How to Become an AI Communications Specialist From General PR

To become an AI communications specialist, you add a domain layer to the PR skills you already have: enough understanding of how models are built and served, how benchmarks work, what the safety vocabulary means and which rules apply, that you can explain a client's claim to a reporter without the founder in the room. Pitching, writing and media relations transfer straight across. What doesn't transfer is fluency, and fluency is what AI reporters test in the first five minutes of a call.

I had to build a domain layer like this once already, when I started in Web3 and had to learn the tech before I could pitch it. Now I'm watching good generalists try the same move into AI. The ones who struggle aren't weak at PR. They just can't tell when a founder's claim is shaky.

What does an AI communications specialist actually do differently?

A generalist's job is to find the story and get it in front of the right reporter. An AI comms specialist does that too, plus three things a generalist usually hands back to the founder:

• Translates the technical claim. "We fine-tuned a 7B open-weight model on proprietary legal data" has to become a sentence a business editor can use, without becoming false.
• Stress-tests the claim before a journalist does. If the release says "state of the art", you need to know against which benchmark, on what date, and whether anyone has beaten it since.
• Anticipates the safety and policy question. Almost every AI story now carries a second question about misuse, data, jobs or regulation. You prepare the answer before the interview, not during it.

That's the gap. The rest of this post is the reading list and the self-test I'd give anyone making the move. If you're still deciding whether comms is a good bet at all with AI in the picture, my sibling piece on PR skills AI can't replace covers that question.

Why does narrow domain fit matter so much?

In Web3, I learned that "crypto experience" on a CV tells you very little. I posted about this when I saw teams hiring marketers on the wrong signal:

DeFi is a niche but there are various niches in this broader niche! If you don’t hire the close marketing fit, you just lose the opportunity to acquire potential users, partnerships & even brand credibility! Shilika Jain on X, 20 Sep 2023

AI is the same, only bigger. "AI company" covers foundation model labs, inference and compute infrastructure, developer tools, AI agents, vertical applications in healthcare or law, safety and evaluation startups, and decentralized AI networks. Each one has its own reporters, its own vocabulary and its own way of getting a story wrong.

So don't try to become "an AI comms person" in the abstract. Pick one or two sub-sectors and go deep. When I worked with GAIA, a decentralized AI project, the story only made sense to reporters once it was framed around AI agents and a concrete proof point, node count, rather than generic "AI meets blockchain". A steady run of founder interviews around that one narrative is what built toward Forbes and Decrypt coverage. Without the domain context, it's hard to even decide which of the two audiences, crypto or AI, a pitch like that should lead with. You can read the full engagement in the GAIA case study.

The AI Literacy Stack: five layers to learn in order

This is the frame I use to check whether someone is ready to sit in on a technical briefing. Learn the layers in order, because each one depends on the one below it.

You pass a layer when you can answer the self-test for a real client without opening a browser. Most people with a PR background can get through layers 1 and 2 in a couple of weeks of focused reading. Layers 3 to 5 take longer because they change.

Layer 1 in practice: start with the AI supply chain

Start with the supply chain. Almost every AI company sits at one point in it, and knowing where tells you which reporters care.

1. Compute. Chips, data centres, GPU clouds. Reporters here care about capacity, cost and power.
2. Foundation models. The labs that train large general models. Coverage is about capability, safety and money.
3. Infrastructure and tooling. Inference platforms, vector databases, evaluation tools, orchestration. Developers and trade press care.
4. Applications. Products built on top of models. Business and vertical press care about customers and outcomes.
5. Agents. Software that takes actions, not just answers. Reporters care about autonomy, reliability and who's liable when it goes wrong.

Then learn the words founders use loosely. "Our model" might mean a model they trained, a fine-tuned version of an open-weight model, or a prompt on top of someone else's API. Each is legitimate. Each needs different wording. If a release says "proprietary model" and it's a wrapper on a third-party API, a technical reporter will find out, and the correction will follow the company for a long time.

How should you handle benchmarks and performance claims?

Benchmarks are where AI coverage goes wrong most often, so this is where a specialist earns their fee. A benchmark is a fixed test set with a scoring rule. A score tells you how a model did on that test, under that setup, on that day. It doesn't tell you how the product behaves for a customer.

My working checklist before any performance number goes into a pitch:

• Name the benchmark and version, not just "industry benchmarks".
• Say what it was compared against and when the comparison was run.
• Ask whether the test data could have leaked into training (contamination). Reporters ask this.
• Prefer a customer outcome over a leaderboard rank when you have one.
• Keep the raw evaluation details ready to share with a technical reporter on request.

I go much deeper on this in how to pitch AI benchmark claims to journalists. The short version: a specialist's value is saying "we can't use that number yet" before the founder's favourite chart goes out.

What safety vocabulary do AI reporters expect you to know?

Safety questions are no longer reserved for the big labs. A small startup announcing an agent that books travel will get asked what happens when it books the wrong flight. Learn these terms well enough to use them precisely:

• Hallucination: a model producing confident output that isn't true.
• Red-teaming: deliberately trying to make a system misbehave before release.
• Guardrails: the filters and rules around a model that limit what it will do.
• Alignment: the broader effort to make model behaviour match intended goals.
• Model card or system card: the document describing what a model is, how it was tested and its known limits.
• Human in the loop: a person approves or reviews an action before it takes effect.

Then make yourself a translator. One of my older posts applies directly here:

If you stop using technical jargons, half of your user onboarding problem is solved! Shilika Jain on X, 31 Oct 2023

The same holds for press. You need the jargon to understand the briefing and to be trusted by a technical reporter. You need to drop it for the business editor. A specialist runs both registers in the same day. I've written about how this plays out for safety-led stories in responsible AI PR.

Which AI regulation should a comms specialist follow?

You don't need to be a lawyer. You need to know which rule a reporter will connect your client's news to, so the answer is ready.

Track four areas:

• The EU AI Act. Its risk-based approach sorts uses into categories with different obligations, and it applies to companies selling into the EU wherever they're based. Know roughly where your client's product would sit.
• Data protection. GDPR in Europe and its equivalents elsewhere shape every question about training data and user data.
• Copyright and training data. Lawsuits between publishers, artists and AI companies are an active story, and any client training on scraped content will be asked about it.
• Sector rules. Healthcare, finance, hiring and legal AI each sit under existing regulators who are already writing guidance.

In the US, the picture is a moving mix of state laws, agency guidance and executive action, so I check it fresh before any launch rather than relying on what I knew last quarter. When a client's announcement touches a regulated use, the legal team signs off on the wording. Your job is to flag it early, not to rule on it.

What does a 60-day transition plan look like?

The plan I'd give a generalist who wants to cover AI clients within two months:

The beat map matters more than people expect. Reading what a reporter has already written tells you which claims they distrust and which questions they always ask. That's the cheapest domain education there is.

When doesn't specialisation make sense?

If your clients are mostly consumer brands that happen to add an AI feature, deep specialisation may cost more time than it pays back. The story there is the product and the customer, and a strong generalist covers it well.

Specialising also narrows your pipeline. AI funding and hiring move in cycles like everything else. I'd keep your general PR craft sharp and treat the AI stack as an added layer, not a replacement. And if you can't get a technical founder to spend an hour a week answering your questions, you'll struggle whatever you read, because the best education is the client's own engineers.

If you work at an AI company and want this layer covered without building it in-house, that's the work I do in AI startup PR.

FAQ

Do I need a technical degree to work in AI communications?

No. You need enough fluency to understand a briefing, ask the right follow-up and spot a claim that won't survive a technical reporter. Plenty of strong AI comms people came from journalism, policy or general PR. What you can't skip is the reading and the habit of asking engineers "how do you know that".

Should I learn to use AI tools as well as understand them?

Yes, and for a different reason than speed. Using models daily shows you where they fail, which makes you better at writing honest claims about them. It also helps you understand what a client's users actually experience, which is often the real story.

How do I show AI domain knowledge on a portfolio if I haven't had AI clients yet?

Write. Rewrite three real AI announcements into clearer, more defensible versions and explain your edits. Publish a short explainer on one benchmark or one regulation in plain language. A hiring manager learns more from that than from a list of tools you've tried.

Is AI comms in Web3 different from AI comms in traditional tech?

Partly. Decentralized AI projects carry two audiences at once: crypto reporters who care about tokens and networks, and AI reporters who care about models and agents. You have to decide which story leads for each outlet. The domain stack above still applies, with on-chain mechanics added on top.

Moving into AI clients and want a second opinion on your first pitch? Book a 30-minute call with me.