AI Detector for Press Releases: Why PR Boilerplate Looks AI-Generated

A communications manager runs the quarterly product announcement through an AI detector before it goes out on the wire. The result comes back at 91% likely AI-generated. Nobody used ChatGPT. She wrote the whole thing herself, the way she’s written press releases for a decade.

So what happened? Nothing wrong with her writing, actually. Press releases are one of the most rigidly formulaic genres in business communication, and that formula looks a lot like what a machine produces. This is the genre most likely to set off a detector even when a person typed every word.

Key takeaways

• Press releases follow such a tight template that they often score high on AI detectors regardless of who wrote them.
• Detectors flag *predictability*, not authorship. Boilerplate, stock quotes, and PR cliches are exactly the low-variety text that reads as machine-made.
• A high AI score is a probability, not proof. On a press release, it says more about the genre’s conventions than about the writer.
• You can cut the false-positive risk by varying sentence length, trimming filler, and writing quotes that sound like a person actually talking.
• The smart move is to find the predictable lines, then revise those specific passages instead of torching the whole draft.

What an AI detector actually measures

This trips people up, so it’s worth being blunt. An AI detector doesn’t have a window into how text was created. It can’t see your keyboard, your prompt history, or your intent. It reads the words and estimates how *surprising* they are.

Two statistical ideas do most of the work. The first is perplexity, roughly a measure of how predictable each next word is given the words before it. The second is burstiness, which captures how much sentence length and rhythm vary across a passage. Machine-generated text tends to be low on both: smooth, even, predictable. Human writing usually jumps around more. If you want the full breakdown, we wrote a whole piece on perplexity and burstiness.

Here’s the catch for PR. A press release is *engineered* to be low-perplexity and low-burstiness. It’s supposed to be smooth, neutral, on-message. You’re not trying to surprise anyone. You’re trying to deliver facts in a familiar shape so a busy journalist can skim it in fifteen seconds. The genre’s whole job is to read like every other press release. That goal happens to be the statistical fingerprint of AI text.

Why boilerplate is the worst offender

Most of a press release can pass as ordinary human writing. The boilerplate almost never does.

Boilerplate is that standardized “About [Company]” paragraph at the bottom. Legal and brand teams approve it once, then it gets pasted into every release verbatim, for years. By design it’s a dense pile of mission-statement language: *a leading provider of innovative, end-to-end solutions empowering organizations to unlock their full potential and drive transformative growth across global markets.*

Read that again. Every clause is the most expected possible continuation of the one before it. “Leading provider” pulls in “innovative.” “Solutions” pulls in “empower.” There’s almost no surprise in the sequence, which means very low perplexity. The sentence is also one long even clause-chain, so burstiness sits near zero. A detector scores that paragraph as overwhelmingly machine-like, and honestly, it’s not wrong about the *style* even if it’s wrong about the *author*.

The irony is hard to miss. Companies spend years sanding their boilerplate down to maximally generic corporate English, and that very polish is what makes it look synthetic. It’s a cousin of the over-polished problem: text so smooth and inoffensive that it loses any fingerprint of a specific human mind.

The quote problem

The other reliable culprit is the executive quote.

Look at a typical one: *”We are thrilled to announce this milestone, which underscores our unwavering commitment to delivering best-in-class value to our customers and stakeholders,” said the CEO.* No CEO talks like that. A PR writer assembled it from approved phrases, and it shows. It’s frictionless, adjective-heavy, and utterly predictable, which is exactly what pushes a detector’s score up.

Compare it to something closer to how a person actually speaks: *”We’ve wanted to ship this for two years. The first version was clunky, and customers told us so. This one isn’t.”* Short sentences next to longer ones, a concrete number, an admission, a fragment of personality. Higher burstiness, higher perplexity, far more human-looking, and frankly a better quote. The writing fix and the detector fix point the same direction.

A mini-scenario: the announcement that scored 91%

Back to our communications manager. Picture her running the release through a detector that gives a sentence-level breakdown instead of one scary number. The body of the release, the part describing the actual product and the rollout dates, scores low and reads as human. Good. The two hot spots are predictable: the manufactured CEO quote and the boilerplate.

She doesn’t rewrite the release. She rewrites two paragraphs. The quote gets a real voice and a specific detail the CEO genuinely said in the kickoff meeting. The boilerplate she can’t fully change because legal owns it, so she at least breaks the single monster sentence into three plain ones with varied lengths.

Re-run, and the score drops sharply, with the remaining flag landing on the boilerplate she wasn’t allowed to touch. That’s the realistic outcome. You won’t always get to zero, because some constraints aren’t yours to override. But you can usually move the needle a lot by fixing the two or three passages doing the damage.

So is a flagged press release a problem?

Usually less than people fear. A few things to keep straight:

• A score isn’t a sentence. It’s a probability. A 91% on boilerplate tells you the prose is generic, not that you cheated. There’s no court of PR that bans you for a number.
• Distribution and search care about usefulness, not detector scores. Wire services and search engines aren’t running your release through GPTZero and rejecting it. What can actually hurt you is publishing thin, near-duplicate, mass-templated content at scale. One templated boilerplate line won’t do that. A hundred copy-paste releases with nothing original might.
• False positives are baked into this genre. Detectors are known to misfire on formulaic writing. OpenAI retired its own AI text classifier in July 2023 because its accuracy was too low to be useful, and a 2023 Stanford study (Liang et al.) found detectors disproportionately flagged writing by non-native English speakers. Press release conventions sit right in that high-false-positive zone. Treat the tool as a writing-quality mirror, not a lie detector.

When you should care, and when you shouldn’t

Care if the release reads thin, if you’re publishing many similar ones and want to avoid looking like a content farm, or if a client specifically asks for a clean detector score. In those cases, the score is a useful proxy for “does this sound like a human bothered.”

Don’t lose sleep if a single approved boilerplate paragraph lights up. That paragraph is doing its job, which is being identical everywhere. Consistency was the point. You can’t have legally-vetted uniformity and high burstiness at the same time, and uniformity often wins for good reasons.

How to make a press release read more human

A handful of moves cover most of it:

• Vary your sentence length on purpose. Put a five-word sentence next to a twenty-five-word one. That contrast alone raises burstiness.
• Cut stacked adjectives. “Innovative, scalable, best-in-class, end-to-end” is four words doing the work of zero. Pick one true thing or none.
• Make quotes sound spoken. Use contractions. Include a real detail or number. Let the person admit something or react, not just affirm the brand.
• Add one concrete fact early. A specific date, figure, or named constraint breaks the predictable flow faster than any amount of rephrasing.
• Diagnose before you rewrite. Find the hot passages and fix those. Rewriting clean paragraphs just wastes time and can make them worse.

If a tool fits into your regular workflow, it’s worth knowing what the free check covers versus what you’d see what a paid plan covers. And if you want to dig deeper into how detection works across different writing situations, there’s more on detection and writing worth a read.

Frequently asked questions

Why does my press release get flagged as AI when I wrote it myself? Because detectors measure how predictable text is, not who produced it. Press releases follow a tight template: the formal lead, “is pleased to announce,” a quote full of corporate adjectives, and identical boilerplate at the end. That’s low perplexity and low burstiness, the same signature models produce. A human writing to convention lands in the same statistical zone as a machine.

Does a high AI score mean Google or the wire will penalize my release? Not directly. Distribution services and search engines don’t reject content over a detector number. They care about whether it’s useful, accurate, and not mass-produced spam. A flagged boilerplate paragraph won’t sink a legitimate announcement, though thin, duplicated, templated content across many releases can erode how a domain is judged over time.

Should I rewrite the boilerplate so it stops getting flagged? You can, but weigh the tradeoff. Boilerplate exists so an approved company description appears everywhere, for legal and brand reasons. If consistency matters more than the score, leave it. If you want it more natural, write a few real, varied-length sentences about what the company does instead of one dense clause-pile.

Can a detector tell a human PR writer from ChatGPT? Reliably, no. Detectors output a probability from statistical patterns, and press release conventions overlap heavily with model output. That overlap is exactly why false positives are common here. A detector can flag a passage as predictable; it can’t prove who or what wrote it.

Closing

A flagged press release is rarely a sign you did something wrong. More often it’s a sign the genre itself has been optimized into something that reads like a machine, decade after decade of approved phrasing and pasted-in boilerplate. The fix isn’t to panic over a percentage. It’s to write quotes a human might actually say, vary your rhythm, and trim the cliches that were never adding anything.

Want to see exactly which lines in your next release read as predictable before it ships? Run your draft through PaperBleach and revise from there.