Why Newsrooms Should Never Let AI Publish Without a Human Gate
My position is simple: no AI system should be able to publish a news story unless a named person has reviewed it and can stop it. I call that person and that checkpoint a human gate. AI can research, summarize, draft and suggest. It should not be the last step before the reader. The evidence for this is stronger than the usual "AI makes mistakes" argument, and the strongest objections to it deserve a fair hearing. This piece covers both.
What the evidence says about AI and news accuracy
The most direct evidence comes from a study coordinated by the European Broadcasting Union and led by the BBC. Journalists at 22 public service media organizations in 18 countries and 14 languages assessed answers from four AI assistants about news. They found that 45% of the answers had at least one significant issue, and 31% had serious sourcing problems such as missing, misleading or incorrect attributions. Tech Brew's summary of the study adds that 20% of responses had factual inaccuracies, and that the assistants struggled most with breaking or complex stories.
One caveat matters here. That study tested general-purpose chatbots answering news questions, not tools that newsrooms have built around verified sources. I will return to that, but the failure types it found are the ones that matter for publishing: invented details, weak sourcing and missing context.
Two failures, two lessons
Chicago Sun-Times, 2025. A summer supplement included a reading list of 15 books, of which 10 did not exist, although the authors were real. According to the paper's own review, the section was produced by a syndication company, was not submitted to Sun-Times journalists for review, and ran without a note saying who produced it. The freelance writer admitted he had failed to vet the AI output. King Features said it ended its relationship with him for not disclosing AI use. The Philadelphia Inquirer also ran material from the same supplement. The lesson: a gate must cover content from third parties, too. "Someone else made it" is not a review.
CNET, 2022-2023. CNET published 77 short finance explainers drafted with an in-house AI tool. According to its editor-in-chief, editors created outlines, then expanded and edited the AI drafts before publishing. After an error was found, a full audit identified incorrect information in 41 of the 77 articles, including problems such as incomplete company names and transposed numbers. CNET paused the program until it was confident that its tool and editorial processes would prevent both human and AI errors.
The CNET case is the most important one, because humans were in the loop and errors still got through. Futurism compared the editors to drivers on autopilot who stop actively controlling the vehicle. That points to the real lesson: a human gate only works if it is a real review, not a rubber stamp.
What readers think
Audiences are not asking for AI to publish on its own. The Reuters Institute's Generative AI and News Report 2025 surveyed six countries and found that 12% of people were comfortable with news made entirely by AI. That rose to 21% when there was some human oversight, and 62% were comfortable with entirely human-made news. The report also found that people have limited confidence that human oversight is routinely applied, which means saying you have a gate is not enough. You need to show it.
Standards are moving the same way. The Associated Press updated its AI newsroom standards in July 2026. The press release says AI-generated output is reviewed and edited by AP journalists before publication in every approved use, and that AI does not replace reporting, sourcing, editorial judgment or verification. The standards also set disclosure rules for cases where generative AI plays a material role in published content.
The best arguments against my position
"Humans make mistakes, too." True. But the failure modes differ. An AI can produce a fluent, confident and entirely invented book list attributed to real authors, at scale, in seconds. One study of reader perceptions found that readers rated AI-generated and human-written news as similar in quality when they did not know who wrote it. That is a reason for the gate, not against it: readers cannot reliably catch these errors for themselves.
"A human gate slows breaking news." It does add steps. But accuracy cannot be recovered after publication, and a correction never travels as far as the original error. Speed is a reason to design a faster gate, with an evidence file ready and a checklist, not to remove it.
"The EBU study tested chatbots, not newsroom tools." Fair. A system limited to verified sources may perform better, and I am not claiming otherwise. But I have not seen evidence that any such system is reliable enough to publish unsupervised, and the cost of being wrong falls on readers and on your reputation.
"Some content is routine." Some newsrooms have long generated formulaic items from structured data. I do not object to automation in principle. My claim is narrower: a named person must own the decision to publish, set the rules, audit the output and be able to pull the plug. For high-risk or novel content, that person should review each item.
What a real human gate looks like
A named owner. Someone is accountable for every published item, including syndicated and freelance material.
A checklist, not a glance. Check claims against sources, numbers, names, quotes, images and the headline, and confirm the original source for every key fact. An evidence file such as a fact pack makes this faster.
Time and authority. The reviewer needs enough time and the power to stop publication without penalty.
Higher scrutiny for higher stakes. Politics, health, crime and finance get slower, deeper review.
Disclosure. Tell readers when AI played a material role, in line with your corrections and AI policies.
A log and audits. Record who approved what, and review error rates and corrections regularly.
A rule for third-party content. Treat it as unreviewed until your gate has checked it.
Where NewsBolts stands
At NewsBolts, we build around this principle: AI helps with research, drafting, SEO and publishing tasks, and a person decides what is confirmed and what goes live. You can read more in our AI newsroom operating system guide.
FAQ
Should AI publish news without human review? No. In my view, a named person should review every story before publication and be able to stop it. Leading organizations such as AP require journalist review of AI output before publication.
What is a human gate in journalism? A human gate is a mandatory checkpoint where an accountable person reviews content, with a checklist, before it is published. It is more than a quick look, and it carries the authority to stop publication.
Is "human in the loop" enough? Only if the review is real. CNET had editors review AI drafts and still corrected 41 of 77 articles. The gate needs time, a checklist and accountability.
Can low-risk automated content skip review? Some formulaic content may be handled with rule-based checks and audits. Even then, a named person should own the rules, sample the output and be able to stop it.
Conclusion
AI is a useful assistant for newsrooms, and I expect its role to grow. But the person who decides to publish is the one readers will hold responsible, and that decision should never be delegated to a system that cannot be held accountable. Build the gate, give it teeth and show readers it exists.
Comments