AI in the Newsroom

None of these tools existed when we first wrote this guide. Now generative AI and machine learning are part of the working data journalist’s kit. A newsroom that ignores them falls behind, and one that leans on them without care ends up publishing mistakes. The trick is knowing what they’re good for, where they fail, and the rules that keep them from embarrassing you.

Start with a distinction that matters more than any product name. “AI” collapses two different things: classical machine learning (classifiers, OCR, entity extraction) and generative models (the large language models behind chatbots). The prize-winning investigative uses so far have mostly been the former. When Missing in Chicago (the Invisible Institute and City Bureau) won the 2024 Pulitzer for Local Reporting, it used a purpose-built machine-learning tool and OCR to make thousands of police-misconduct records searchable, not a chatbot writing prose. Be able to say which kind of “AI” you mean. It shapes both your method and your readers’ trust.

What it’s good for

The pattern that holds up is AI as a needle-in-a-haystack tool for structured, investigative work, with a human checking every output:

• Document analysis at scale. OCR plus entity extraction turns an unsearchable pile of scanned PDFs into something you can query. This is what Google Pinpoint and DocumentCloud are built for.
• Transcription. Whisper-based tools transcribe interviews, hearings, and foreign-language audio fast enough to change what a small team can cover.
• Entity search across records. OCCRP’s Aleph and ICIJ’s Datashare let you find the same person or company across leaks and public records, with self-hosting for sensitive material.
• Structured extraction. An LLM constrained by a schema (a defined set of fields, like a standardized form) can pull structured data out of messy documents far more reliably than a free-text prompt.
• Coding assistance. LLMs can draft scrapers, regexes, and SQL, which saves real time as long as you read what they produce before you run it.

Where it goes wrong

The blockers are well documented, and they’re the reason none of the above is “set it and forget it.”

Hallucination and overconfidence. LLMs produce fluent, confident text that is sometimes simply wrong, and they rarely flag their own uncertainty. So treat any model output as an unverified tip, and check every claim against the source before it informs a word of your reporting.

Data security. Confidential or source-protected material often cannot go to a commercial cloud model at all. For high-risk document sets, prefer self-hosted tools (Datashare, local models) so the material never leaves your control.

Provenance and bias. Be aware of where a model’s training data came from, and audit AI-assisted outputs for the biases they can launder into your work.

The ground rules

You don’t need to invent an ethics framework. The major published guidelines (AP, Reuters, The Guardian, Poynter, and the international Paris Charter on AI and Journalism, which GIJN has signed) all land in the same place:

Put your own version in writing. A 2024 Associated Press report found that nearly 70 percent of newsroom staff were using generative AI, while fewer than half had any guidelines for it. A one-page policy, treated as a statement of your newsroom’s values first and a rulebook second, is one of the most useful things a small data team can produce. For templates, see the Resources page.

If your newsroom is unionized, the contract is a place to lock these rules in. Grist’s 2023 contract bars using AI to replace or reduce bargaining-unit work, and dozens of newsrooms have won similar protections. If you’re near the bargaining table, it belongs on the list.

The cautions here aren’t a reason to avoid these tools. Handled carefully, they change what a small newsroom can take on: a five-person shop can work through a document dump that used to take a foundation grant and a summer of interns.