We Asked Four Frontier Models to Write a Press Release. Only One Lied.

We do this thing internally where, before a launch, we throw the brief at every frontier model we can get an API key for and see what comes back. Not because we're going to ship it — because watching them fail is the fastest way to learn what the story is actually missing.

So last week: one brief, four models, one prompt. "Write the announcement press release." Same context, same facts, same word count. Then we read all four like we were the journalist on the other end, the one with 300 unread pitches and a 4pm deadline.

Here's what happened.

Model one wrote the press release every press release is afraid of becoming. "Leading provider." "Robust, scalable solution." "Excited to announce." It was grammatically perfect and said absolutely nothing. If you fed it into a shredder, the shredder would also be bored. This is the default failure mode of AI comms: it has read ten million press releases and learned to write the average of all of them. The average press release does not get covered. The average press release gets archived.

Model two overcorrected into a LinkedIn thought-leadership post. Em-dashes everywhere. "Here's the thing." A rhetorical question every third sentence. It had opinions the company never expressed and a tone the founder would never use. Confident, punchy, completely off-voice. This is the other failure mode — personality with no grounding. It sounds like someone. Just not your someone.

Model three was genuinely good. Tight lede, real news up top, quote that sounded like a human said it. We'd have shipped 80% of it. Credit where due.

Model four made up a customer.

Not a vague claim — a specific one. A named enterprise logo, a percentage improvement, a quote from a VP of Engineering who, as far as we can tell, has never existed in this or any timeline. It was the most convincing draft of the four. It was also the one that, had it gone out, would have ended a client relationship and possibly involved a lawyer.

That's the whole thing about hallucinations: they don't show up looking like errors. They show up looking like your best work. The fabricated quote read better than the real ones, because the model wasn't constrained by what actually happened. Reality is a constraint. Constraints are annoying. Models, given the chance, route around them.

Which is why we named this blog what we named it.

The job isn't to stop the model from being imaginative — imagination is the useful part. The job is grounding: every wild idea tethered to something true. A press release can be bold, weird, funny, contrarian. It cannot contain a customer who doesn't exist. The creativity and the fact-check are not in tension; they're the same craft. You hallucinate the angle, then you ground every claim in it.

Here's the part the AI-will-replace-PR crowd keeps missing. Three of those four drafts were unusable, and the fourth was unusable in the most dangerous way possible — and you only know that if you already know the story, the client, the voice, and what's actually true. The model is a phenomenal first-draft engine. It is a catastrophic last-line-of-defense. The value didn't move from humans to machines. It moved from writing to judgment — knowing which of four confident drafts is the one that gets someone fired.

We still use all four models. We'll use more next quarter. But every draft runs through the same filter it always has: is this true, is this them, and would a journalist actually care. The tools got faster. The bar didn't move.

Anyway — model three got the byline. Model four is in time-out. And the fake VP of Engineering, wherever he is, remains available for comment.

— The WestComms team

(Want the version of your launch that's grounded enough to survive a fact-check and weird enough to get covered? That's literally the job. Let's talk.)