Risks of Using AI to Write Press Releases: A Review Checklist

The main risks of using AI to write press releases are invented facts, invented or reworded quotes, inflated claims, tone that sounds like every other release, and legal exposure when any of those end up in a public, dated document with your company's name on it. None of these risks mean you shouldn't use AI for a first draft. They mean nothing AI-drafted goes out without a structured human review, and the review has to be designed for the specific ways models fail.

I use AI in drafting. I also read every line of every release before a journalist sees it, because a press release is the one document reporters assume you've checked.

Why is a press release riskier to draft with AI than a blog post?

A blog post can be edited quietly. A press release can't.

Once a release goes out on a wire or lands in a reporter's inbox, it gets copied. Aggregators republish it word for word. Reporters lift lines into their stories. In crypto especially, a release can be on dozens of sites within hours. If one sentence is wrong, you aren't correcting one page. You're chasing copies.

A release is also a statement of record. It names investors, partners, numbers and people. It often carries a quote attributed to a named executive. And for companies with a token, it can be read as a statement about something people trade, which makes accuracy a legal matter, not just a reputational one.

So the question isn't whether AI writes decent prose. It does. The question is whether you can catch what it gets wrong before the copies multiply.

What are the specific ways AI gets a press release wrong?

These are the failure modes I check for in any AI-drafted release. Each one looks plausible on a first read, which is exactly why it's dangerous.

The invented quote deserves its own warning. A model asked to "add a CEO quote" will write one. If that quote goes out and the CEO later says "I never said that", you've created a credibility problem with every reporter who used it. Quotes in a release should start from something the person actually said, in a call or a message, then get tightened and approved by them.

Why does generic AI tone hurt a release even when it's accurate?

Because reporters read hundreds of releases and the AI rhythm is now instantly recognisable. A correct release that sounds generated gets the same treatment as a wrong one: it gets skimmed and deleted.

I've posted about this confusion more broadly:

Most crypto founders confuse press releases with PR. That’s why they get ignored. Shilika Jain on X, 6 May 2026

AI makes that confusion cheaper to act on. If writing a release takes ten minutes, founders send more of them, about less. The release becomes the whole strategy. I wrote about why that fails in the wire services trap, and I've said the same thing on X more bluntly:

that’s because your running your press coverages on wire services rather than building up your credibility from scratch! Shilika Jain on X, 9 Aug 2024

An AI-written release on a wire service is the fastest possible version of a tactic that already didn't work.

What fixes tone is not a better prompt. It's specifics only you have: the real number, the customer's actual words, the reason this happened now. Models can't supply those, so a release built around them reads as human because it is.

The Six-Pass Release Review

This is the checklist I run on any AI-assisted release. Each pass looks for one kind of error, which is faster and more reliable than reading the whole thing six times hoping to catch everything.

1. Source pass. Highlight every fact: names, numbers, dates, amounts, titles, product names. Next to each, write where it came from (term sheet, product doc, partner email). Anything with no source gets deleted or confirmed. This pass catches invented facts.
2. Quote pass. For each quote, confirm the named person said or approved it, in writing. Partner quotes need the partner's comms team to approve them, not just your contact there. This pass catches invented quotes.
3. Superlative pass. Search for "first", "only", "leading", "largest", "fastest", "revolutionary", "best". Keep a word only if you can name the comparison and defend it to a sceptical reporter. This pass catches claim inflation.
4. Freshness pass. Re-check any market data, competitor mention or regulatory reference against today's sources, not the draft. Models confidently repeat things that were true a year ago. This pass catches stale context.
5. Read-aloud pass. Read the release aloud. Cut every sentence that could appear in any company's release. Replace at least two with something specific to this one. This pass catches tone drift.
6. Exposure pass. Send the final text to whoever owns legal and compliance, with forward-looking statements, token references and named partners flagged. This pass catches legal exposure.

Only after pass six does the release go to journalists or a distribution service. If any pass changes a fact, run the source pass again on that paragraph.

What does this look like on a real draft?

A worked example helps. This one uses an invented company, so nothing here describes a real client. A Web3 infrastructure startup asks a model to draft a funding release from a few bullet points: amount raised, three investors, a product launching "soon".

Say the draft comes back clean and confident. Running the passes would look like this:

• Source pass: the draft says the round was "led by" the first investor on the list. The term sheet shows no lead, so it becomes "with participation from".
• Quote pass: the CEO quote calls the company "the backbone of the next generation of Web3". The CEO never said it. Replace it with a line from a recorded call where the CEO explains why customers kept asking for the product.
• Superlative pass: "the first decentralized solution for" comes out, because two competitors exist.
• Freshness pass: a line about a regulation "coming into force next year" is out of date. Remove it rather than rewrite it, because the release doesn't need it.
• Read-aloud pass: the opening paragraph is interchangeable with any funding release. Rewrite it to lead with what the money is for.
• Exposure pass: "soon" becomes no date at all, and legal strikes a phrase that implied token utility.

None of these errors is dramatic. Every one of them would be embarrassing in a reporter's inbox.

Who should own the final review?

One named person, not "the team". AI-assisted drafting tends to spread responsibility thin: the founder thinks marketing checked it, marketing thinks the founder approved it, and the model is nobody's responsibility.

I'd split it like this:

This is where honesty costs something in the short term. Telling a founder that their favourite line has to go, because there's no source for it, is part of the job. A model will never refuse to write the line. That refusal has to come from a person.

If you'd rather have that person be me, release drafting and review sits inside my content writing work.

When is AI drafting fine and when should you skip it?

AI is fine for structure, first drafts of boilerplate, cleaning up a founder's rambling notes, producing a short version for social, and translating an approved English release as a starting point for a native-speaking reviewer.

I'd skip it, or keep it to formatting only, for:

• Crisis statements, where every word is weighed and context changes by the hour.
• Anything involving a token, pricing or financial results, where wording carries legal weight. My guide to writing a crypto press release covers what that wording should look like.
• Partner announcements, where the other side's comms team has its own approval process.
• Releases for markets where you don't have a native-speaking reviewer.

And sometimes the right answer is no release at all. If the news is small, a targeted pitch or a founder post does more. My press release template covers the structure when you do need one.

FAQ

Can I just tell the AI not to make things up?

You can, and it helps a little, but it doesn't remove the risk. Models still fill gaps with plausible details, especially when the prompt implies something should be there, like a quote or a lead investor. Treat the instruction as a convenience and the source pass as the actual control.

Do journalists penalise releases that were drafted with AI?

They don't usually know or care how a draft started. They penalise releases that are generic, inflated or wrong, and AI drafts are more likely to be all three unless someone edits them hard. A release that's specific and accurate won't be discounted because a model helped with the first version.

Is it safe to paste unannounced news into an AI tool?

Check your company's policy and the tool's data terms first. Unannounced funding, partnerships or incidents are confidential, and some tools and plans handle submitted data differently. If in doubt, draft the structure with placeholder names and add the real details in your own document.

How long should the six-pass review take?

For a standard announcement, an experienced reviewer can do it in under an hour, plus waiting time for quote approvals and legal. Build that waiting time into your timeline. The usual reason releases go out unchecked is that the review started the night before distribution.

Have an AI-drafted release you're unsure about? Book 30 minutes and I'll review it with you before it goes out.