Gemini Passive voice fixer limits

One constrained generation. Your text, 1 to 12,000 characters, plus an optional 300-character instruction, goes to our AI server. A model rewrites it under a prompt directing it to preserve the original meaning and factual claims, and to return the rewritten text alone with no preamble or commentary.

That last clause determines what this route can be. There is no parse tree, no rule engine, no sentence classifier and no annotation layer. Nothing marks a sentence as passive, because the system emits prose rather than analysis. Every limitation below follows from that, and so does the one advantage: the rewrite sees the paragraph, not just the sentence.

Purdue OWL's procedure is the canonical one and it has three steps: "find the agent in a 'by the...' phrase, or consider carefully who or what is performing the action", then "Make that agent the subject of the sentence, and change the verb accordingly", and then, when neither of the first two supplies an agent, "Sometimes you will need to infer the agent from the surrounding sentences which provide context."

Read as a specification, those three steps are not equivalent. The first is retrieval from the sentence. The second is a syntactic transformation. The third is inference, and inference produces content that was not in the input. Purdue's own worked example, converting "mistakes were made" to "we made mistakes", supplies a subject that the original sentence was constructed precisely to omit.

Any automated fixer is therefore performing two different operations under one name, and only one of them is safe by construction.

The instruction field is the only control, and using it to bound the operation is what separates a useful run from a destructive one.

Step 1. Submit the paragraphs you have already triaged

Up to 12,000 characters. Submitting an entire document delegates the retention decision to a model that has no stake in your meaning, and the retention decision is the one that carries risk.

Step 2. Bound the transformation

An instruction such as convert only sentences that already name their agent, and leave agentless constructions unchanged, restricts the run to the safe operation. Without it, the model performs both.

Step 3. Audit the output for new subjects

Read for grammatical subjects that are absent from the source. This is a targeted check with a specific target, and it is more reliable than rereading the passage for general quality.

Agentless passives are not a uniform category. They divide by why the agent is absent, and the cost of supplying one differs sharply between them.

Agent obvious from convention

Two burglars were arrested implies police. Just Publishing Advice notes that adding by police is unnecessary rather than wrong, so conversion here is merely redundant.

Agent genuinely unknown

The house was burgled. Converting yields A burglar burgled the house, which Just Publishing Advice flags as clumsy and which also asserts a single actor the sentence declined to posit.

Agent deliberately withheld

Incident reports, regulatory findings and legal summaries omit agents because attribution is contested or not yet established. Supplying one converts an unresolved question into a stated fact.

Agent structurally irrelevant

Markus was born in 1923. The sentence is about the subject, not the action, and an active version changes what the sentence is about rather than how it reads.

The strongest evidence against blanket conversion comes from the vendors themselves. Grammarly, which sells a passive voice checker, states that "Passive voice isn't a grammatical mistake, but it tends to make your writing unclear and hard to follow", and that the passive "does a better job of presenting an idea" when the performer "is very general, unknown, or should get less emphasis."

Just Publishing Advice adds the register argument: "In business, technical, and academic writing, there is more scope to use the passive voice", which describes most of the writing a tool like this is used on. Its closing advice, "Don't let your writing become unnaturally contorted just to satisfy a grammar checker", is the correct posture toward every product in this category, this one included.

Purdue, notably, does not discuss retention at all. The canonical instructional page on conversion contains no guidance on when not to convert, which is how a generation of writers acquired the belief that the passive is an error.

The market splits cleanly and the pages do not always make the split obvious. Sapling's passive checker is a detector, and describes its output precisely: given a list of sentences, it outputs whether each is passive along with the passive auxiliary. Rewriting lives at a different URL. Grammarly marks passives in place inside the surface you are writing in, requiring an account for full features and distributing through extensions and desktop applications.

This route is a rewriter with no detection layer attached. It cannot tell you which sentences were passive, cannot name an auxiliary and cannot mark anything, because the engine returns prose. If your process needs identification before transformation, a detector belongs upstream of this and the two jobs should not be expected from one request.

The compensating property is scope. A detector evaluates sentences; this rewrite sees the paragraph, which is the context Purdue's third step requires. That makes the inference better informed, and it does not make it safe.

No count, no proportion, no score appears anywhere in this route's output, and none is published on this page. Two reasons, one technical and one substantive.

The technical reason is that the engine emits rewritten text without commentary, so there is nothing to count from. The substantive reason is that no source in this result set publishes a correct target. Just Publishing Advice explicitly refuses a threshold in favour of contextual judgement; Grammarly names situations rather than a figure.

A proportion also measures the wrong thing. Two documents at identical passive percentages can differ entirely in whether those passives are load-bearing. A metric that cannot distinguish an evasive institutional passive from a deliberate methodological one is not measuring writing quality, and printing one would give a false impression of rigour.

Three domains concentrate the risk, and they share a property: the absence of a named agent is information rather than an omission. Incident and safety reporting, where attribution may be under investigation. Scientific method sections, where the convention exists to foreground procedure over personnel. Legal and regulatory drafting, where naming a party constitutes an allegation.

In all three, a converted sentence is not stylistically different from the original, it is evidentially different. The audit described earlier is not optional in these contexts, and for documents of any sensitivity the safer instruction is to restrict conversion to sentences that already carry their agent.

Elsewhere, in policy text, internal communication, marketing copy and general prose, the calculus reverses and the passive is usually concealing nothing worth concealing. That is the material this tool is genuinely useful on.