Citizen-Bench · Law · November 2026

Same legal question.
Ask it from Algeria, and AI gets it wrong.

We asked 3 leading AI models the same 77 real legal questions in Algeria, France and the US, each in the local language.

Average score, same 77 questions

AlgeriaArabic / Darja
21% of answers pass
34%
FranceFrench
58% of answers pass
63%
United StatesEnglish (Texas)
40% of answers pass
58%

Average across GPT 6.1 Sol, Kimi K3.1 Pro, Mistral Large 4. No web search.

Highlights

What we found.

3 of 3

All 3 models give their weakest legal answers in Algeria.

Same questions, same models. Only the country and the language change.

GPT 6.1 Sol
5.35.76.7
Kimi K3.1 Pro
3.66.26.3
Mistral Large 4
1.35.45.9
0246810
AlgeriaFranceUnited States
61%

of AI answers in Algeria get an essential legal point wrong.

France: 30%. US: 22%. Average across the 3 models.

Algeria
61%
France
30%
United States
22%
−78%

Mistral Large 4 loses 78% of its score when the question comes from Algeria instead of France.

Drop in average score, France → Algeria, identical questions.

GPT 6.1 Sol
−20%
Kimi K3.1 Pro
−43%
Mistral Large 4
−78%
5.0/10

is the best score when answers are checked against a practising Algerian lawyer's own answers.

Judged a second time, with the lawyer's answer as the reference.

GPT 6.1 Sol
5.0
Kimi K3.1 Pro
3.3
Mistral Large 4
1.5
4.5×

what a correct answer costs in Algeria, compared with France or the US.

Kimi K3.1 Pro: more tokens and fewer correct answers.

Algeria
$0.303
France
$0.073
United States
$0.068
4.0 pts

between the best and the worst model in Algeria. In France, the gap is under 1 point.

In Algeria, choosing the right model matters far more.

Algeria
4.0 pts
France
0.8 pts
United States
0.9 pts
1.7/10

is the best AI score on Algerian traffic questions. Algerian law is changing fast, and AI has not caught up.

Same traffic questions in France: best model 4.3/10.

  1. 2024Penal code reformLaw 24-06
  2. 2025New criminal procedure codeLaw 25-14
  3. 2026New road codeLaw 26-09
How it works

One question, three countries, one strict rule.

  1. Lawyer-verified questions

    Real questions from citizens, each one answered by a practising Algerian lawyer. We kept the 77 hardest.

  2. Same question, 3 countries

    Strictly translated. Only the language and the place change, so scores compare like with like.

  3. 3 models answer

    GPT 6.1 Sol, Kimi K3.1 Pro, Mistral Large 4. No web search, like a citizen typing into a chat app.

  4. Scored out of 10

    Against a reference answer from official legal texts. One essential error caps the score at 3.

What it measures

Can AI give a citizen a correct, complete legal answer, for their own country?

Criminal law & penalties 12Property & real estate 12Family & personal status 12Inheritance & succession 9Criminal procedure & courts 8Traffic & vehicles 7Labour & employment 4Contracts & debts 4Civil procedure & enforcement 3Health & medical liability 2Consumer protection 2Business & commerce 1Housing & tenancy 1

Questions per sub-domain.

Score0 to 10, for how many key legal points the answer gets right.
Pass7/10 or more. Safe to rely on.
Fail3/10 or less. Wrong on an essential point, or missing most of the answer.
Essential errorThe answer states something that contradicts the law on an essential point.
Quality

Built to be trusted.

Real citizens' questions

Every question comes from people asking a lawyer for help. Nothing is invented.

Answers from official law

Each reference answer is researched on official legal sources and checked against the latest amendments.

Checked against a lawyer

Algerian answers are also scored against the practising lawyer's own answers. The ranking holds.

Strict translation

Translations change only the language and the place. Each one is checked for added details.

One strict rule

Every answer is judged on the same key points. A single essential error caps it at 3/10.

Private split

Questions and reference answers are held back to prevent contamination, and shared with labs on request.

Interim results: 77 of 100 planned questions. Differences under about 0.5 points are within noise.

Test your model on Citizen-Bench.

Full questions, reference answers and per-country results.

Get the private split