Public relations

Can AI write comms copy that anyone remembers?

Not from scratch, says Sam Polstein: models have no emotion and a recognisable tone. Three mitigations he uses, and the fifty percent draft to never send.

No — not from scratch. Sam Polstein, who leads communications and content for deep tech startups at HAUS, says the models fail at exactly the thing comms is for: "My God, they have no emotion." Ask one for a client deliverable with none of your accumulated context and, in his estimate, it arrives around fifty percent of the way there. Presentable. Average. Average is the failure.

Why does AI-written copy fail to land?

Because nothing it produces unaided sticks, and sticking is the job.

"My God, they have no emotion. There's like nothing they create from scratch is emotionally resonant. It doesn't stick with people. And that's basically our job."

That is disqualifying rather than merely disappointing, in his framing.

"You could create that super average comms campaign, but if it's not resonating with anybody, it's just as bad as the campaign that didn't have any strategy behind it."

Galina Fendikevich offered a possible mechanism: emotional language often lacks enough statistical support for a model to select it. Her example came from her own messaging work — a user describing her workflow as "document chaos," a phrase that immediately became the positioning. Polstein's response: "I think that's an oxymoron for any LLM. Those two things always live in harmony."

The words that land are frequently the ones that should not go together. A system predicting likely next words is biased against precisely that.

Why does everything sound the same now?

Because each model has a recognisable voice, and readers are learning it.

"Everybody sounds the same now. Claude has a tone, it has a voice that I think we're all coming to recognize."

Fendikevich noted she can tell Claude from ChatGPT from Gemini by voice alone. Polstein's addition: "I think folks are picking up on it." For a profession whose product is distinctiveness, a house style you did not choose is a commercial problem, not an aesthetic one.

What actually helps?

Three things, all of which Polstein describes as helpful without ranking them, and none of which he claims solves it:

A good writing folder. He keeps a context folder of writing he rates plus a list of tropes to avoid, and has the model check output against it before returning anything. "It's editing its own work, essentially. And that's helpful — but it doesn't remove that Claude tone."
A defined client voice, built once. Worth the setup cost because it is reusable: "Once you've built it for a client, then you don't necessarily have to rebuild it."
Your outline, your headline. The highest-leverage of the three, in his description: "Rather than letting Claude build from scratch based on the context, maybe take it a step further and provide an outline of the thing you're trying to do. That's something that's quicker to create than if I was to write every single sentence, but allows me to heavily influence the voice and the tone, so that it's coming across in my writing rather than Claude's writing."

On headlines specifically: "If I give it a strong headline, I find that the output is a lot better, and it'll build more from the headline than it will from its own knowledge about how to write."

He is clear this is management rather than resolution: "This is a broad problem, and it's something we'll just continue to edit on."

Is the bigger risk slop or laziness?

Laziness with context, in Polstein's view — and he raises it in place of the slop conversation, not alongside it.

"There's a lot of talk in communications, in content marketing, about AI slop ... I think that we're not often talking about some of the other downsides of AI use. And I think being lazy with context is probably the worst one."

What that looks like in practice is the shortcut anybody can take on a busy afternoon.

"I could just jump into ChatGPT or Claude now and ask it to create a deliverable for one of my clients without any project knowledge, without any of the markdown files that we've created and curated over the years. And it would create something that's like, I don't know, fifty percent of the way there, just based on searching the web about my client. And that's tempting, but the output is mediocre."

"It becomes tempting to just put in the prompt, get the output, send it to the client, because you can send them something and it'll be halfway decent. But they don't pay us to be halfway decent."

The risk compounds because the relationship is already fragile. "Clients are questioning whether they need a PR agency, whether they can't just do it on their own. That's always a thread through comms." A halfway-decent deliverable is an argument for doing it without you.

What to do this week

Change the order you work in. Write the headline and the outline yourself, then hand those to the model with the client context attached — that single change is the one Polstein rates most highly for keeping your voice in the output rather than the model's.

Then build the two files that make the rest possible: a folder of writing you would be proud to sign, and a list of the phrasings you never want to see again. Have the model check itself against both before it hands anything back.

And set a rule about the fifty percent draft. It is a starting point, never a deliverable. The question to put to your own team: when a client reads your work, can they tell which model wrote it?