Why use AI at all?

Because the parts of the job it's good at were never the journalism.

1. It gives the hours back. Transcribing a 90-minute interview, sorting 400 pages of records, pulling every date out of a court filing. Those are the tasks that eat a reporting day and burn people out of the business. A machine does them in minutes. You spend the time on interviews, verification and the story, and you go home at a reasonable hour.
2. It's a second set of eyes. Ask it to list every claim in your draft, find the sentence a reader will trip on, or spot the question you forgot to ask. It doesn't replace an editor. It catches what you'd catch yourself with more time.
3. It finds patterns in things too big to read. Thousands of comments, years of meeting minutes, a data set with no obvious shape. AI can tell you where to look. You still go look.
4. You can't cover it or catch it if you don't understand it. Synthetic images, AI-written press releases and chatbot-sourced claims are already in your inbox. Knowing how the tools work is now part of verifying anything.

The headline is not the job. The fifth headline is.

Reporters hear "AI can write your headlines" and think, I can write a headline. You can. Here's a real day on a story: you pitched it, reported it, shot it, wrote it, cut the video and filed the web version. Now the desk wants a main headline, an alternate for testing, a short one for the app, an SEO version with the keywords in front, and a social line that sounds like a person. Five versions of the same sentence about a story you are already sick of, due in 10 minutes, with the next assignment waiting.

That's where AI earns its place. Not the first headline, the one you were going to write anyway. The fourth and fifth, the ones that come out of a tired brain sounding like the first. Write the main headline yourself. Let the tool draft the alternates, then check every one of them before it publishes.

If you don't know what that day feels like yet, you will. This is what burnout is made of: not one hard task, a hundred small ones you've done before. Use the tool on the small ones. Keep the hard one for yourself.

What it is not for: writing your story, standing in for a source, or filling a gap in your reporting. All three are unethical.

Rules that hold up in any newsroom

The tools will be different next year. These won't.

1. You are the byline. AI can summarize, organize, transcribe and check. It should not write your story. Every fact, quote, number and name gets verified by a human before it publishes, and the human who verifies it owns it.
2. Disclose when it matters. If AI shaped what the reader sees or hears, say so in plain language. When in doubt, disclose. Specific disclosure language is available in the AI Premium Toolkit.
3. Never feed it what you would not print. Unpublished sources, embargoed documents, internal memos and anyone's personal data stay out of tools that may train on your inputs. Read the data policy before you paste.
4. Treat output like a tip, not a document. Models carry the gaps and biases of their training data, and they state guesses with the same confidence as facts. Confirm everything against a primary source you can name.
5. Keep the receipts. Save your prompts, the outputs and each version you edited. If a story is challenged, you can show exactly what the tool did and what you did.
6. Real events, real images. Do not generate or alter images of real people or real news events and present them as photography. Before you create anything synthetic, ask why you're making it and whether it belongs in the story. If it does, label it clearly every time it appears, including in social clips.

If you don't trust the cloud, don't use it

Every chatbot you've heard of runs on someone else's computer. You paste, it uploads.

For most work that's fine. For a source who could lose a job, a visa or worse, it isn't. You have three options.

1. Turn off training. Every major tool has a switch. It stops your chats from training the model. It does not stop the company from storing them or handing them to a court.
2. Run the model on your own machine. Free apps let you download an open model to your laptop and run it with Wi-Fi off. Nothing leaves the room. These models are smaller and slower than the big chatbots, but they summarize, translate and transcribe well enough for sensitive work. Your phone already does some of this. Most phones now transcribe voice recordings on the device, no upload needed, as long as you turn off cloud backup for the recorder app.
3. Build your own, like the big newsrooms. The New York Times doesn't type into a chatbot. Its staff use tools the Times built, such as Cheatsheet for sorting records and Echo for searching its archive. The tools still use outside AI models, but under contracts that keep Times data out of training and inside the paper's control.

The rule: the more a source has to lose, the closer the AI should be to your own hands.

Questions to ask before you open the tool

Thirty seconds, every time. If any answer makes you pause, that's the answer.

1. Is this tool cleared for my newsroom? If your outlet has a policy, follow it. If it doesn't, check with your editor.
2. Am I about to share something I shouldn't? Names of sources, unpublished documents, anything under embargo.
3. Can I verify the result myself? If you can't check it against a source you can name, don't use it.
4. Does the reader need to know AI was involved? If it substantially changed what they read, see or hear, yes. Spellcheck-level fixes don't need a label. A generated image or an AI-written summary does. What counts as substantial is your newsroom's call, and it should be written down before the story, not argued about after.
5. Would I defend this choice on the record? To an editor, a source or the subject of the story.
6. Did I turn off model training and check the security settings? If you haven't, do it before you paste anything. Here's how.

Before you publish

Tap each line as you clear it. Nothing is saved, so it resets when you leave the page.

• Every fact has a primary source I verified myself, whether AI touched it or not.
• Every quote came from a real interview or document I can produce.
• No confidential material or personal data went into a tool.
• Any AI-generated or AI-altered visual is labeled where the reader sees it.
• The disclosure line says what the AI did, in language my mother would understand.
• I saved the prompts and outputs with the story file.
• I would be comfortable explaining my AI use to the subject of the story.

Prompts that respect the reader

Copy, paste, adjust. Each one is built so the tool cannot quietly add facts.

Fact-check my draft

Why it works: it turns the model into a claims extractor, not an editor. You do the verifying.

Interview prep

Why it works: you do the thinking first. The tool fills gaps and catches assumptions, not the other way around.

Readability check

Why it works: the tool points, you fix. The words stay yours.

Free resources worth your time

• Olivia's ethics and standards page. A working example of an AI disclosure policy for an independent journalist. Copy the structure, make it yours.
• Online News Association. The AI in Journalism Initiative: practical training on prompting, transparency, copyright and verification.
• Poynter. A step-by-step guide to writing an AI ethics policy for your newsroom.
• Trusting News. Ready-made language for explaining your AI use to readers, and research on what audiences actually want disclosed.

The AI Premium Toolkit

The one thing on this page that costs money, and why.

Bring this into your company, newsroom or classroom

Olivia runs hands-on AI training for reporters, editors, producers and comms teams through Media Consulting Coaches. Every session is built around your workflows and your standards. Clients include USC and the United Nations.

The rules on this page are written for newsrooms, but the method works for any organization. Olivia adapts every session to the industry, the tools already in use and the standards the team has to meet.

About Olivia Smith

Emmy Award-winning journalist, AI specialist and media trainer. Olivia has worked for numerous outlets, including a decade at Disney and ABC News as a reporter/producer for the network and Good Morning America, and as ABC7's first Executive Producer of Next Generation Storytelling. She then worked on the Generative AI Content team at Meta. Olivia teaches at USC Annenberg, where she created the school's first journalism and AI course and has developed others since, and co-hosts the Prompt Response podcast. She has contributed to several organizations, including CNN, 60 Minutes, Al Jazeera and the Columbia Journalism Review.

More at oliviasmithonline.com. Instagram: @livnews.