AI Press Release: Fix Copy That Reads Machine-Made

If your AI press release reads machine-made, the fix isn’t a full rewrite. It’s finding the handful of sentences that carry the machine signal and rewriting only those. Here’s how to do it, why it matters more this year, and where a draft editor fits.

Key takeaways

• A press release gets flagged when its sentences are uniformly smooth, evenly paced and free of specifics, not because a detector “knows” who typed it.
• Detector scores are probability estimates, not proof of authorship. Treat them as a triage signal, not a verdict.
• In our own measurement on 2026-10-04, across n = 1504 hand-written DeAIze guides, the mean AI score was 27% and 487 of 1504 were flagged at a 30% threshold. Human writing is not automatically clean.
• Fix the flagged sentences, not the whole document. Two or three edits usually move the reading more than a global paraphrase.
• Keep the news. A press release that loses its date, quote or attribution to sound human is a worse press release.
• Check the draft at sentence level before you send it to a journalist, a client or a wire service.

Why does your AI press release read machine-made?

A press release has a narrow job: announce something, attribute it, give the reader a reason to care, and provide contact details. That structure is rigid by design, which is exactly why AI-generated press releases look plausible at first glance and hollow at second.

Detectors don’t read meaning. They look at statistical texture. Uniform sentence length, low variation in how clauses start, predictable transitions, and a lack of concrete nouns all push a passage toward a high machine probability. When every paragraph opens with the same rhythm and every claim is balanced against a matching counter-claim, the text reads like a template filled in.

The tell isn’t one word. It’s the absence of friction. A real announcement has an awkward detail: the product name that doesn’t fit the sentence, the number the spokesperson actually said, the date that changes the tense halfway through. Machine drafts smooth all of that away.

That’s why the same passage can score high on one detector and borderline on another. They’re measuring overlapping signals with different weights. If you want the mechanics, our guide to how AI content detectors work walks through the signals without the marketing gloss.

What changed outside your draft that makes this urgent now

You’re not fixing this in a vacuum. Two things shifted recently.

First, disclosure rules. The EU AI Act’s transparency obligations for AI-generated content have been phasing in, and platforms from search engines to social networks now ask publishers to label synthetic media and, in some cases, synthetic text. A press release is a published artefact. If it was drafted with a model and not edited, you may have a disclosure question on your desk that didn’t exist a couple of years ago. Our piece on the AI content disclosure label covers what to say and when, including the cases where no label is required.

Second, journalists and editors are now running their own checks. Not always with a formal detector, but with the same instinct: does this pitch sound like it was written by a person who knows the story? A wire-service editor who receives a release that reads machine-made doesn’t need a score to bin it. They just don’t reply.

So the practical question isn’t “will I get caught?” It’s “does this release do the job?” A machine-made release fails at the job long before any detector gets involved.

If you want to see which lines are actually carrying the machine signal before you touch a word, run the draft through the AI detector first. It gives you a sentence-by-sentence reading, so you’re editing three lines instead of rewriting a whole page.

The sentence-level method (and why whole-document rewriting backfires)

Most writers respond to a flag by paraphrasing the entire release. That’s the wrong move, and it usually makes things worse. A global rewrite flattens the few specific details that were doing the work, and it often produces the same uniform rhythm in new words.

Do this instead:

1. Score the draft as a whole. Get a document-level reading so you know whether you have a problem or a single noisy sentence.
2. Read the sentence-level breakdown. DeAIze highlights each sentence by AI probability, so you can see which lines carry the signal. Usually it’s a cluster, not the whole page.
3. Rewrite only the flagged sentences. Keep the ones that score clean. They’re already doing what you need.
4. Check that meaning survived. If a rewrite changes what the release claims, revert it. A press release that misstates a date or a figure is a liability, not a stylistic win.
5. Re-score and compare. Look at the before and after. If the number barely moved, the problem is structural, not lexical.

That last point matters. If every sentence is flagged, you don’t have a wording problem. You have a drafting problem: the release has no specifics in it. No amount of synonym-swapping fixes a release with nothing to say.

A worked example: before and after

Here’s an illustration, not a real client release.

Before (reads machine-made):

“We are thrilled to announce the launch of our innovative new platform, which represents a significant milestone in our journey to deliver cutting-edge solutions that empower businesses to unlock their full potential in an ever-evolving digital landscape.”

That sentence is doing a lot of motion and saying nothing. There’s no product name, no date, no user, no number, no verb that commits to anything. Every phrase is a placeholder.

After:

“We’re launching [Product] on 14 March. It lets small retailers sync stock across three sales channels from one screen. We built it because our own shop kept overselling.”

Same announcement. One sentence. The second version gives a journalist something to quote, a reason the product exists, and a detail they can verify. It also reads like a person wrote it, because a person did.

Notice what didn’t happen: no thesaurus, no “elevate”, no restructuring for the sake of variety. The fix was replacing abstraction with specifics. That’s the whole trick, and it’s the one thing a humanizer can’t do for you.

Where a draft editor helps, and where it can’t

The honest version: a draft editor is good at the mechanical layer and useless at the editorial layer.

Problem	What a draft editor can fix	What only you can fix
Uniform sentence length	Yes — varies rhythm across flagged lines	—
Predictable transitions	Yes — removes or rewrites them	—
No product name or date	No	You have to add the fact
Quote that sounds like a brochure	Partly — tightens phrasing	You have to get the real quote approved
Wrong claim in the release	No	You have to verify it
Missing contact block	No	You have to supply it

The pattern: a tool fixes texture, a writer fixes substance. If your release has real news in it and bad texture, a sentence-level editing pass will get you most of the way. If it has no news, no tool will rescue it.

This is also why we don’t promise a particular score on any detector. DeAIze’s own site reports an average AI-score reduction of 60%+ after humanizing (best case 71%), and on the example paragraph used on the site, a 92% score becomes 4%. Those are the product’s own predicted scores, not an official verdict from any third-party detector. A journalist’s gut check is not a score at all.

The disclosure question you’ll get asked

If you used a model to draft the release, someone will eventually ask whether you did. Answer before they ask.

Three practical rules:

• Disclose the tool, not the workflow. “Drafted with AI assistance and edited by our team” is usually enough. You don’t need to describe your prompt history.
• Never let a model invent a quote. An attributed quote that the spokesperson didn’t say is a fabricated source. That’s a different problem from a detector flag, and a much worse one. Our guide on AI writing and fake citations covers verification, and the principle is the same for quotes.
• Check your client’s or employer’s policy. Agencies increasingly have a written rule about AI use in client-facing copy. If you don’t know what it says, find out before you send, not after.

None of this is about hiding anything. A press release is a public document with your name on it. The goal is that it reads like your organisation wrote it, because your organisation is accountable for every claim in it.

A pre-send checklist for press releases

Run this before you hit send. It takes about ten minutes and catches most of what makes a release read machine-made.

1. Is there a date, a name and a number in the first two sentences? If not, the release is too abstract to pitch.
2. Does every paragraph start the same way? Vary the openings. “The company” three times in a row is a tell.
3. Could this release have been written about any company? If yes, it says nothing. Add the specific.
4. Is the quote something a human would actually say out loud? Read it aloud. If you stumble, rewrite it.
5. Have you scored it at sentence level? Fix flagged lines only, then re-score.
6. Does the contact block work? Test the email and the phone number. A dead contact on a live release is embarrassing.
7. Is the disclosure line accurate? Say what was used, and only what was used.

The first-party measurement is worth keeping in mind here. When we scored n = 1504 hand-written DeAIze guides on 2026-10-04, the mean AI score was 27% and the median was 20%, with 487 of 1504 flagged at a 30% threshold. Human writing scores high sometimes. That doesn’t mean it was machine-written, and it doesn’t mean you should rewrite a clean sentence to chase a number. It means the score is a starting point for editing, not a verdict on authorship.

What to do when two detectors disagree about your release

They will. Detector output is a probability estimate built on different signal weightings, so a release can read borderline on one tool and clean on another. If you’ve already checked two tools and got two answers, read are AI detectors accurate before you start rewriting again. The short version: neither number is proof, and the disagreement itself tells you the text is borderline rather than clearly machine-made.

What matters more than the score is whether a human editor would keep reading. That’s the test your release actually faces. A pitch that opens with a specific, dated, verifiable claim survives a skeptical editor even if a detector gives it a middling reading. A pitch full of “innovative solutions” fails regardless of what any tool says.

Ready to see which sentences read machine-made?

If you’ve got a release sitting in drafts and you’re not sure whether it sounds like you or like a template, the fastest way to find out is to look at it line by line.

• Run the release through the AI detector to get a document score and a per-sentence breakdown, so you can see exactly which lines carry the signal.
• Open the draft editor and work only on the flagged sentences. Rewriting runs a closed loop: it scores the text, rewrites the flagged lines, keeps a rewrite only when semantic similarity says the meaning survived, then re-scores and shows you before and after.
• If you’re editing a Word release, upload the .docx. Rewriting a Word document preserves fonts, styles and tables, so you’re not reformatting after every pass.
• Want to try it first? The free tier gives you 200 words on signup, no credit card. Register here.
• Doing this across a backlog of releases? Pricing is pay-as-you-go, credits never expire, and there’s no subscription.

Paste it in, look at the flagged lines, and decide for yourself whether the fix is a rewrite or a missing fact. Either way, you’ll know where to start.