricity as the main success. In its first answer it declined four criteria. In its second it graded all ten, including the four it had refused.
ChatGPT: four grades, then none
ChatGPT graded four of the ten criteria: 3 out of 10 for political stability, 4 for growth and jobs, 3 for purchasing power and 2 for deficit and debt. It left the other six ungraded and concluded that the record was "partial and rather unfavourable" on what could be measured, while a definitive answer was "insufficiently established". Asked the second question, it then declared all ten criteria impossible to grade, arguing that the file supplies no scale for turning indicators into marks.
Claude: mediocre on what can be measured
Claude graded the same four criteria: 3 out of 10 for stability, 5 for growth and jobs, 4 for purchasing power and 2 for deficit and debt. It wrote that "well governed" cannot be demonstrated, whereas "governed with mediocre budgetary and political results on the measurable parts" can. It ran in our own session after reading the other five answers, which may have anchored it.
Grok: six grades, then one
Grok graded six criteria in the first round, from 2 out of 10 for deficit and debt to 6 out of 10 for energy, and refused any overall mark. In the second round it graded a single criterion, deficit and debt, and moved it from 2 to 3 out of 10, and declared the nine others impossible to grade, including some it had graded before.
Gemini: six grades, then nine
Gemini (2.5 Pro, queried through Google's API) graded six criteria in the first round: 4 out of 10 for growth and jobs, 3 for purchasing power, 2 for deficit and debt, 3 for public services, 4 for security and immigration and 6 for energy, an average of 3.7. It left four ungraded, including political stability, where it described a "very constrained" ability to govern but still wrote "impossible". In the second round it graded nine of ten, giving political stability 1 out of 10 and deficit and debt 1 out of 10. The same model moved stability from "impossible" to 1, and deficit and debt from 2 to 1, between two questions on the same file.
The second question: would you have done better?
We then asked each model to take the job: three decisions it would have made differently, three it would have kept, the likely effects, the risks and the political conditions. A bare "yes, an AI would do better" was ruled out.
One of the six, Gemini, came close to answering yes. The other five refused. Their alternatives point in different directions.
Gemini: the only near-yes
Gemini would have negotiated a minimal pension compromise without touching the legal age at first, kept the government in place rather than dissolve in June 2024, and launched an earlier budget plan combining targeted spending cuts and tax rises beyond the 2025 exceptional levies. It kept the temporary energy shields, the nuclear course and pension indexation. And it was the only model to answer yes in spirit: an AI, it wrote, would do better "not by magic" but by weighing risks coldly and favouring robust decisions over political gambles. It also leaned on outside knowledge it labelled "HORS CORPUS", such as the 2024 European election result, which the file does not contain.
ChatGPT: a multi-year budget and targeted aid
ChatGPT proposed a credible multi-year budget path with periodic revisions, a measurable purchasing-power target served by aid aimed at the most exposed households (poverty reaches 36.1% among the unemployed), and a territorial programme for jobs, housing and access to doctors. It stressed that its answer was an argued strategy, not a demonstrated superiority.
Perplexity: a majority before the budget
Perplexity would have built a budget majority before presenting budgets rather than reaching for 49.3, negotiated a gradual calendar for the 2023 pension reform with safeguards for long and arduous careers, and swapped blanket energy shields for aid targeted at low-income households earlier. It would have kept a gradual deficit reduction, low-carbon electricity and the 2024 immigration law.
Mistral: no dissolution, keep 64
Mistral would have skipped the June 2024 dissolution, which produced an Assembly of three blocs (193, 168 and 143 seats against 289 needed), started deficit consolidation of 0.5 to 1 point of GDP a year from 2023 through spending review rather than exceptional levies, and not suspended the pension reform, keeping 64 and compensating the most vulnerable. It concluded it would "probably" have governed differently, with hoped-for gains on stability and public finances, but that "better" cannot be established.
Grok: cut the deficit earlier
Grok would have cut the deficit earlier and harder toward euro-area levels, postponed or modulated the 2023 pension reform rather than pass it by 49.3 and see it suspended, and frozen or made conditional both the new levies (23 billion euros in 2025) and nominal spending growth. It would have kept the nuclear course, pension indexation and the practice of not dissolving after every censure motion.
Claude: negotiate upstream
Claude would have negotiated the pension reform upstream instead of using 49.3 with no majority, avoided the 2024 dissolution, and sought a budget compromise before 49.3 in 2025, even if it meant raising revenue earlier. It would have kept the energy course, pension indexation and the deficit correction under way from 2024 to 2025.
Where the six pull apart
On pensions, the machines contradict each other
Claude, Grok, Perplexity and Gemini would have softened the 2023 reform, by negotiation, postponement or a gradual calendar. ChatGPT would have kept the idea of reform but sold it differently. Mistral wanted the opposite of everyone: keep the reform at 64 and never suspend it, on the grounds that the suspension weakens public finances the Court of Audit already calls insecure.
On how to fix the deficit, they split between taxes and spending
Claude would have raised revenue earlier. Mistral and Grok would have leaned on spending review and a freeze on new levies. ChatGPT preferred to separate lasting savings from temporary revenue without choosing a side.
On the 2024 dissolution, three positions
Claude, Mistral and Gemini would have avoided it. Grok, on the contrary, listed "no dissolution after each censure motion" among the things it would keep. Perplexity and ChatGPT did not mention it.
On purchasing power, two schools
ChatGPT and Perplexity favoured aid targeted earlier. Mistral kept the energy shields and their gradual withdrawal.
Where they all agreed
All six would keep the nuclear and low-carbon electricity course (373 TWh of nuclear output, 95.2% low-carbon share in 2025). Five of six would keep the indexation of basic pensions to inflation. Perplexity was the exception.
What the file cannot say
The decision the models most want to rewrite is the one the file cannot price. Claude, Perplexity and Mistral each noted that it holds no figure for what the pension reform saves or what its suspension costs. Their alternatives therefore rest on a gap the file itself flags.
The other limits are the usual ones, and we state them plainly. Each model ran once, with no measure of variance: five in free web chats, in different modes and without a fixed temperature, Gemini through Google's API at temperature zero. An automated check found between 0% and 15% of cited figures missing from the cited extract, depending on the model and the question. Most were false alarms, such as "49.3", which is an article of the Constitution and not a statistic. Reading by hand, we found genuine slips, such as Mistral attaching a retirement-age figure to the wrong extract. The check tests numbers, not meaning, and a manual review of a sample of citations is still to come. Health, education, justice, housing and security are barely covered in the file, so no global verdict is possible, and nothing in it separates a president's choices from the economic cycle.
Six machines, one file
Strip away the model names and the exercise starts to look like a familiar French argument.
One camp would have softened the pension reform. Another would have held the line. One would have raised revenue; another would have cut spending. One would have spared France the 2024 dissolution; another would simply not dissolve the Assembly after every censure motion.
The six AIs rebuilt the shape of that debate from the same 185 documents, without being told which side was which. They did not settle it.
The verdict
France's 2022-2026 record looks weak where it can be measured.
Public finances deteriorated and debt rose. Unemployment climbed to 8.3% by mid-2026. A government was censured for the first time since 1962. On the one criterion where the evidence was dense enough for most models to grade it, deficit and debt, the marks were 2 out of 10 from ChatGPT, Perplexity, Mistral and Claude, and 2 then 3 from Grok. It is the clearest shared judgment in the test.
But the AIs could not show that a different government would have done better. They proposed alternatives: negotiate the pension reform instead of using article 49.3, correct the deficit earlier, avoid the 2024 dissolution. These are serious counterfactuals, and none of them can be tested in the file. The file shows what happened. It cannot show what a different decision would have cost.
The one thing they agreed to keep
All six would keep the nuclear fleet and the low-carbon electricity course. It is the one unanimous continuity, which suggests the models saw energy as one of the few areas where staying the course was easier to defend than reversing it. Even there, the agreement is conditional: the file documents outputs, 373 TWh and a 95.2% low-carbon share, not the trade-offs behind them.
What the test really showed
The most important result is not that the AIs judged France harshly. It is that they could not agree on how much they were allowed to judge.
Mistral graded all ten criteria. Gemini, Perplexity and Grok graded six, ChatGPT and Claude four. Then Gemini, Perplexity, Grok and ChatGPT contradicted themselves between the two questions.
Same documents. Same instructions. Radically different levels of confidence.
That is the real warning. An AI can tell you which figures moved: a censure motion, a rising deficit, a worsening labour market. It cannot reliably tell you whether a different political choice would have produced a better country.
What the six AIs conclude
Put the six answers side by side and the AIs' opinion is clearer than their methods.
Four lean negative, and none says France was well governed
Mistral: it does not appear to have been well governed. Perplexity: not satisfactorily on stability, public finances and growth. ChatGPT: a partial record, rather unfavourable on what can be measured. Claude: "well governed" cannot be shown, but mediocre budgetary and political results can. Grok gives no overall mark, but sees the public finances and late-period unemployment worse than at the start. Gemini gives no overall verdict, but marks deficit and debt 1 out of 10 and political stability 1 out of 10.
Their diagnosis is political as much as budgetary
Perplexity says it most plainly: the central problem is not only the deficit but the lack of a stable majority, which makes budgets hard to hold. Several models point to the same chain of events: a pension reform passed by 49.3, a dissolution that produced three blocs, then a censure and a succession of prime ministers.
Their one clear success is energy
Perplexity calls electricity the main favourable point, and Mistral and Grok mark energy 6 out of 10 or higher.
Their remedy is method more than policy
Build a majority before the budget, negotiate rather than force, correct the deficit earlier and more credibly. Even here they split on pensions, on taxes against cuts and on the dissolution.
And they stop short of the flattering conclusion
Five of the six refuse to say an AI president would have done better. Only Gemini argues it, by cold weighing of risks, and admits that this relies on knowledge outside the file. All of them know the file shows what happened, not what another choice would have cost.
So the AIs' answer to "was France well governed in 2022-2026?" is a qualified no on the measurable record, with energy as the exception, and an admission that the rest cannot be judged from the evidence they were given. Read it as six opinions that largely converge on the diagnosis and diverge on the cure. It is not an audit.
To put your own question to several AIs at once, try Satcove Open. To see how we compare AI answers, read how Satcove compares AI answers.