It’s not longer possible to manage AI in the newsroom with simple rule, according to a Reuters report. Here we look at the steps editors need to take.
Artificial intelligence (AI) is now part of everyday journalism. It helps with writing, editing, translating, and connecting to audiences. But as AI tools get more powerful, they get harder to control.
Reuters Institute researcher Ramaa Sharma spoke to 20 senior editors, technology leaders, and academics in 13 countries. She found that most newsrooms still rely on simple guidelines and a promise to keep a human in the loop. But, according to her research, for many that’s no longer enough to manage the real risks AI brings. Read Ramaa’s report on the Reuters Institute site here.
Media Helping Media (MHM) reviewed the findings of Ramaa’s report, and with the help of AI, extracted the ways newsrooms are said to be managing artificial intelligence in their workflows. We then wove that information into the following assessment of what the Reuters research revealed.
Three ways newsrooms are managing AI
- No formal rules: Journalists use tools such as ChatGPT on their own, without any newsroom policy. This is common in newsrooms with fewer resources. It can lead to mistakes, and to private information being shared with outside bodes.
- Written principles and committees: Some newsrooms have written down their AI principles. A group of people from different departments oversees AI use, and a human checks every AI-assisted story before publication. This is better, but the rules can go out of date quickly as AI tools change.
- A dedicated team: The largest organisations, including the BBC, Finland’s YLE, and Czech Television, have appointed a named person or team to lead on responsible AI. Their work combines legal knowledge, technical checks, and day-to-day editorial support.
Worth remembering
Whichever level your newsroom is at, someone needs to own AI decisions. Without clear ownership, AI use tends to grow in an unplanned way.
Checking AI’s work is getting harder, not easier
Almost every newsroom in the study agreed that a human should check AI-generated content before it is published. But this “human in the loop” idea is coming under strain.
AI does not save time in the way many hoped. Senior editors and copy editors, who are often already busy, must now check text line by line for AI mistakes. This adds work rather than removing it.
Constant close checking is tiring. When people get tired, they are more likely to miss a mistake, a made-up fact, or biased wording. India’s Scroll news site has responded by setting a daily limit on how much AI-generated text its staff are asked to check.
Two hidden risks worth understanding
Beyond tiredness, newsroom leaders pointed to two risks that are harder to see:
- Switching off your own judgement: When journalists get used to reading polished AI text, they can start to trust it without checking the original sources. Researchers call this “cognitive surrender” – simply accepting what the AI produced, rather than testing it.
- AI that agrees too much: Many AI chatbots are built to be helpful and agreeable. This means they may support a reporter’s idea, or a leading question, rather than challenging it the way a good human editor would. AI tool will often tell you what you want to hear.
Speed is not the most important factor
Many newsrooms judge AI success mainly by time saved – one study found 42% do this. But leaders in the Reuters Institute study warned that speed should not be the main goal.
Denmark’s JP/Politikens Group built a framework called the “Values Compass”. Before buying or using any new AI tool, editors must first agree on their purpose as a publisher and their journalistic principles. The technology decision comes after that, not before.
Fact-checking groups Chequeado, in Argentina, and Factchequeado, in the United States, follow a similar approach. Both built closed AI chatbots to answer questions from migrant communities – using AI to serve a clear public need, rather than using it simply because it exists.
New rules are coming
Governments are moving from voluntary guidance to firm law. Under Article 50 of the EU’s AI Act, anyone publishing AI-generated content on a matter of public interest in the EU must say so, unless a named human editor has properly reviewed it first. Newsrooms will need to be able to show, if asked, what that human review actually involved.
Three big questions every newsroom must answer
Who is responsible when AI gets something wrong?
The newsroom is responsible, not the company that made the AI tool. AI systems have no legal standing of their own, and software agreements usually rule out the AI company being blamed for mistakes or defamation. Publishers such as the BBC handle this by folding AI risk into their normal editorial policy process, so a risky decision goes through the same accountability chain as any other story.
How do you protect your brand and retain audience trust?
By keeping control of the whole process, rather than letting an outside AI system assemble your content unwatched. Ole Reissmann, Director of AI at Der Spiegel, has warned that breaking a story into pieces, feeding it to an AI system nobody controls, and then putting your name on the result, is a serious risk to audience trust.
Some newsrooms manage this with a technique called “closed retrieval”: instead of letting an AI search the whole internet, it is only allowed to draw on the newsroom’s own verified archive. Newsrooms are also being clear with readers about which graphics, translations, or audio were AI-assisted.
What does responsible AI buying look like?
It means asking hard questions of any AI vendor before you sign a contract, not just after.
- Data and jurisdiction: Check where the vendor stores your data, and which country’s privacy laws apply. Factchequeado chose a Spanish technology partner partly so that EU data protection rules would apply to its users.
- Training rights: Make sure the contract says the vendor cannot use your stories, leaked material, or confidential reporting to train their own AI models.
- Ongoing testing: Agree on regular, independent testing of the tool, to check it is not drifting towards inaccurate or biased output over time.
A checklist for newsroom managers
Use this as a starting point to check your own newsroom’s AI governance.
Foundation and accountability
- Write down which AI uses are allowed (for example, transcription or translation) and which are not (for example, unchecked AI-written copy).
- Name a person, committee, or policy lead who owns AI decisions.
- Make sure a named human editor carries final responsibility for AI-assisted content.
Workflow and oversight
- Set out what proper human review actually means in your newsroom – for example, checking original documents and source links.
- Set a daily limit on how much AI-assisted content one person is expected to check, to guard against tiredness.
- Train staff to spot when an AI tool is simply agreeing with them rather than challenging their thinking.
Brand and structure
- Where possible, limit AI tools to your own verified archive rather than the open internet.
- Tell your audience clearly when graphics, voiceovers, or translations are AI-assisted.
- Check that your own reporting and reader data are not being used to train outside AI models.
Buying and compliance
- Confirm your AI vendors follow data protection law (such as GDPR in the EU) and store data in a suitable location.
- Put a clause in every AI contract stopping the vendor from using your content to train their own models.
- Arrange regular checks – ideally every two weeks – on how accurate and unbiased the tool’s output remains.
Worth remembering
Start with the basics: agree what AI is and is not allowed to do, name who is responsible, and set a limit on how much AI work one person checks each day. Everything else can follow from there.
Getting started
If this feels like a lot, begin small. In the first month, focus on writing down your basic rules and agreeing who owns AI oversight. Over the following months, look at your AI contracts, and consider whether your newsroom’s own archive, rather than the open internet, should be the main source your AI tools draw on.
Source: This article is based on reporting by Ramaa Sharma for the Reuters Institute for the Study of Journalism, ‘From guidelines to architecture: how newsrooms are rethinking AI governance’, 11 August 2027. Media Helping Media (MHM) used Claude AI to extract the main learning points from Ramaa’s research. That material was then reworked into the article above. You can read the MHM AI policy here.
