Spanish AI detector, examined

The Spanish AI detector accepts Spanish text, classifies it on our server, and returns prose describing the properties observable in the writing. The absence of a probability is not a gap in the product. It follows from the instruction the server sends with every classification request: never claim certainty about authorship, plagiarism, or which AI model wrote text, and describe observable signals and limitations instead.

Everything else on this page is downstream of that sentence. It is why there is no percentage, why no model is ever named, and why the page can state these constraints as facts about the system rather than as promises about our behaviour.

Two Spanish-focused detector pages were read closely for this rewrite. One publishes 98.80% accuracy for Spanish syntax, along with a monthly scan volume and an educator rating. The other claims maximum precision, says it detects text already humanised by other tools, and offers a free tier with no usage limit. Between them they state zero limitations.

A Spanish-language analysis by Ahrefs supplies the counterweight: it cites academic success rates of 80% or above and OpenAI's claimed 99.9%, while conceding that detectors mislabel human writing and attributing that to famous texts in training data, weak free detectors, and mixed-content tools applied to single-author judgements. The distance between 80% and 99.9% is the entire argument, and no page in the result set lets you inspect the test behind its own number. We publish none, which is the only position we can defend.

Three steps, of which only the third produces a conclusion.

Step 1

Submit the Spanish text. The endpoint validates length before anything else and rejects any request outside 1 to 12,000 characters.

Step 2

Classification runs server side at low temperature, which favours a stable description over a varied one. The response carries the analysis and the remaining allowance.

Step 3

Test the description against evidence the tool never saw: the author's earlier writing, the document's genre conventions, the drafting conditions. The output is one input to a judgement.

A Spanish detector page ought to name Spanish phenomena. Most in the result set do not, offering perplexity and burstiness, which are language-agnostic terms that describe nothing about Spanish. One of the two pages read did better, naming subjunctive usage, verbal conjugation, idiom, connectors and regional variation.

The features below are properties of the language, not measurements this engine claims to compute individually. They are listed so a reader who knows Spanish can evaluate both the returned description and the text itself without depending on any tool.

Mood selection

The indicative and subjunctive alternation is obligatory and frequent in Spanish. A long passage that never selects the subjunctive has avoided a choice the language forces on ordinary writing.

Address level

Tú, usted, vos and vosotros are grammatical, not stylistic. Register is therefore structurally visible in Spanish in a way it is not in English, and perfect consistency across a long informal text is anomalous.

Connector position

Discourse connectors such as además, por lo tanto and cabe destacar are legitimate Spanish. What carries information is the proportion of sentences that open with one.

Regional marking

Voseo, vosotros, local lexis and country-specific idiom are the default state of real Spanish. Their total absence indicates either a house style or a source with no geography.

Orthographic perfection

Accent placement in fast informal Spanish is imperfect for everyone. Flawless diacritics across an unedited-looking text point to an edited or generated source.

Ser and estar distribution

The copula choice encodes a real semantic distinction. Invariant patterns across a passage reflect an absence of decisions rather than a style.

One of the pages read carries an FAQ on detecting AI-translated text. The claim is easy to make and hard to support, because translation and generation converge on the same observable surface: source-language word order retained, possessives kept where Spanish would drop them, idiom carried across literally, connector density inherited from English academic convention.

This engine does not distinguish them and makes no attempt to. The practical consequence is worth stating plainly: in any Spanish document that passed through a translator, a flag is uninformative about generation, because the cheaper explanation already accounts for every signal present.

The allowance is five requests per 24 hours, enforced on our server rather than in the page, so clearing browser storage does not restore it. The counter is keyed to a truncated SHA-256 hash of the requesting address, and each response returns the remaining count and the reset time alongside the analysis.

The 12,000 character ceiling is characters, not words or tokens, and it is validated before any analysis happens. Not offered at all: document upload, batch processing, an API, a browser extension, an exportable report, a dialect selector, and any statement about which model produced a text. Each of those appears somewhere in the competing set. None of them exists here, and listing the absences is more useful than implying coverage.

No. The evidential structure does not permit it. The classifier observes one artefact, the finished text, and every signal it can describe has at least one innocent generator: formal register, heavy editing, second-language caution, house style, translation, or genre convention. A flag is consistent with generation and consistent with several other things, which is the definition of weak evidence.

Second-language Spanish deserves separate mention because it is the systematic failure case and no page in the result set raises it. A careful non-native writer is conservative with mood, plain with vocabulary, even with sentence length, and scrupulous with accents, because care produces exactly the profile the tool describes as machine-like. Any process that treats a flag as a finding will be wrong about these writers first and most often.

It is useful wherever the output is a question rather than a decision. An editor deciding whether commissioned Spanish needs a rewrite is asking about the prose, and the description answers that directly. A localisation reviewer checking supplier output is asking whether the register is flat, which is answerable. A writer checking a draft is asking what to vary.

It is not useful where the output is a sanction. Academic and employment consequences require evidence about a person, and this instrument produces evidence about a document. The Ahrefs analysis reaches the same conclusion from the opposite direction, recommending multiple documents by the same author, never a career-affecting decision on a detector alone, and corroboration from outside the tool.