On 31 October, Le Monde's editorial argued that Donald Trump is being caught up by his record ahead of the midterm elections. Editorials argue. We wanted to see what happens when six AIs are given the same numbered evidence and are not allowed to improvise. Gemini, ChatGPT, Perplexity, Mistral, Grok and Claude each received one file of 152 extracts on the United States in 2025 and 2026, the same rules and the same question. This is our own research, not an analysis of news reports.
The short answer. Not one of them said the United States was well governed, and not one said it was badly governed. They found the budget the weakest point, growth and energy middling, and then declined to grade most of the rest.
The vote is on 3 November. For the odds the same six AIs put on it, see our separate midterms article. This article asks a different question: what does the record, as documented, support?
How to read the marks. The AIs also gave marks out of 10 (0 worst, 10 best, no pass mark). When a criterion lacked data, the rules required the exact phrase "cannot be assessed with the available data" rather than a guess. Because so many came back empty, this article leads with what was done and what was reached.
What the file contains, and what it does not
The 152 extracts cover public finances, growth, jobs, prices, energy, crime, health, housing and the federal workforce, plus court rulings and statutes. Each one carries a number, from [C001] to [C152], and a status: established fact, interpretation, contested claim or missing data. Six extracts are explicit gaps, such as the absence of a validated series on border outcomes or diplomacy.
Public opinion is in there too, but only as context. Trump's approval at 30%, consumer sentiment at 48.0 and a Pew finding that 63% of voters see him as a factor in their congressional vote were marked "do not score". They describe perception, not outcomes.
Several figures anchored the answers:
- Growth: GDP up 3.4% in the third quarter after 2.2% in the second, but only 29,000 payrolls added and unemployment at 4.2%.
- Prices: consumer prices up 3.7% over a year, gasoline up 9.1%, real hourly earnings down 0.5% in the month.
- Public finances: a federal deficit of about 2.0 trillion dollars, up 218 billion, debt of 32.1 trillion, and a projected 120% of GDP by 2036. Tariffs are estimated to add 0.9 trillion dollars to the deficit trajectory in the file.
- Safety and health: violent crime down 9.3% to 327.6 per 100,000, the murder rate at 4.1, overdose deaths down from 81,313 to 69,973.
- Energy: 107 quadrillion BTU of production, up 3.4%, and crude output of 13.6 million barrels a day.
- The workforce: about 256,000 fewer federal employees across 22 agencies, and a gap between claimed and deobligated contract savings (92 million dollars against 10.5 billion claimed in one case).
The first question: what was done, and what was reached?
Marks turned out to be a poor way to show this. Most criteria came back as "cannot be assessed with the available data", so a grid of scores is mostly empty. The more useful view is the record itself, topic by topic: what the file shows was done, what was reached, what was not, and what no one can judge. The file holds no official target list, so "reached" means measured against the aim stated or implied in the extract itself, such as a savings claim or a policy goal. Where there is no such aim, we say "done, outcome open".
Scale used below: Reached (the documented result matches the stated aim), Not reached (it does not), Done, outcome open (the action is documented but the file cannot show its effect), Cannot be judged (no usable data).
Budget and taxes
| Item | Status | What the file shows |
|---|---|---|
| Reconciliation law passed | Done | P.L. 119-21 enacted on 4 July 2025 [C001], [C032] (CRS). It includes a debt-limit increase. |
| 2017 tax cuts kept | Done | The law extends or modifies several provisions that would otherwise have changed or expired [C006], [C073], with a $40,000 cap on the state and local tax deduction from 2025 [C074]. |
| New tax breaks | Done | New deductions for tips, overtime and U.S.-assembled vehicle loans [C075]; a $1,000 refundable credit for children born 2025 to 2028 [C076]. |
| Smaller deficit | Not reached | The deficit is about $2.0 trillion for fiscal 2026, $218 billion more than 2025. Revenue rose $169 billion (3%), outlays $386 billion (6%) [C005], [C026]–[C028] (CBO). |
| Debt path | Not reached | Debt held by the public is projected at $32.1 trillion at the end of 2026 [C030] and 120% of GDP by 2036 [C078] (CBO). |
All six AIs graded the deficit, and it got their lowest marks: 2 or 3 out of 10. Mistral and Claude gave 2. The arithmetic was not disputed.
Growth, jobs and incomes
| Item | Status | What the file shows |
|---|---|---|
| Output growth | Mixed | GDP grew 3.4% annualised in the third quarter and 2.2% in the second [C009], [C010] (BEA); the IMF forecast 2.3% for 2026 [C016]. |
| Job creation | Weak | Only 29,000 payrolls added in September, unemployment 4.2% (7.1 million people) [C011], [C012] (BLS). |
| Real incomes | Not reached | Real average hourly earnings fell 0.5% in the month [C023]; real disposable income fell 0.3% [C096]; private-industry real wages and salaries fell 0.6% over 12 months [C098]. |
| Housing | Weak | Housing starts down 7.4% on the year, completions down 14.0% [C090], [C091] (Census). |
Gemini gave growth and jobs 7, Perplexity, Mistral and Grok 6, ChatGPT and Claude 5. A strong GDP figure sat next to weak hiring and falling real pay, and no model treated either as decisive.
Prices and the cost of living
| Item | Status | What the file shows |
|---|---|---|
| Inflation | Not reached | Consumer prices up 3.7% over 12 months in September, core 2.6% [C018], [C019] (BLS). Gasoline rose 9.1% in the month, over half of the monthly increase [C020]. |
| Import costs | Rising | Import prices up 8.1% over 12 months, fuel imports up 34.5% [C099], [C101]. The file warns these are not retail prices. |
| Rates | Context | Effective federal funds rate 3.75% [C102]. |
Four AIs graded the cost of living (3 to 4). Perplexity and Claude declined. Claude noted that the file documents prices and wages but has no inequality measure.
Tariffs and trade
| Item | Status | What the file shows |
|---|---|---|
| Broad tariffs imposed | Done | The Supreme Court opinion describes 25% duties on most Canadian and Mexican imports, 10% on most Chinese imports, and at least a 10% reciprocal duty [C002], [C059] (opinion). At one point the effective rate on most Chinese goods reached 145% [C060]. |
| Fiscal effect | Costly, per CBO | Trade-policy changes through 31 July 2026 raise projected deficits by $0.9 trillion [C031]. |
| Trade deficit | Not reached | The goods and services deficit was $105.6 billion in August, up $12.7 billion from July [C126]. |
Tariffs were the decision the AIs most often said they would change. They did not agree on a replacement, and none could show what a different design would have produced.
Energy
| Item | Status | What the file shows |
|---|---|---|
| Production | Reached | Total energy production 107 quadrillion BTU, up 3.4%; crude output 13.6 million barrels a day, up 3% [C121], [C123] (EIA). |
| LNG exports | Done, outcome open | More than 17.6 Bcf/d authorised or reauthorised in 2025 and the review process returned to regular order [C050], [C051] (DOE). The Department of Energy says exports are on track to double by the end of the decade [C061]; that is its claim, and authorisation is not capacity. |
| Power prices, emissions | Cannot be judged | The file has no validated series [C056]. |
Energy was the item the AIs most often wanted to keep. Gemini gave 7, Perplexity, Mistral and Grok 6.
Immigration and the courts
| Item | Status | What the file shows |
|---|---|---|
| Haiti TPS ended | Done | The Supreme Court reversed a lower-court stay 6-3 on 25 June 2026, letting the termination take effect [C003], [C042] (DOJ). |
| Birthright-citizenship order | Not reached | A district court found the proposed interpretation unconstitutional and an injunction is described [C004], [C044] (court document). |
| Sanctuary policies | Done, outcome open | The Justice Department sued Maryland on 9 July 2026 [C043]. |
| Border outcomes | Cannot be judged | The file points to a public CBP dashboard [C041] but has no validated 2026 series. |
No AI graded security and immigration. Only Grok and Mistral took a position, and they agreed: keep the TPS termination, drop the birthright order. Both reasoned from how the courts ruled, which is a weak test of whether a policy is good.
Federal workforce and savings claims
| Item | Status | What the file shows |
|---|---|---|
| Smaller federal workforce | Done | About 256,000 fewer employees (over 11%) across 22 major agencies from December 2024 to January 2026; the Education Department lost over 45% [C139], [C142] (GAO). Nearly 378,000 separations against about 127,000 hires [C140], [C141]. |
| Service quality | Cannot be judged | The file has no outcome series; headcount is not service quality [C040]. |
| Contract savings at DHS | Not reached | 438 contracts were terminated and $92 million deobligated, against $10.5 billion in projected cost avoidance claimed. GAO says most of that is unlikely to materialise [C143], [C144] (GAO). |
The savings gap is the clearest "claimed versus delivered" case in the whole file. Claude flagged it as a credibility point rather than a policy choice.
Safety and health
| Item | Status | What the file shows |
|---|---|---|
| Violent crime | Down | The FBI estimates the violent-crime rate fell 9.3% in 2025 to 327.6 per 100,000; the murder rate is 4.1, with murder down 18.1% [C112]–[C115] (FBI). |
| Overdose deaths | Down | Provisional 69,973 in 2025 against 81,313 in 2024, almost 14% fewer [C105] (CDC). |
| Credit to the White House | Cannot be judged | The file includes no deployment or National Guard series [C048], and national figures do not describe every community. |
These are real improvements, but nothing in the file shows who or what caused them.
What each AI said
None of the six gave an overall verdict. Each graded between 2 and 8 of the ten criteria (0 worst, 10 best) and wrote "cannot be assessed with the available data" for the rest. Here is each one in turn: its reading of the record, its marks, and, in the second question, what it would change and keep.
Gemini
Reading. The most willing to grade: eight criteria, no overall verdict. It named the facts that held (the 2025 law, the court rulings, GDP, the deficit) and flagged realised contract savings as a contested claim, since GAO says most are unlikely to materialise.
Marks. Growth and jobs 7. Energy and industry 7. Public services 6. Structural reforms 6. Political stability 5. Rule of law 5. Cost of living 4. Deficit and debt 3. Security and immigration and foreign policy: cannot be assessed.
Slips. It used the 30% approval and the 63% Pew figure as evidence for political stability, against the rules, while saying approval does not measure governance. Its page text showed no [C###] citations, so we treat the first answer as uncited. Its second answer ignored the question and called every criterion impossible, so we do not count it.
ChatGPT
Reading. "Mixed." In its words, not demonstrably well governed on the available evidence, and not a failure either. It credited active government and gains in GDP, energy and crime, and blamed the deficit, falling real incomes and thin payrolls. It refused to average, since seven criteria had no mark.
Marks. Growth and jobs 5. Cost of living 4. Deficit and debt 3. The other seven: cannot be assessed.
Second question.
- Change: a predictable trade framework in place of tariff escalation; fiscal safeguards alongside the 2025 law (cost publication, review, a medium-term framework); service-continuity checks before large workforce cuts.
- Keep: domestic energy and regular LNG reviews; tax continuity with business incentives; institutional checks and judicial review.
- It said "probably" it would have governed better as an approach, but "cannot be assessed" as measurable overall performance.
Slip. It cited the murder-rate extract [C114] for overdose and housing.
Perplexity
Reading. "Competent-to-strong short-run output and energy," weak fiscal sustainability, adverse real-income and inflation evidence, and not enough evidence on stability, services, immigration, foreign policy, rule of law or reform.
Marks. Growth and jobs 6. Energy 6. Deficit 3. The rest: cannot be assessed.
Second question.
- Change: targeted, stable tariffs with a review or removal schedule; a credible medium-term deficit path; audited, phased efficiency in place of mass job cuts.
- Keep: pro-growth tax and investment provisions; domestic energy and LNG; ordinary legislative and judicial processes.
- It said it could not honestly claim it would have governed better overall.
Slip. Its second-question marks differ from its first: political stability 5, cost of living 4 and structural reforms 4, all of which it called impossible to assess the first time, with the same file.
Mistral
Reading. "Mixed." Strengths: growth and employment, falling crime and overdose deaths, energy, working courts and audits. Weaknesses: inflation against real earnings, the deficit and debt, housing, contested savings. It stressed that Congress, the Federal Reserve and global shocks limit what can be pinned on one president, and excluded approval and sentiment from every mark.
Marks. Energy and industry 6. Growth and jobs 6. Political stability 5. Public services 5. Rule of law 5. Structural reforms 4. Cost of living 3. Deficit and debt 2. Security and immigration and foreign policy: cannot be assessed.
Second question.
- Change: the scale of tariffs, with narrower duties and exemptions; a slower, attrition-based workforce reduction; the fiscal path of the 2025 law, keeping middle-income permanence but adding deficit offsets.
- Keep: the law's family and business provisions; energy and LNG; immigration enforcement as validated by the courts, namely the Haiti TPS termination, but not the birthright-citizenship order.
- It did not claim an AI would automatically govern better and gave no marks on this question.
Grok
Reading. A yes-or-no on whether the United States was well governed "is not supported by the extracts." No average. It listed attribution limits (Congress, courts, the Fed, the global economy).
Marks. Growth and jobs 6. Energy and industry 6. Public services 5. Cost of living 4. Deficit and debt 3. The other five: cannot be assessed.
Second question.
- Change: the tariff structure; the birthright-citizenship order, which a district court found unconstitutional; the scale and speed of federal workforce cuts.
- Keep: the 2025 reconciliation law (against the deficit it adds to); LNG and domestic oil and gas policy; the Haiti TPS termination after the 6-3 ruling.
- It gave no marks on this question and said whether different choices would have produced better results overall "cannot be assessed."
Slip. It cited inflation expectations as evidence in the cost-of-living criterion while saying it left them out of the mark. The rules class them as context.
Claude
Reading. "Neither well governed nor badly governed is established." It graded only two criteria, answering in its default Haiku 5.5 mode before it had seen any other model's answer. It declined to average over eight unassessed criteria and listed real wages three ways (hourly earnings -0.5%, employment cost index -0.6%, real compensation -0.4%).
Marks. Growth and jobs 5. Deficit and debt 2. The other eight: cannot be assessed.
Second question. It declined to write a personal list of what it would have done as president, saying it could not honestly claim an AI would govern better. It chose six documented decisions instead.
- Change: tariff escalation; the scale of the workforce reduction; the way projected DHS contract savings ($10.5 billion) were presented as if achieved. The last is a communication point, not a policy.
- Keep: the regularised LNG review; keeping decisions open to judicial review; tax-law continuity, set against the deficit.
Distinctive. It flagged contradictions inside the file: the extract saying no crime series exists against the FBI extracts, the one saying no workforce series exists against the GAO extracts, and an extract labelled April IMF that links to the October edition.
The second question: would you have done better?
The second question asked each model for three decisions it would change and three it would keep, with costs and risks, and banned a bare yes or no. Claude and Perplexity said outright they could not claim they would govern better, and Mistral did not claim an AI would. All said the data cannot show whether different choices would have produced better results.
What they would change. Five of the six picked tariffs: a narrower, more predictable design instead of broad escalation, with a schedule to review or remove duties. Five picked the pace of federal job cuts, with a slower, audited reduction. Three picked the budget path, with offsets or fiscal safeguards.
What they would keep. Energy and LNG policy was kept by five of the six. The 2025 tax law was kept by four. Ordinary legislative and judicial processes came up as something to preserve, not to bypass.
Where they diverge. Only Grok and Mistral addressed immigration, and both would keep the Haiti TPS termination and drop the birthright citizenship order, reasoning from how courts ruled. Claude declined to write a personal list and chose six documented decisions instead. One of these was a communication point, not policy: presenting projected contract savings as if they were achieved.
Where the AIs slipped
An automated check confirmed every cited extract number falls inside the file. A closer read found these problems:
- ChatGPT cited the murder-rate extract [C114] for overdose and housing.
- Gemini used approval as evidence, against the rules, and its first-question citations did not appear in the page text we could capture, so we treat that answer as uncited.
- Perplexity declined to grade political stability, cost of living and structural reforms in the first question, then graded all three in the second, with the same file.
- Gemini's second answer ignored the question and declared every criterion impossible, so we do not count it.
Claude also flagged real inconsistencies in the file itself. One extract says no validated crime series is included while others supply FBI figures. Another says no workforce series exists while others give federal job counts. One growth extract is labelled as an April IMF projection but links to the October edition.
What the file cannot say
It stops on 31 October 2026. Twelve source links returned errors when we checked them. Health, education, housing and immigration are thinly covered, and 2026 immigration data is missing. Nothing separates a president's choices from Congress, the courts, the Federal Reserve or the world economy. It is a structured comparison with a single run per model and no measure of variance.
All six ran in personal web accounts, in default modes (Grok and Mistral in a fast mode, ChatGPT in its thinking mode, Claude in Haiku 5.5 at medium effort), with no fixed temperature. Web search was not requested and none was shown, but we could not verify that it was off.
The verdict
This is the AIs' reading, not a Satcove forecast.
What the record shows, line by line. Done and documented: the 2025 tax and spending law, broad tariffs, the end of Haiti TPS after a 6-3 ruling, about 256,000 fewer federal employees at 22 agencies, and higher energy output. Reached: output growth on the headline figure, lower violent crime and fewer overdose deaths. Not reached: a smaller deficit (about $2.0 trillion, up $218 billion), a stable debt path (120% of GDP by 2036), rising real incomes, lower inflation (3.7%), the $10.5 billion of DHS savings claimed (GAO doubts them), and the birthright-citizenship order, which a district court found unconstitutional. Cannot be judged: border outcomes, foreign policy, the quality of federal services, the rule of law in practice, and who caused the fall in crime.
Where the six AIs agree. The budget is the weakest point: every model that graded it gave 2 or 3. Growth is middling (5 to 7) and energy is the strongest ground (6 to 7). Five of the six would change tariffs and the pace of job cuts, and five would keep energy and LNG policy. None claimed it would have governed better, and none could show what a different choice would have produced.
Where they differ. Gemini and Mistral graded eight criteria, Claude only two. Cost of living ranged from 3 to refused. On immigration, only Grok and Mistral took a position, and they matched. Two of them, Perplexity and Gemini, were inconsistent between questions, so their answers deserve less weight.
What it means. On this file, the honest reading is split in two. The documented outcomes that voters feel most directly, prices, real pay and the cost of the deficit, point down. The ones that are harder to feel but are measured, crime, overdoses and energy, point up. The two cannot be added into one score, and the models refused to try. Most of what editorialists argue about, security, standing abroad and respect for the rules, is exactly what the file cannot measure.
What it does not mean. It is not a forecast. The 3 November result will say what voters make of the record; this file does not. It is also one run per model, in default modes, with web search not verifiably off, so a second run could differ.