Perplexity grammar checker and its limits

There are two kinds of tool in this category and they are frequently confused. A corrector marks your document, proposes a replacement and applies it. An analyser reads the document and reports. This is the second kind, and the distinction is not cosmetic: the correction never touches your text, so the judgement about whether to make it never leaves you.

Mechanically, your submission is sent to our server, analysed under an instruction to produce a concise result with clear findings, and returned as prose in a read-only box. Your original is unmodified. There is no document view, no inline marking and no apply action anywhere in the interface.

Grammar checking is narrower than most people assume, and knowing the boundary is what stops a clean result being mistaken for a good draft.

Inside: agreement

Subject and verb, pronoun and antecedent in number. Failures cluster where a long qualifying phrase separates the two.

Inside: sentence boundaries

Comma splices, run-ons and fragments. All three are boundary errors, which is why a writer who makes one often makes the others.

Inside: modifier attachment

Dangling and misplaced modifiers. Formally detectable, and the error most often preserved through a rewrite because it is grammatical in isolation.

Inside: punctuation convention

Including the restrictive and non-restrictive comma, which is a punctuation mark that changes what a sentence asserts.

Outside: reference

A pronoun whose antecedent was deleted is still a well-formed pronoun. No grammar check can know what it was supposed to point at.

Outside: warranted claim

Whether the sentence overstates its source is a question about evidence, not syntax, and no grammatical test reaches it.

Short procedure, with the constraint stated where it actually binds.

Step 1

Submit 1 to 12,000 characters. The ceiling is enforced in the request handler, so an oversized submission is refused rather than truncated without notice.

Step 2

Optionally scope the analysis with up to 300 characters of instruction: an error class, a register, or which English variant governs the punctuation.

Step 3

Apply the findings yourself. Each one is a proposal about your text, and you are the only party with access to what you meant.

Text produced with an answer engine tends to return few grammatical findings, because syntactic well-formedness is exactly what a language model is good at. The tempting inference, that the prose is therefore sound, does not follow.

Consider what a grammar check certifies. It certifies that sentences are well formed. It cannot certify that a sentence carries information, that a claim is supported by the citation attached to it, or that a pronoun refers to something still present in the document. Those are the three most likely defects in a research draft assembled from retrieved passages, and all three survive a grammar check unremarked.

This asymmetry is worth holding onto. Grammar checking has high precision on a narrow class of defects, and says nothing whatever about anything outside it. A clean result narrows the space of possible problems. It does not empty it.

The word in this route's name does two unrelated jobs. Perplexity is a commercial answer engine. Perplexity is also a measure of how well a model predicts a continuation, which is the concept the company named itself after.

Grammarly's own explainer of the term describes the product as an AI-powered research and conversational search engine, and glosses the concept as a measure of how well a model predicts the correct response to a prompt. It draws no connection to grammar checking anywhere, and there is none to draw. Predictability and grammaticality are orthogonal: a highly predictable sentence can be ungrammatical, and a startling one can be flawless. Any page that treats a perplexity figure as a proxy for writing quality has conflated two measurements.

Automatic application is the default in this category and it has a specific failure mode. When a correction is one click away, it stops being read, and some grammar corrections are semantic. The comma before a relative clause is the clearest case: adding it converts a restrictive clause to a non-restrictive one, which changes the proposition the sentence expresses. A tool that applies that on your behalf has edited your claim under the heading of punctuation.

The absence of an apply button here is partly principle and partly simply what exists: this route renders a remote analyser with two text areas. Stating both is more useful than dressing the second as the first. It is worth noting that the more careful commercial tools have converged on something similar from the opposite direction, promoting their editorial explanation of each change rather than the speed at which changes are applied.

Listed against what the category advertises, including the features that are simply absent here.

No inline correction

No underlines, no suggestion cards, no accept or reject. Your text is returned exactly as submitted.

No corrected output

There is no clean copy to download, because none is produced. The deliverable is a description.

No real-time checking

The analysis runs once, on demand, on the contents of the box at that moment.

No modes or document types

No Academic, Business or Creative presets and no strength setting. The 300-character instruction field replaces them and is more precise.

12,000 characters

Per run, enforced server-side. Roughly 1,800 to 2,000 words of English prose.

5 runs per rolling 24 hours

Counted by our server per visitor, rolling from the time of each run rather than resetting on a calendar day.

No accuracy figure

Competitors publish percentages. We have measured nothing of the kind and will not print a number we did not produce.

Academic writers at the stage where the argument is fixed and the prose is being finished, running a check over a section after the structural edits that create agreement and reference errors in the first place. Writers working in English as an additional language who want the rule identified rather than the sentence substituted, because the identification is the part that generalises.

And editors who have grown wary of accept-all. Anyone who has watched an automatic correction change a claim while presenting itself as punctuation has a reason to prefer a tool that can only tell them things.