AI Detection for PR Professionals: Press Releases That Sound Authored, Not Auto-Generated

A press release is one of the few documents that’s supposed to sound official and a little stiff. That’s exactly the problem. The format is so formulaic that AI detectors — and sharp-eyed journalists — often read it as machine-generated even when a human wrote every word.

For a PR professional, the stakes aren’t a failing grade. They’re whether a reporter opens your next email, whether an editor trusts your byline on a contributed piece, and whether your client’s announcement lands as news instead of noise. A detector is one gauge worth watching. It isn’t the whole story.

Key takeaways

• Press releases score high for AI unusually often because the format itself — boilerplate, dateline, the obligatory CEO quote — is predictable by design.
• Detectors estimate how predictable your text is. They grade style, not authorship, so a high score on your own writing is common rather than damning.
• The real reputational risk isn’t a scan; it’s a journalist who can tell a quote was machine-written and quietly stops covering you.
• Check the release in pieces — you’ll usually find the news body is fine and only the quote and boilerplate flag.
• Fix the flagged parts with real specifics, not a one-click rewrite that flattens your meaning.

Why press releases trip AI detectors so easily

Most people assume a detector “knows” who wrote something. It doesn’t. It estimates how predictable your text is — how closely each word follows the patterns a language model tends to produce. Two stats carry most of the weight: perplexity (how surprised a model is by your word choices) and burstiness (how much your sentence length and rhythm vary). Smooth, even, predictable prose reads as more likely AI. If you want the math without the jargon, we cover the statistics behind AI detectors in plain English.

Now think about what a press release actually is. A dateline. A lead built from a standard template. A quote engineered to sound “on-brand.” A boilerplate paragraph you reuse across every release. A block of contact details at the bottom. Almost none of that is meant to be surprising — it’s meant to be conventional, scannable, and safe.

That’s the trap. The conventions that make a release look professional are the same ones that make it read predictable. So a release you drafted entirely by hand can still light up red, because the genre rewards exactly the kind of low-perplexity, evenly paced prose detectors flag.

The takeaway: the detector is grading texture, not origin. It’s also why no detector should be treated as proof. OpenAI retired its own AI Text Classifier in July 2023, citing a low rate of accuracy. A 2023 study from Stanford researchers (Liang et al., published in *Patterns*) found several popular detectors disproportionately flagged writing by non-native English speakers — likely because that writing tends to be lower in perplexity, the same way boilerplate is. These tools are signals, not verdicts.

What’s actually at risk for a PR pro

The detector score itself isn’t the threat. Nobody’s denying your wire distribution because GPTZero returned 80%. The threat is what a high score tends to track: copy that reads hollow, interchangeable, and obviously templated — the kind of release a journalist skims for two seconds and deletes.

Reporters have been reading press releases for years. Plenty of them can spot an AI-drafted quote without any tool, because the giveaway is so specific: an executive who supposedly speaks in flawless, perfectly parallel marketing cadence. “We are thrilled to leverage this transformative milestone to empower our valued customers.” No human says that. A detector flags it as predictable; a journalist flags it as fake. Same problem, two readers.

That’s where credibility erodes. Not in a percentage, but in the moment an editor decides your pitches aren’t worth the time. The score is a cheap early warning for an expensive problem.

A workflow that keeps releases sounding authored

Don’t run the detector on a rough draft, and don’t run it once at the end and panic. Here’s a sequence that fits how PR copy actually gets built.

1. Draft and lock the facts first

Get the release to near-final — the news is right, the numbers are confirmed, the quote is approved. Checking too early wastes effort, because you’ll rewrite anyway once legal and the client weigh in.

2. Check by section, not as one block

A single score for the whole release hides too much. Run it in chunks: headline and lead, the news body, the quote, the boilerplate. You’ll usually find the body — the part with real facts and figures — scores low, while the quote and boilerplate score high. That’s normal. It tells you precisely where the writing went on autopilot.

3. Rewrite the quote like a person talking

This is where most releases live or die. A flagged quote is almost always one drafted to sound impressive rather than human. Read it out loud. If your CEO would never say it at a dinner table, rewrite it until they might. Cut the stacked adjectives. Let it carry one real idea. A quote that sounds spoken reads lower to a detector and far better to a reporter.

4. Give the boilerplate something specific

Boilerplate repeats by design, so it’ll always flag a little. But a lot of it is also vague filler — “a leading provider of innovative solutions.” Swap one generic line for a concrete fact: founding year, a real customer count, the actual category you compete in. Specifics lower the score and make the company sound like it exists.

5. Do a human pass last

Tools find the flat spots. A person decides whether the release is accurate, on-message, and worth a journalist’s attention. No score gives you that judgment — and a release still has to be true no matter how the first draft was produced.

A quick mini-scenario

Say you’re handling comms for a Series B fintech announcing a new payments feature. You drafted the release with an AI assist to save time, then ran it through a detector. The whole thing comes back at a high AI likelihood, and you’re tempted to scrap it.

Instead, you check by section. The news body — the part explaining what the feature does, the rollout date, and a figure on transaction volume — scores low. The detector basically agrees it reads human. But the lead sentence and the CEO quote sit near the top of the scale.

The fix takes fifteen minutes. The quote was: “We are excited to revolutionize the payments landscape and deliver unparalleled value to our customers.” You rewrite it around something the CEO actually said in a meeting: “Our merchants told us settlement times were killing their cash flow, so we cut them from three days to one.” You tighten the lead so it opens with the news, not the adjectives. You leave the strong body alone.

The overall score drops, but that’s not really the win. The release now opens with a fact and quotes a human being. That’s the version a reporter pastes into their story.

What not to do with the score

A few traps PR teams fall into:

• Chasing 0%. Forcing weird phrasing into a document that’s supposed to be clean and quotable backfires. A natural, human release can land in the mid-range and be perfectly fine.
• Trusting one tool as truth. Two detectors will disagree on the same quote. Treat the number as directional.
• One-click “humanizing” the whole release. Auto-rewriters can loosen a stubborn paragraph, but run unattended they introduce odd word choices and can quietly change a fact — dangerous in a document that has to be accurate. Use them on a specific flat passage, then edit by hand.
• Fabricating to sound human. Adding a vivid “quote” the executive never said lowers the score and crosses an ethical line. Humanize the prose, never invent the facts.

If you’re building this into a team process, it’s worth seeing what fits your release volume before you commit a workflow — check the pricing first. And if you want to go deeper on detection mechanics and editing tactics, you can browse more on AI detection and writing.

Frequently asked questions

Do journalists actually run press releases through AI detectors? Some newsrooms and freelancers do, especially after a wave of obviously machine-written pitches. But the bigger risk is informal: an experienced reporter doesn’t need a tool to spot a quote that no human executive would ever say out loud. The detector is a proxy for a judgment they’re already making by eye. Treat a high score as an early warning that your copy reads generic, then fix the writing — that protects you whether or not anyone runs a scan.

Why does my press release score high for AI when I wrote every word? Because the press release format is inherently predictable, and detectors grade predictability. Datelines, boilerplate, and the standard “We are thrilled to announce” quote follow patterns a language model produces easily, so they read low on perplexity and trip the score. It usually means your release is formulaic, not that you cheated. Rewrite the flat parts with specifics and the number tends to fall.

Is it unethical to humanize an AI-drafted press release? No. Editing an AI draft so it reads like something a real communicator wrote is normal craft, not deception — you’re improving clarity and voice, the same thing a good editor does by hand. The line to avoid is fabricating quotes, stats, or events. A press release has to be factually true regardless of how the first draft was produced; humanizing the prose doesn’t change that obligation.

What part of a press release usually flags hardest? The executive quote and the boilerplate. Quotes drafted to sound “on-brand” often collapse into the same hollow, parallel phrasing across every release in the industry, which is exactly what a detector reads as machine-like. Boilerplate is repetitive by design. The news body — the part with actual facts, numbers, and a real announcement — usually scores lowest because it carries genuine information.

Should I aim for a 0% AI score on a release? No. Chasing zero forces awkward phrasing into a document that’s supposed to be clean and quotable. Detectors disagree with each other and output probabilities, not verdicts. The better target is a release that reads like a knowledgeable person wrote it — a quote that sounds like a human speaking, a body with concrete detail — and lets the score land where it lands.

Before it goes on the wire

Run the check, but read it like a communications lead, not a compliance officer. The score points at the lines that sound auto-generated — usually the quote and the boilerplate — and your job is to make them specific, spoken, and real. That’s what gets a release quoted instead of deleted. When you’re ready to see where a draft reads flat, run it through PaperBleach and start with the quote.