The old question was whether artificial intelligence could write like a journalist. The new question is whether readers would know if it already had.
Across the news business, artificial intelligence has moved from experiment to infrastructure. It is used to transcribe interviews, suggest headlines, summarize stories, translate copy, generate alt text, search archives, assemble newsletters and, in some cases, draft or produce entire articles. What remains harder to know is where assistance ends and authorship begins.
The best evidence does not support the most dramatic version of the claim: that the front pages of elite publications are now largely being written by machines. But it also no longer supports the comforting opposite: that AI is merely a distant, back-office tool with little bearing on what readers see.
Reuters says its journalists may use generative AI in “reporting, writing, editing, production and publishing,” while adding that it discloses when it relies primarily or solely on generative AI to produce news content. The Associated Press, which has long used automation for corporate earnings stories, updated its standards in 2024 to allow experiments with AI-assisted Spanish translations, summaries and headlines, while saying each begins with AP journalism and is edited or vetted by AP journalists before publication. The Financial Times says its journalism will continue to be reported, written and created by journalists and editors, even as it experiments with AI for story discovery, accuracy, quality and secondary content.
The point is not that every major newsroom is secretly replacing reporters. It is that the category “written by a human” is becoming less clean than it once was.
Automated journalism is not new. The AP said in 2015 that it was automatically generating more than 3,000 U.S. corporate earnings stories each quarter, a tenfold increase over what reporters and editors had previously produced. The Washington Post used its Heliograf system in 2016 to cover nearly 500 election races, with editors able to add reporting, analysis and color to bot-written text.
Those early examples were mostly structured-data stories: earnings, sports results, election returns. The newer generation of AI can do more than fill templates. It can draft fluent paragraphs, rewrite in house style, summarize documents, imitate conventional news prose and produce endless variations of service journalism. That has changed the economics of publishing.
In a 2026 survey of media leaders, the Reuters Institute found that back-end automation, including transcription, copy-editing assistance and automated metadata, remained the most widely mentioned AI use case. It also reported that headline generators, automated alt text, integrated style guides and automated summaries had been widely adopted. Yet the same survey found some cooling of expectations: only 13 percent of respondents described current AI initiatives as transformational, while 42 percent called them limited.
The evidence suggests an uneven future. AI is most likely to spread first through the content that is easiest to commoditize: weather updates, earnings reports, sports recaps, event listings, traffic stories, product explainers, local service articles, SEO-driven guides and rewritten summaries of information that already exists elsewhere.
That is already happening outside the most prestigious newsrooms. One large-scale audit of 186,000 articles from 1,500 American newspapers, published as a 2025 preprint, estimated that about 9 percent of newly published articles were partly or fully AI-generated. The study found heavier use at smaller local outlets, in topics such as weather and technology, and in certain ownership groups. It also found disclosure was rare: in a manual review of 100 AI-flagged articles, only five disclosed AI use. Because the study relied on an AI detector, its article-by-article classifications should not be treated as proof. But as a broad signal, it is evidence that undisclosed AI-generated or AI-assisted publishing is no longer hypothetical.
The problem is compounded by the fact that readers are not especially good at telling the difference. A cross-national study from CISPA and partner universities, using cleaned data from 2,609 participants in the United States, Germany and China, found that most participants classified AI-generated media as human-made. The text samples in that study were news items.
That makes disclosure more important and more complicated. If a reporter uses AI to transcribe an interview, few readers would expect a label. If an editor asks AI for ten headline options, a disclosure may feel excessive. If an article is substantially drafted by AI and then edited by a person, many readers would probably want to know. But newsroom policies often reserve disclosure for substantial, primary or visible AI use, leaving a gray zone where AI may shape the final product without being obvious.
Public opinion is wary. In a 2025 Reuters Institute report across six countries, only 12 percent of respondents said they were comfortable with news made entirely by AI. Comfort rose to 43 percent when a human led with some AI assistance, and to 62 percent for entirely human-made news. People were more accepting of back-end uses like grammar editing and translation than front-facing uses such as artificial presenters or authors.
There is also a quality paradox. AI writing is often better than people expect. In a large study published in Scientific Reports, ChatGPT-generated argumentative essays were rated higher than human-written student essays; GPT-4 scored about one point higher on a seven-point scale. Other research has found that readers may rate AI-generated or AI-assisted short news articles similarly to human-written ones on measures such as readability, credibility and expertise.
But polished prose is not the same as journalism.
A well-written paragraph can still be wrong. A balanced-sounding summary can omit the most important fact. An article can be clear, structured and confident while relying on weak evidence. The strongest work at major publications is not valuable merely because the sentences are smooth. It is valuable because someone made calls, checked records, interviewed people, weighed uncertainty, understood context, accepted accountability and knew what could not yet be said.
That is where the case for human journalism remains strongest. AI can imitate the visible surface of news writing more easily than it can reproduce reporting. It does not cultivate sources, attend public meetings, notice nervous pauses in interviews, challenge evasive officials, file records requests, or bear reputational responsibility for what appears under a publication’s name.
Still, the incentives pushing publishers toward more AI-generated content are real. Generative AI lowers the cost of producing text at scale. NewsGuard reported in June 2026 that it had identified 3,749 AI content-farm news and information sites across 16 languages, many publishing dozens of articles a day with little or no human oversight and no clear disclosure. These are not the same as reputable news organizations, but they show how easy it has become to flood the information system with news-like material.
The cautionary tale is CNET. In 2023, the technology site came under scrutiny for publishing AI-written personal finance articles. The Verge reported that CNET issued corrections on 41 of 77 AI-written stories after an internal review found errors. The episode did not prove that AI cannot be useful in journalism. It did show that fluent automated writing can create old-fashioned editorial problems at new speed.
So where is this headed?
The most likely future is not a clean replacement of journalists by machines. It is a layered system in which AI writes more first drafts, more summaries, more sidebars, more explainers, more local briefs and more search-friendly service copy. Human editors will review some of it carefully, some lightly and some not enough. Elite publications will probably continue to emphasize human reporting, partly because that is their brand and partly because generic information is becoming less defensible as a business.
For readers, the question will become less “Was AI used?” and more “How was AI used, who checked it, and what evidence supports it?”
A fully human article can be sloppy. An AI-assisted article can be accurate. A machine-generated summary can be useful. A beautifully written story can be misleading. Authorship matters, but it is not the only thing that matters.
The better test is provenance: Is there original reporting? Are the sources named? Are claims linked to documents, data or direct observation? Does the article distinguish fact from inference? Does the publication explain when AI played a substantial role? Is someone accountable when errors appear?
AI has already entered the newsroom. The evidence says it will spread further. But the more serious question is not whether machines can write. They can. It is whether the institutions publishing that writing will be transparent enough and careful enough for readers to keep trusting what they read.
Evidence & Source Transparency
Evidence First shows its work. The article ends above; this section is included so readers can inspect the main sources behind the factual claims.
The list below does not source every sentence. It focuses on the factual claims most important to the argument.
1. Major news organizations’ stated AI policies
Claim or topic:
Major news organizations are using AI in newsroom workflows, but often describe it as assistance rather than a replacement for journalists.
Source:
Reuters: Reuters and AI
Associated Press: Updates to generative AI standards
Financial Times: AI standards
Source type:
Primary document.
What it supports:
These sources show how major news organizations publicly describe their own AI use, including AI in reporting, writing, editing, summaries, translations, headlines, and secondary content.
Important caveat:
Policies describe stated standards. They do not independently prove how every newsroom employee or contributor uses AI in practice.
2. Earlier forms of automated journalism
Claim or topic:
Automated journalism existed before the current generative AI boom, especially for structured-data stories such as earnings, elections, and sports.
Source:
Associated Press: Automated earnings stories multiply
The Washington Post: Heliograf election coverage
Source type:
Primary document.
What it supports:
These sources support the article’s point that automated article production did not begin with ChatGPT. AP described generating thousands of earnings stories per quarter, and The Washington Post described using its Heliograf system for election coverage.
Important caveat:
These examples involve structured or semi-structured information. They are not the same as AI independently producing investigative or enterprise journalism.
3. Newsroom AI adoption trends
Claim or topic:
AI adoption is increasing across newsroom workflows, especially in back-end automation, headlines, summaries, metadata, transcription, and related production tasks.
Source:
Reuters Institute: Journalism, media, and technology trends and predictions 2026
Source type:
Analysis.
What it supports:
The report supports the article’s claim that AI is increasingly embedded in newsroom infrastructure, while also noting that many publishers still describe current AI initiatives as limited rather than fully transformational.
Important caveat:
This is based on a survey of media leaders and industry analysis. It reflects reported expectations and practices, not a direct audit of every newsroom.
4. Estimates of AI-generated newspaper articles
Claim or topic:
Some AI-generated or AI-assisted newspaper articles appear to be published with limited disclosure, especially in parts of the local newspaper ecosystem.
Source:
arXiv: AI-generated content in U.S. newspapers
Source type:
Academic research / estimate.
What it supports:
The study estimated that about 9 percent of a large sample of newly published U.S. newspaper articles were partly or fully AI-generated, with heavier use in smaller local outlets and some topic areas.
Important caveat:
The study relies on AI-detection methods. AI detectors can make mistakes, so the estimate is useful directional evidence, not definitive proof for each individual article.
5. Whether readers can identify AI-generated content
Claim or topic:
Readers are not reliably able to tell whether media, including news-like text, was generated by AI.
Source:
CISPA: AI-generated media detection study
Source type:
Academic research.
What it supports:
The study supports the article’s claim that people often misclassify AI-generated media as human-made, including AI-generated news items.
Important caveat:
Detection performance depends on the type of content, the quality of the AI output, the audience, and the study design. It does not mean nobody can ever detect AI writing.
6. Public comfort with AI-generated news
Claim or topic:
The public is much less comfortable with fully AI-generated news than with human-led journalism that uses some AI assistance.
Source:
Reuters Institute: Generative AI and news report 2025
Source type:
Analysis / survey research.
What it supports:
The report supports the article’s claim that public trust and comfort vary sharply depending on whether AI is used entirely, partially, or only as a back-end tool.
Important caveat:
Survey responses measure stated comfort and perception. They do not necessarily predict how readers behave when reading unlabeled AI-assisted content.
7. Quality of AI-generated writing compared with human writing
Claim or topic:
AI-generated writing can be rated highly on surface qualities such as structure, clarity, and fluency, but that does not automatically make it good journalism.
Source:
Scientific Reports: ChatGPT-generated essays compared with human essays
Source type:
Academic research.
What it supports:
The study supports the article’s point that AI-generated writing can perform well in human evaluations of written quality.
Important caveat:
The comparison involved argumentative student essays, not elite reported journalism. It does not prove that AI is better than top human journalists at reporting, judgment, originality, or factual verification.
8. AI content farms and low-oversight publishing
Claim or topic:
AI makes it easy to produce news-like content at large scale, especially outside traditional, high-oversight journalism.
Source:
NewsGuard: AI Tracking Center
Source type:
Analysis.
What it supports:
NewsGuard’s tracking supports the article’s claim that thousands of AI-generated news and information sites have emerged, often producing large volumes of content with little or no human oversight.
Important caveat:
AI content farms are not the same as reputable major newsrooms. This evidence shows the scalability and risk of AI-generated publishing, not that elite outlets are operating in the same way.
9. CNET as a cautionary example
Claim or topic:
AI-written articles can sound fluent while still containing factual or editorial errors.
Source:
The Verge: CNET corrected AI-written stories
Source type:
Reputable journalism.
What it supports:
The reporting supports the article’s use of CNET as a cautionary case in which AI-written personal finance articles were later corrected after errors were found.
Important caveat:
This is one publication and one episode. It does not prove that all AI-written journalism is inaccurate, but it shows why oversight and transparency matter.
How to read this evidence
This article is the author’s analysis. The sources above are provided so readers can see where the factual claims come from and judge the evidence for themselves. Some sources support direct facts, while others provide context, estimates, or background evidence.
Corrections and updates
If a factual error is identified, this post will be corrected in the web version with a dated note explaining the change. Because email versions cannot be edited after sending, the web version should be treated as the current version.