How to humanize an AI-written press release
Put the news in the first sentence and rewrite the quote. Generated releases bury the story under positioning language and produce a quote nobody would print — an executive saying the announcement is exciting. Lead with what happened, and make the quote say something only that person could say.
Typical length · 300–500 words, of which a journalist reads the first forty before deciding. · Last reviewed 16 August 2026
Before and after
Before
TechCorp, a leading provider of innovative solutions, today announced the launch of its groundbreaking new platform. “We are thrilled to bring this exciting product to market,” said CEO Jane Doe. “This represents a significant milestone in our ongoing commitment to excellence.”
After
TechCorp has cut its warehouse picking software down to a phone app, and is giving it away to sites under fifty staff. “We built the enterprise version first and it was the wrong order,” said CEO Jane Doe. “The small sites are the ones running on paper, and they were never going to buy a six-month implementation.”
The positioning language went. The first line now says what happened and who it is for. The quote states a specific judgement about the company's own earlier mistake, which is the kind of sentence that gets printed.
The first sentence is the whole decision
A journalist on a news desk scans the opening line and moves on. That is not impatience; it is the only way to process the volume. So the question a release has to answer immediately is what happened, and generated releases almost never answer it — they open with who the company is and how it positions itself, which is the information a reader needs last.
The test is mechanical. Cover everything after your first sentence and ask whether someone could tell you what the announcement is. If the answer involves the words leading, innovative or solutions, the sentence is describing the company rather than the news.
This is also where the strongest single edit lives. Most releases can delete their first paragraph entirely and promote the second, and they get better immediately.
The quote is the part worth writing by hand
Executive quotes in generated releases are interchangeable. Thrilled, excited, significant milestone, committed to excellence — none of it is attributable to a person, and no journalist has ever printed a sentence about being thrilled.
A usable quote does one of three things: it explains a decision, it concedes something, or it says what happens next in concrete terms. All three require the speaker to have an opinion, which is why this part cannot be automated and should not be attempted at speed.
It also has to be true. A quote is a claim about what a named person said, and inventing one is a different category of problem from writing a flat sentence. Where a rewrite changes a quote's wording, the speaker approves the new version or it does not ship.
Facts are the failure that costs you
A release is a document journalists quote from directly, which means an error in it propagates into coverage you cannot correct. Generated drafts are fluent about numbers, and fluent is not accurate.
The specific risks are the ones a language model produces confidently: a funding figure, a customer count, a job title, a launch date, a percentage improvement. Every one of them needs checking against the source document rather than against the draft that came back.
Boilerplate deserves a mention because it looks like prose and is not. The paragraph about the company at the bottom is usually legally reviewed, and a rewrite that improves it may have changed a description someone signed off.
Formats with the same problem
Blog posts — where the same opening-paragraph waste costs you a reader rather than a journalist.
Product descriptions — the other format where an invented fact reaches the public as a claim.
Questions
- Do journalists check press releases for AI?
- Not in any systematic way anyone has documented, and a release is short enough that a detector score on one would be unreliable anyway. What journalists do is stop reading, which needs no software and happens faster.
- What should the first sentence contain?
- What happened, to whom, and why it is different from yesterday. If the sentence would be true a year ago it is not news, and if it describes the company rather than the event it belongs further down or nowhere.
- Can I use AI to draft the quote and get it approved after?
- You can draft a suggestion, and it must be genuinely approved rather than nodded through — the speaker is being made to say it. A quote invented and attributed without approval is a misrepresentation regardless of how it was produced.
- Is the boilerplate worth rewriting?
- Rarely, and check who owns it before you do. Company descriptions are often legally reviewed, and an improved sentence may no longer be the one that was approved.
What to watch for
- Never let a rewrite touch a quote without the speaker approving the new wording. An attributed sentence is a claim about what someone said.
- Check every figure, date, name and job title in the output against the source document.
- Embargo lines, boilerplate and contact details are structural. Rewriting them causes practical problems rather than stylistic ones.
If your writing gets flagged
Rewriting for rhythm and specificity tends to lower detection scores, because that is what detectors read as human. It is not a guarantee — detectors disagree with each other and change without notice, and we do not promise a result from any of them.