We Tested Four AI Models on One French Sentence. Here’s What Happened.

The French Register Paradox

French is the most-tested language in AI translation benchmarking. It is a tier-one target language for every major model, included in every evaluation dataset, and cited in most capability claims. When an AI Models translation tool says it supports European languages, it is implicitly saying: French is covered.

That assumption is worth testing directly.

French has a formality system with no clean English equivalent. The vous/tu distinction encodes the relationship between speaker and listener at the grammatical level, not as a stylistic choice the writer makes separately from the sentence structure. French also uses gendered nouns, agreement rules that ripple across adjectives and past participles, and a set of register markers that distinguish administrative, commercial, literary, and conversational French in ways that translate-then-review workflows routinely miss.

The result is the same structural divergence we documented for the same structural variance we documented for Korean: models that are all technically correct but interpretively inconsistent, producing outputs that cannot be interchanged without changing meaning, tone, or brand voice.

The French Register Test for AI Models

Consider the English sentence: “We will keep you informed of any changes to your account.”

It is grammatically neutral, professionally standard, and common in financial services, SaaS, and customer communications. In French, it requires a choice the English source does not make: is the communication formal (vous) or familiar (tu)? Does the construction use a future tense, an aller + infinitive construction, or a nominalized form? Does the passive imply institutional distance, or should the verb be made active to read as more direct service communication?

Four leading AI models, each tested independently on the same input, returned four distinct French outputs:

• GPT-4o: “Nous vous informerons de tout changement concernant votre compte.”
• Claude Sonnet: “Nous vous tiendrons informe(e) de toute modification apportee a votre compte.”
• Gemini 1.5 Pro: “Nous vous informerons de tout changement lie a votre compte.”
• DeepL: “Nous vous informerons de toute modification de votre compte.”

All four are grammatically correct. All four use vous, which is the appropriate formal choice for account communication. But they are not equivalent. GPT-4o and DeepL use a direct statement of future action. Claude’s output includes a gendered agreement marker (informe/informee) that the others omit, which matters for accessibility and demographic representation in French-language legal communications. Gemini uses “lie a” (linked to) rather than “concernant” (concerning) or “apportee a” (brought to), producing a different conceptual framing of the account relationship. DeepL drops “changement” in favor of “modification,” a distinction that carries different legal weight in contract and regulatory contexts.

In a newsletter, these differences are editorial preferences. In a banking notification, a legal agreement, or a healthcare data consent form, they are interpretive choices that affect meaning and could affect compliance.

Why the Four AI Models Diverge

The divergence follows the same training-driven pattern that model variance across tasks produces in multimodal evaluation contexts. GPT-4o’s fine-tuning emphasis on direct, professional English prose carries forward into French as a tendency to flatten register choices into clean transactional language. Claude’s reinforcement training rewarding natural warmth produces the gendered agreement inclusion that the other models skip. Gemini’s closer mapping to source structure results in “lie a” as a calque of “related to” rather than a native French construction. DeepL’s domain-specific tuning for business language in European pairs produces the cleanest output for standard business correspondence but drops the nuance that distinguishes “changement” from “modification” in legal registers.

These are not errors. They are different models making different reasonable choices in response to the same interpretive gap. As distinct interpretive tendencies that widen rather than converge as models scale, the divergence on ambiguous source text does not resolve itself. It becomes more structured, more consistent within a given model, and therefore harder to detect through fluency checks alone.

French compresses this problem because it appears solved. Unlike Korean, which signals its complexity through a structure that is overtly foreign to European language speakers, French reads as familiar. An English-speaking reviewer who is not a native French speaker will not catch the informe(e) omission. They will not notice that “lie a” is subtly less idiomatic than “concernant.” The fluent output passes the review that Korean would fail.

What This Means for Teams Building French-Language Workflows

The Lokalise 2026 research on the best LLMs for translation reaches the same conclusion as the Korean analysis: no single model wins across all language pairs and task types. For French, the practical implication is specific. Teams that choose one model for all French content are not choosing accuracy. They are choosing a consistent interpretive framework that may or may not match the register their French-speaking audience expects, and that will produce outputs their English-speaking reviewers cannot evaluate.

The 2026 Crowdin enterprise translation survey found that 1 in 5 organizations reported quality incidents after introducing AI translation. French was not flagged as a high-risk language in that survey, which is part of the problem. Korean, Japanese, and Arabic generate review requirements because teams know they cannot evaluate the output. French generates assumptions that the output is fine because teams believe they can.

The verification cost is hidden in French in a way it is not in structurally foreign languages. Internal data from teams tracking translation workflow time shows that non-linguists using single-model AI for French spend a significant portion of their AI translation time manually comparing or correcting outputs rather than publishing them directly. That is not a speed advantage. It is a verification backlog that does not get measured because it looks like editorial time.

How Running Multiple Models Against Each Other Changes the Reliability Equation

The account notification sentence above has a correct French translation. It depends on context that none of the four models had access to: the formality convention of the sending institution, the legal jurisdiction governing the account, the demographic characteristics of the recipient, and the branding guidelines of the product. Without that context, each model makes a reasonable assumption. With that context, only one of the four outputs is right.

The divergence between the four outputs is not a problem to resolve by picking the best model. It is information. When GPT-4o, Claude, Gemini, and DeepL agree on a French output, that agreement is stronger evidence of correctness than any single model’s confidence score. When they diverge on a word choice that carries legal or tonal weight, that divergence is a signal that the source text has an ambiguity requiring human resolution.

MachineTranslation.com, which recorded a 230% spike in English-to-French translation volume in a single week in June 2026, applies this logic through its SMART mechanism, which compares outputs across 22 AI models and selects the translation that the majority agree on. For French, where register divergence is structural and fluency checks do not surface it, the consensus approach surfaces disagreements rather than silently resolving them in favor of one model’s interpretive default. When the platform processed a formal French sentence against five AI models, they reached 100% agreement in 6.7 seconds with zero disputed terms. Internal benchmarks show critical translation error rates below 2%, compared to the 10% to 18% range reported for individual models on complex content.

Ofer Tirosh, CEO of Tomedes, a translation company that developed the platform, frames the principle directly: “The question teams should be asking is not which AI model is most accurate in the abstract for French. It is: when the models disagree on a vous/tu choice, a legal term register, or a gendered agreement that only shows up in formal French, which disagreement structure tells you where the translation risk actually sits?”

Building a More Defensible French AI Workflow

Three principles follow from the analysis above, consistent with the guidance derived from the Korean case:

• Test your specific content type, not a benchmark ranking. French marketing copy and French financial notifications are different translation tasks. A model that produces excellent French product descriptions may handle legal register poorly. Benchmark rankings do not differentiate at this level.
• Treat gendered agreement and register as review triggers. Content containing gendered recipient references, formal/informal register choices, or legal terminology in French should flag for human review regardless of output fluency. The cases where models diverge most significantly on French are precisely these cases.
• Make the disagreement visible. Single-model workflows hide the interpretive choices that multi-model comparison would surface. If four models agree on a French output, you have stronger evidence than one model’s confidence score. If they disagree in a way that matters, you have a review case that a fluency-only check would pass silently.

The French sentence that opened this piece has a correct translation. It depends on context the model does not have unless you provide it. The four divergent outputs are each a hypothesis about what that context might be. The workflow question is whether you want to see those hypotheses or trust that the first one is right.