Four AI models, one Spanish saying, four completely different answers: What just happened?

There is a Spanish expression that most English speakers, with a little thought, could probably map to something familiar:

“No hay mal que por bien no venga.”

Literally: there is no bad from which good does not come. The closest English equivalent most people land on is “every cloud has a silver lining.” But that mapping, while intuitive, is not the only defensible one. Depending on the context, register, and intended audience, the phrase could also be rendered as “it is an ill wind that blows nobody good,” “things happen for a reason,” “something good comes out of every difficulty,” or simply as a direct translation that preserves the Spanish construction entirely.

This is precisely what makes it a useful test input for AI translation. Not because it is obscure, it is anything but, but because it sits in a category of language where multiple correct answers exist, and the choice between them is not arbitrary. It reflects a judgment call about register, cultural familiarity, and translation philosophy. And AI models, it turns out, do not all make that judgment call the same way.

Last month, I ran the phrase through four major AI translation models simultaneously: ChatGPT, Claude, Gemini, and DeepSeek. What came back was not four versions of the same answer. It was four distinct choices, each defensible, each different in ways that matter depending on where the translation ends up.

The test: One phrase, four models, four different answers

The expression no hay mal que por bien no venga appears throughout Spanish literature, everyday speech, journalism, and business writing. It is well-represented in any corpus of Spanish text. Any model trained on a significant volume of Spanish-English parallel data or cultural commentary should, in theory, have encountered it many times.

Here is what each model returned when given the phrase in isolation, with no additional context:

• ChatGPT (GPT-4o): “Every cloud has a silver lining.”, Standard English cultural equivalent, idiomatic, widely recognized.
• Claude (Sonnet): “There is no misfortune that does not bring some good.”, Closer to the literal structure; preserves the original logic without adopting an English idiom.
• Gemini (1.5 Pro): “There is no bad from which good does not come.”, Near-literal, transparent to the Spanish structure, reads as slightly formal in English.
• DeepSeek (V3): “Every setback is a setup for a comeback.”, Interpretive and contemporary; conveys the optimistic spirit but departs significantly from the original register.

None of these outputs is wrong. All four are grammatically correct English. All four communicate something close to the meaning of the original Spanish. But they are not equivalent, and the differences between them are not cosmetic.

A translator working on a formal business document would not choose DeepSeek’s contemporary paraphrase. A literary translator would likely reject ChatGPT’s idiomatic substitution on the grounds that it replaces Spanish cultural content with English cultural content. A journalist writing for a general audience might find Claude’s near-literal version slightly stiff. Each output is optimized for a different context, and each model made that optimization silently, without asking which context applied.

Why culturally embedded phrases are a hard test for AI

Proverbs and sayings occupy a specific category within translation that is distinct from both technical terminology and everyday prose. Technical terms have correct and incorrect translations. Everyday prose typically has a fairly narrow range of idiomatic equivalents. But proverbs carry cultural meaning that can be legitimately transmitted in at least four different ways: through a culturally equivalent expression in the target language, through a literal translation that preserves the source structure, through an interpretive paraphrase that conveys the meaning, or through a hybrid that splits the difference.

The challenge this creates for AI is not primarily one of knowledge. A February 2026 study by Appen, which tested seven major AI models across 20 languages, found that models performed poorly on idiomatic language and puns regardless of how well those expressions were represented in training data. The problem was not recognition. It was consistent decision-making under interpretive ambiguity: models did not fail to understand the expression, they failed to apply a consistent strategy for handling it.

Slator’s 2026 analysis noted that even GPT-5, tested across five European languages, produced “recurrent issues in naturalness and idiomaticity.” Spanish, with its rich proverbial tradition and its status as the second most spoken native language in the world, presents exactly this challenge at scale. The Spanish-speaking world encompasses significant regional variation in how expressions like no hay mal que por bien no venga are understood and used, and that variation compounds the decision-making problem for AI.

A translation tool that feels reliable enough to depend on and one that actually produces consistent outputs across content types are not the same thing. The gap tends to surface exactly where cultural judgment and linguistic knowledge converge: in proverbs, sayings, and expressions where the surface meaning diverges from the operative one.

The divergence problem is not an error problem

What the four outputs above have in common is that none of them constitutes a translation failure in the conventional sense. There is no factual error. There is no hallucinated content. There is no grammatical breakdown. What exists instead is a divergence in translation philosophy that produces four different outputs from the same input.

The distinction between ChatGPT’s cultural substitution (“every cloud has a silver lining”) and Gemini’s structural preservation (“there is no bad from which good does not come”) is a genuine choice between two legitimate translation strategies. Translators have debated the relative merits of domestication versus foreignization, roughly, replacing source-culture content with target-culture equivalents versus preserving the source structure, for as long as translation theory has existed. These are not fringe positions. They represent fundamentally different views about what translation is for.

The four AI models did not reveal which strategy they were applying. They did not surface the fact that a strategic choice was being made. A user who does not speak Spanish received a fluent English sentence and, in most tools, had no mechanism for knowing that three other fluent English sentences representing different strategic choices were also possible.

This is the specific problem that this test surfaces: not that AI translation produces bad outputs, but that it produces outputs whose strategic orientation is invisible to the user. The variation between models is real and consequential. Its invisibility is the problem.

The opacity problem: When the choice is made for you

There is a broader pattern at work here that the AI trust problem in enterprise settings keeps circling back to. AI tools perform reliably on well-defined tasks with stable correct answers. They produce inconsistent, often invisible variation on tasks where multiple defensible outputs exist and the choice between them depends on context the tool does not have.

For translation teams working at scale, this inconsistency is not a theoretical concern. A content team producing Spanish-to-English materials for a global audience, marketing copy, employee communications, customer-facing documentation, is not agnostic between “every cloud has a silver lining” and “there is no misfortune that does not bring some good.” Those outputs carry different tonal registers, different cultural resonances, different suitability for different audiences. The team needs to know which output a given tool is producing and why.

When MachineTranslation.com processes the same input through its SMART system, running 22 AI models simultaneously and surfacing the output the majority of those models agree on, the phrase test produces something useful: the consensus output is visible alongside the distribution of individual model outputs. A user can see that the majority of models landed on the near-literal rendering while a minority chose the cultural equivalent. That visibility is not decorative. It is the mechanism by which the user can apply their own judgment about which output fits their context.

“It is no longer about finding the single best AI model,” said Ofer Tirosh, CEO of Tomedes, the company behind MachineTranslation.com. “It is about orchestrating a consensus among them to eliminate error.”

For technically precise content, that framing addresses the hallucination and accuracy problem. For culturally embedded content like no hay mal que por bien no venga, it addresses something different: the problem of invisible strategic divergence. Making the range of outputs visible does not eliminate the need for human judgment. It creates the conditions under which that judgment can actually be applied.

What consistency actually requires at scale

The standard benchmarking conversation in AI translation focuses on accuracy scores: percentage agreement with reference translations, performance on domain-specific terminology, fluency ratings across language pairs. Those metrics are meaningful, but they measure something different from what the phrase test measures.

A multi-model test published in 2026 that ran 22 AI translation models against the same inputs found that no single model was consistently reliable across all five target languages and all four content types in the test. Models that scored highly on composite accuracy benchmarks still failed on specific content categories. The composite score masked the per-category variance.

For proverbs and culturally embedded expressions, the relevant question is not which model scores highest in aggregate. It is which model, or which combination of models, produces the output that fits the specific use case. A content team publishing materials in multiple registers, for multiple audiences, across multiple channels, is not looking for the single best translation of no hay mal que por bien no venga. They are looking for the right translation for each specific placement.

That requirement is not met by picking a model and trusting its default behavior. It is met by surfacing the range of defensible outputs and preserving the user’s ability to select between them. Consistency, in content that admits multiple correct answers, does not mean every model producing the same output. It means every user having access to the full picture.

What this test leaves open

I want to be precise about what one test can and cannot show. Four models, one phrase, one language pair, one session. That is not a study. It is an observation that points toward a category of questions worth examining more systematically.

What it does demonstrate is that the divergence between major AI models on culturally embedded language is not marginal variation in phrasing. It is strategic divergence with real consequences for the outputs that end up in real documents for real audiences. And that divergence is, in most single-model tools, entirely invisible to the user.

The questions worth sitting with from here:

• Does the pattern of strategic divergence, cultural substitution versus near-literal preservation versus interpretive paraphrase, hold consistently across AI models for other Spanish sayings, or does it shift depending on the specific phrase?
• Does the model’s training emphasis on either fluency or fidelity predict which translation strategy it applies by default for this category of content?
• And at what point does making the range of AI outputs visible, rather than selecting a single output and presenting it as the answer, become the expected standard for translation tools working with culturally embedded language?

I do not have data-backed answers to those questions yet. But the phrase test suggests they are worth asking, and the finding that sparked them, four major AI models, one familiar Spanish expression, four meaningfully different English sentences, was clear enough to be worth putting into the conversation.