Portuguese AI detector, signals only

One server call carrying your text and an analysis instruction. That instruction asks the model to analyse only the supplied text for the qualities implied by the tool name, to give a concise result with clear findings, and then constrains it: never claim certainty about authorship, plagiarism, or which AI model wrote text, and describe observable signals and limitations instead.

Two consequences follow, and both are architectural rather than editorial. The refusal to assert authorship is imposed before generation, not appended as a caveat under a confident answer. And the response is prose, because nothing in the request asks for a number and no numeric field is parsed from what returns.

Sampling temperature on this route is 0.2, lower than the 0.55 used by the rewriting tools here, but not zero. Two analyses of the same text will broadly agree and will not be word for word identical.

The category standard is a number. One Portuguese detector returns a probability on a zero to one scale. Others return percentages. The number is the product, and the qualifications arrive afterwards in the FAQ.

Those qualifications are worth reading, because the vendors are more honest than the interface implies. One states outright that no AI detector is 100 percent reliable and describes its own results as indicative rather than definitive. Another warns against using detection alone for career or academic decisions. Both are correct, and both are placed several screens below a figure that will be screenshotted without them.

A number is portable and unfalsifiable by its reader. A finding is neither: it names a property of the text, so the reader can go to the text and see whether that property is there. That is the entire argument for emitting prose, and it costs the convenience of a figure to defend against, which is a trade this route makes deliberately.

Sample size dominates everything else about the result quality here.

Step 1

Paste the text, within the 1 to 12,000 character range the server enforces.

Step 2

Optionally constrain the analysis through the instruction field, up to 300 characters appended verbatim.

Step 3

Falsify each finding against the source text. Any signal you cannot locate should be discarded, not averaged in.

The tells reported in Portuguese-language writing on this subject are lexical and structural rather than statistical, which is precisely why they are checkable by hand. The connector stack is the most cited: além disso, por outro lado, portanto and em resumo, appearing at a density no unselfconscious writer sustains.

The second class is the contentless declarative. Two quoted as typical: este item é ideal para quem busca qualidade e durabilidade, and a inovação tecnológica transforma o futuro de maneira positiva. Each is well formed and carries no proposition a reader could disagree with, which is the diagnostic property. A third class is confident factual drift, with garbled detail about Portuguese history offered as the example, the Treaty of Tordesillas among them.

None of these is decisive alone. Each is countable, which makes the set more useful to a careful reader than any aggregate figure.

Error is not distributed evenly across writers, and that is the fact the category most consistently obscures. One Portuguese-language guide reports roughly 15 percent of natural Portuguese texts misclassified by automated detectors, and identifies the affected group precisely: authors with concise, repetitive or formal styles frequently trigger false alerts.

Read that as a description of populations rather than of individuals. Academic register is formal by convention. Technical documentation is repetitive by design. Second-language writers construct more regular sentences because regularity is what a learned grammar produces. All three are systematically more likely to be flagged, and none of the three is more likely to have used a model. Any process that applies a detector uniformly across such a mixed population is importing that bias wholesale.

No variant detection, no variant switch and no variant-specific tuning exists on this route. One prompt handles both. A competing detector advertises optimisation for Brazilian Portuguese; another makes no mention of the distinction. We state ours rather than leaving it to be inferred.

The distinction remains analytically interesting to a human reader. Internal inconsistency of variant, a text using você constructions throughout and then reaching for autocarro, is the kind of incoherence a single writer's habits do not usually produce, and it is available to inspection without any tool at all.

Five checks per rolling 24 hours, counted from first use against a hash of the requesting address, with no account and no paid tier in the product. Input is rejected outside 1 to 12,000 characters. Output is capped at 1,400 completion tokens, independently of input length, so a long document returns a short analysis rather than a proportionate one.

For calibration against the field: one Portuguese detector's free tier permits three uses per day at a 1,000 character maximum, with unlimited use behind a subscription, while its own copy notes that reliability improves with longer texts. Those two statements are in tension. Ours are not, though the 1,400 token output ceiling is a real constraint and is stated as one rather than omitted.

No. Not probabilistically, not by elimination, and not by naming a model. The instruction forbids claiming certainty about authorship, about plagiarism and about model attribution, and a competing Portuguese detector independently states that it does not identify which model generated a text.

Nor does any result here predict another service's verdict. No detector is run on our side beyond this single analysis call, no affiliation exists with GPTZero, Turnitin, Copyleaks or any other product named in this category, and no page here will forecast what one of them would return. What is produced is a reading of observable properties, which is a legitimate input to a judgement and is not itself a judgement.