How can journalists use AI without losing trust?

Journalism sells one thing above all: the reader's confidence that what they see is true and that someone stands behind it. Generative AI, software that writes text or makes images on request, can save reporters hours, but it can also produce confident errors and convincing fakes. The question for a newsroom is not whether to use it, but where, under whose control, and how openly.

The good news is that newsrooms do not have to invent the answer. Several large organizations have published their rules, and survey research shows fairly clearly what audiences will and will not accept.

What audiences say they accept

In mid-2025, Oxford's Reuters Institute, a journalism research centre, polled people in Argentina, Denmark, France, Japan, the United Kingdom and the United States. It found a wide "comfort gap":

The task matters too. People were most at ease with AI for back-office work, such as correcting grammar and spelling (55% comfortable) or translating (53%), and least at ease with AI-generated presenters or authors (19%). Just as telling: only a third believed journalists always or often check AI output before publishing. Among people who strongly distrust the news, that fell to under one in five. In other words, newsrooms start with a credibility deficit on this topic, and quietly doing the right thing is not enough; readers need to be shown.

What two newsrooms decided

The Associated Press published standards in August 2023. Its staff may experiment with tools like ChatGPT, but not to produce publishable content. Any AI output counts as "unvetted source material," which means it goes through the same judgment and sourcing rules as an anonymous tip. AP does not let AI add or remove anything in its photos, video or audio, and it tells staff to keep confidential information out of AI tools.

CBC/Radio-Canada, Canada's public broadcaster, published preliminary guidance in June 2023, promising that audiences would never have to wonder whether a CBC story, photo, audio or video was real. In practice, nothing is published without direct human oversight; AI-generated content is disclosed before people read, watch or listen; AI research is never the only source for a fact; identification tools such as facial recognition need advance permission from the standards office; and confidential sources are protected with familiar techniques like voice distortion and blurring, not with AI-generated faces or voices.

Both policies share three ideas: a human is accountable, AI output is raw material rather than a finished product, and the audience is told.

A worked policy and fact-check workflow

Here is a one-page policy a small or local newsroom could adapt, built on those three ideas.

1. Allowed without approval: transcribing interviews, translating source documents for the reporter's own understanding, spelling and grammar checks, suggesting headlines that an editor rewrites.
2. Allowed with editor approval and disclosure: AI summaries or audio versions of published stories, data analysis whose results are checked by hand.
3. Not allowed: publishing AI-written articles, altering photos or audio, creating images of real people or events, entering confidential or unpublished material into public AI tools.
4. Disclosure line, placed before the content: "This summary was produced with an AI tool and checked by [editor's name]."

And a fact-check routine for anything AI touched:

Here is my transcript of an interview and the draft story below it.
List every factual claim in the draft (names, titles, numbers, dates,
quotes). For each one, say whether it appears in the transcript, and
quote the exact transcript line. Mark anything not found as NOT IN
TRANSCRIPT. Do not correct or rewrite the draft.

The AI here does the tedious matching; the reporter then checks every "NOT IN TRANSCRIPT" item and every quote against the recording itself. The tool narrows the search, but it does not decide what is true.

Which steps must stay human?

Deciding what is news, and what is true, must stay with people. That covers choosing stories, interviewing sources, confirming facts with original documents or a second source, and judging whether material is authentic. AP's guidance is blunt on the last point: if a journalist has any doubt that a photo, video or recording is real, it should not be used, and checks such as finding the original source and running a reverse image search come first.

Final responsibility must also stay human. CBC's rule that nothing reaches the audience without a journalist vetting it, and AP's rule that AI output is unvetted material, both mean a named person signs off. Protecting sources belongs here too: deciding what to share with any outside tool, and how to hide an identity, is an ethical judgment.

Finally, the relationship with the audience stays human. Corrections, explanations of how a story was made, and replies to readers who ask whether AI was involved are where trust is rebuilt, and the Reuters Institute data suggest readers are watching closely.