Every writing teacher says use active voice. Most writers know the rule. Switch "the ball was kicked by John" to "John kicked the ball." Simple.
That's not what this post is about.
This is about something more specific: the way AI models use passive voice in consistent, predictable positions for predictable rhetorical purposes — and how that pattern is one of the tells that AI detection systems look for. It's not that AI uses too much passive voice. It's that AI uses passive voice in the wrong places, for the wrong reasons, at exactly the moments a human writer wouldn't.
If you understand this, you'll edit AI output more effectively, write more convincingly, and watch your TextSight Humanization Score move in the right direction.
Two Kinds of Passive Voice
First, a quick frame.
Human writers use passive voice intentionally — for specific rhetorical purposes. The passive voice isn't wrong. It's a tool that serves particular goals:
- Shifting emphasis. "The report was submitted on time" — where the important thing is the report, not who submitted it.
- De-emphasizing agency. "Mistakes were made" — a classic. The speaker is declining to assign blame.
- Scientific convention. "The samples were analyzed using gas chromatography" — because the methodology matters, the analyst doesn't.
- Creating distance or formality. "It has been determined that..." — which feels impersonal by design.
These are legitimate rhetorical choices. The human writer is deciding to use passive because it serves a purpose.
AI passive voice is different. It tends to appear mechanically, in consistent positions, serving a specific and narrow rhetorical function: hedging. Softening the claim. Diffusing the sentence's commitment to any particular position.
The Specific Passive Constructions That Flag in Detectors
Here are the passive patterns that most reliably correlate with AI generation and pull scores down in detection:
"It is widely believed that..." — Nobody in particular believes it. AI uses this to introduce a common claim without committing to its truth or identifying who holds it.
"...has been shown to..." — Shown by whom? In what context? AI uses this to reference supporting evidence without linking to any.
"...can be seen as..." — Rather than asserting what something is, the AI frames it as a possible interpretation. Low commitment, maximum apparent thoughtfulness.
"...is considered to be..." — Again: considered by whom? The passive erases the agent.
"...are often used to..." — "Often" doing a lot of work here alongside the passive.
"...needs to be addressed..." — The problem exists. Someone should address it. Who? Nobody specific. AI uses this construction constantly in problem-solution writing.
The pattern is obvious once you see it: passive voice + vague frequency adverb + absent agent. This combination lets AI make claims that feel substantive without being falsifiable. It's rhetorical padding dressed up as careful writing.
What the Data Shows on Score Impact
We've tested this directly. Taking flagged paragraphs (scoring below 40 on TextSight's Humanization Scale) and systematically converting passive AI-hedges to active, specific claims.
The results: switching from passive to active in flagged sentences raises the Humanization Score by 3–8 points per rewritten sentence, on average. That's not huge on a per-sentence basis, but a typical 500-word AI paragraph contains 6–9 of these constructions. Convert them all and you're adding 18–50 points to a flagged passage.
Why does this work? Because the underlying signal isn't just grammatical structure. It's commitment. Active sentences with named agents are harder for AI to generate convincingly at scale because they require specific, falsifiable claims. "Research suggests" is easy. "A 2023 Stanford study found that 43% of..." is harder.
How to Tell AI Passive From Intentional Human Passive
The diagnostic isn't "passive = bad." It's "passive in service of what?"
AI passive: Appears in the middle of a claim to soften it. Often accompanied by hedges ("may," "can," "often," "generally"). The passive is hiding the agent because specifying the agent would expose the claim's vagueness.
Human passive: Appears for a deliberate reason — to shift emphasis, follow scientific convention, decline to assign blame, or create a specific rhetorical effect. The human writer could tell you why they chose passive here instead of active.
Test: find a passive construction in your text. Ask: "Could I add an agent here without the sentence becoming false or weird?" If yes — "this problem needs to be addressed" → "your IT department needs to address this problem by Q3" — then the passive was probably hiding something. Active is better.
If adding an agent would change the sentence's meaning in a way you don't want — "the treaty was signed in 1847" vs "James Polk signed the treaty in 1847" — then passive may be the right choice.
Before and After: Real Examples
Let me show you what this looks like in practice.
Original AI paragraph:
"It is widely believed that remote work has been shown to improve productivity, though this view has been challenged by recent research. The issue of work-life balance is often cited as a key consideration. More flexible arrangements are now being explored by many companies."
This passage has four passive constructions in 44 words. Count the absent agents: nobody specifically believes it, nobody specifically challenged it, nobody is doing the citing, and "many companies" is doing the exploring but tucked away at the end after the passive setup.
Rewritten:
"A 2024 Stanford study found a 13% productivity gain from remote work — but a follow-up from MIT complicates that picture. The real fight isn't about productivity. It's about whether employees and managers can trust each other when nobody's watching. That's the question companies are actually struggling with."
Same topic. Specific claims. Named sources. Active agents. No hedges doing structural work. The paragraph has a position.
TextSight scored the original at 31 (flagged). The rewrite: 74 (lower risk). Same information, different commitment level.
Second example — academic writing:
"The relationship between social media use and adolescent mental health has been studied extensively. Mixed results have been reported. It is suggested that individual factors may mediate the relationship."
Rewritten:
"Researchers have studied social media's effect on teen mental health for a decade and still can't agree on what they're finding. Some studies show harm. Others show the opposite. The honest takeaway is that it probably depends on the kid — how they use it, whether they're already struggling, and who's in their feed."
The passive constructions in the original were generating sentences that said nothing specific. The rewrite has a position: we don't know, here's why, here's what we can say.
What Active Voice Rewrites Actually Do to Writing
It's not just about detection scores. The passive-to-active edit is one of the most effective ways to improve writing because it forces you to think more clearly.
You can't write "it has been shown that X" in active voice without either naming who showed it or admitting you don't know. That's a productive constraint. The original passive construction let you skip that decision.
Similarly, you can't write "X needs to be addressed" in active voice without naming who should address it. "Your marketing team needs to address this by Q3" is a much more useful sentence than "this needs to be addressed." The active version is also more honest about what you're actually recommending.
Here's the thing: AI passive voice is a symptom of a deeper problem. It's not just a grammatical pattern. It's the absence of editorial commitment. Active rewriting forces you to have opinions and name them. That's what detection systems respond to. It's also what readers respond to.
The Passive Voice Patterns That Appear in Different AI Models
One interesting wrinkle: passive voice patterns differ somewhat across AI systems, and understanding this helps you spot the specific model's signature.
GPT-4o leans heavily on the "It is widely believed / it has been shown" family of constructions when generating explanatory content. These appear in roughly 1 of every 8 sentences in analytical writing — far more often than any human writer producing the same type of content.
Claude models (Anthropic) tend to use passive voice differently: less in the form of "it is believed" hedges, more in the form of passive conclusions. "More research is needed." "These findings should be interpreted carefully." The conclusion or recommendation often ends up passive even when the rest of the paragraph is active. It's a specific tell.
Gemini Pro generates what I'd call passive transitions — the passive construction appears at the start of a new paragraph to connect back to the previous one. "As was discussed above..." "Building on what has been established..." These backward-looking passive constructions are almost purely a Gemini signature.
This model-level variation matters practically: if you're editing AI output and you know which model generated it, you can run a targeted audit for that model's passive voice signature. But if you don't know — or if the text has been through multiple models — running the general passive audit covers all the variants.
The Quick Diagnostic
Run through this checklist on any text you're checking:
-
Count constructions starting with "It is..." or "It has been..." — These are almost always AI hedges. Flag every one.
-
Find every passive verb and ask "by whom?" — If the answer is vague or absent, it's probably an AI passive.
-
Look for passive + frequency adverb combinations — "is often used," "are frequently cited," "has been widely studied." These clusters are reliable signals.
-
Check the endings of paragraphs — AI often wraps up paragraphs with passive constructions: "more work is needed," "further research should be conducted." Humans end paragraphs with opinions, not deferrals.
If you're working on writing that a detector has flagged — or that you suspect might flag — a passive voice audit is one of the fastest edits you can make. It costs nothing, it improves the writing, and it moves detection scores meaningfully.
You can check the impact directly at TextSight. Paste your original, note the score. Make the passive-to-active edits. Paste the revision. The score difference will tell you whether the constructions you were using were dragging the humanization signal down.
They usually are.
Related reading: