What we counted
In our parse of Wikipedia's perennial sources list we captured the full summary text of all 497 classified entries on the live list. This report searches that text for entries where AI or large language model use is given as a reason bearing on the source's reliability.
Twelve entries qualify, or 2.4% of the list. That is a small number in absolute terms and a striking one in context, because the concept barely existed as a sourcing question three years ago and now appears in the reasoning for a dozen publications.
One candidate was excluded deliberately. The entry for Amnesty International matches a naive search for the letters AI because that is the organisation's own abbreviation, and it has nothing to do with generative content. Automated text matching would have returned 13; the correct figure is 12.
We report the entries as summaries of community discussion, which is what they are. The list states it is “not a policy or guideline”, and its statuses record arguments editors have already had rather than rules applied mechanically.
Naming AI is not automatically disqualifying
The status spread is the first surprise. Of the twelve, five are rated no consensus, four generally unreliable, two generally reliable and one deprecated. If merely touching AI ended a publication's standing, the distribution would not look like this.
The two generally reliable cases are instructive because the carve-out is surgical rather than general. Gizmodo remains “generally reliable for technology, popular culture, and entertainment”, and The A.V. Club remains generally reliable for film, music and TV reviews. In both entries editors then record a specific exception: “There is consensus that AI-generated articles are generally unreliable”.
Both entries attribute those articles to the same parent, noting that G/O Media “began releasing such pieces in July 2023, usually under the byline” The A.V. Club Bot or Gizmodo Bot respectively. The byline is what makes the carve-out enforceable: a reader or editor can tell which articles are covered.
That is the shape of the reasonable outcome. The publication keeps its standing for the work its journalists do, and loses it precisely where machines wrote the copy, because the boundary is visible.
VentureBeat shows the softer end of the same logic, and the consequence that matters commercially. Its entry records no consensus overall, noting that “some editors expressed concern over the site's publication of churnalism and AI-generated articles, and argued against using the site to establish notability”. The publication is not barred; it simply stops counting for the one purpose most buyers of coverage care about. The Times of India's entry is milder still, cautioning only that editors “should be aware that TOI may have published at least one AI-generated article”.
Entries naming AI or LLM use as a reliability factor, of 497 classified entries
Source: Media Morpheus, September 2026. Original parse of Wikipedia:Reliable sources/Perennial sources, subpages 1 to 8.
The Amnesty International entry is excluded: AI there is the organisation's abbreviation. The two generally reliable entries carry explicit carve-outs for their AI-generated articles.
Concealment is the aggravating factor
The harshest outcomes track disclosure rather than use. CNET's entry is the clearest case. It records that in November 2022 the publication “began deploying an experimental AI tool to rapidly generate articles riddled with factual inaccuracies and affiliate links, with the purpose of increasing SEO rankings”.
The decisive sentence is the next one: CNET “never formally disclosed its use of AI until Futurism and The Verge published reports exposing its actions”. The entry adds that the tool “wrote more than 70 finance-related articles and published them under the byline CNET Money Staff, over half of which received corrections after mounting pressure”.
Note what compounds the damage. The articles were generated for ranking rather than for readers, they carried a byline implying human staff, and the correction rate exceeded half. Each of those is a separate breach of the thing a source is trusted for.
Sports Illustrated's entry follows the same arc, recording that in November 2023 “it was revealed that artificial intelligence (AI) had been used to write product review articles with little to no human oversight” — again revealed rather than announced.
A third mechanism appears in the ScienceDirect entry, and it has nothing to do with deception or accuracy. Elsevier's machine-generated topic pages are rated generally unreliable because they “change over time, presenting a challenge to verifiability”, with editors directed to cite “the actual, underlying publications quoted by the topic page” instead. Generated text fails here simply by being unstable: a citation is a promise that a reader can check the same words later, and a page that regenerates cannot keep it.
Disclosure alone does not protect you either
It would be convenient to conclude that disclosure solves the problem. The evidence does not support that, and this is the finding most likely to be misreported by people selling AI content services.
Business Insider's entry records that in a 2025 request for comment, “editors determined that any LLM-generated writing disclosed as written using AI is generally unreliable”. Disclosure there does not rehabilitate the content; it identifies content that is then classified as unreliable.
Newsweek's entry notes as a live consideration that “as of April 2024, Newsweek has disclosed that they make use of AI assistance to write articles”, sitting alongside the other reasons its post-2013 material is not generally reliable.
So the two mechanisms operate differently. Concealment attracts the severest downgrades because it breaches the reader's basic expectation about who wrote something. Disclosure avoids that breach, but it also labels the material, and labelled generative content is then assessed on its own merits — which in Business Insider's case meant being declared unreliable as a class.
It can consume an entire publisher group
The most consequential entry on the list is not a publication at all. Red Ventures, a company, carries its own entry rated generally unreliable, recording consensus that “the online properties of Red Ventures are generally unreliable post-acquisition”.
The stated reason is corporate policy: “Editors express concern that Red Ventures, as a matter of policy, uses AI-authored content on its properties in a non-transparent and unreliable manner.” That judgement then propagates. ZDNet's entry applies it directly, and The Points Guy's entry cites the acquisition by “Red Ventures, a publisher of LLM-generated content which is itself considered generally unreliable” among the reasons it was deprecated in a 2026 request for comment.
This is a genuinely new risk for anyone buying coverage. A publication's standing can now be damaged by what its corporate parent does on entirely different properties, without any change in the newsroom you dealt with.
The entries also record the reverse possibility. Both CNET and ZDNet note purchases by Ziff Davis in August 2024, with editors observing that this “may mean that the reasons for considering it unreliable may no longer apply”. Ratings follow ownership and policy, and can move back.
What this means for anyone buying coverage
The practical test to apply before commissioning is not whether an outlet uses AI. It is whether you can tell, from the published page, which articles were machine-written. Where the boundary is visible — a bot byline, a standing disclosure — the publication's human reporting has generally kept its standing.
Check the corporate parent as well as the masthead, because one entry on this list attaches to a company rather than a publication and flows down to its properties. That is not something a media kit will tell you.
If your objective is encyclopaedic standing, coverage in a source currently rated unreliable does not help you regardless of how prominent the logo is, and ratings tied to AI policy have moved faster than most reputational judgements ever have. Two of these entries changed within the past two years.
Media Morpheus sells media and podcast placement. We publish this because the same logic applies to anything we might produce for a client: content whose authorship a reader cannot determine is the category being penalised here, and that is a standard worth holding ourselves to rather than a risk to disclose in small print.
Methodology and limitations
We searched the full summary text of all 497 classified entries on the live perennial sources list, captured in our parse of 18 September 2026, for terms indicating that AI or large language model use featured in the reasoning: AI-generated, AI-authored, AI assistance, AI tool, artificial intelligence, LLM-generated, large language model, machine-generated, generative, ChatGPT and related phrases. Twelve entries matched on substance. We manually excluded the Amnesty International entry, where AI is the organisation's abbreviation, which is why our figure is 12 rather than the 13 a naive text match returns. Quotations are verbatim from the entries. Limitations: the list is selected for controversy rather than representative of publishing, statuses record consensus at a point in time and change as new discussions conclude, and the page states explicitly that it is not a policy or guideline. A count of 12 is small enough that a handful of new discussions could shift the distribution materially. We did not assess whether any publication's AI practices have since changed.
Conclusion
Twelve of 497 entries on Wikipedia's perennial sources list now cite AI use in their reasoning, and the pattern within them is consistent. Publications that concealed generative content, notably CNET and Sports Illustrated, were downgraded hardest, and in CNET's case the concealment, the ranking motive, the human-implying byline and a correction rate above half compounded each other. Publications that disclosed did not thereby earn a pass: Business Insider's 2025 discussion classified disclosed LLM writing as unreliable outright. Most consequentially, Red Ventures is rated unreliable as a company over its AI policy, and that judgement reaches its properties. Where a publication kept its standing, it was because the boundary between human and machine writing was visible on the page.
Frequently Asked Questions
Does Wikipedia consider AI-written articles unreliable?
In the cases recorded, yes, though usually as a carve-out rather than a blanket rule. Gizmodo and The A.V. Club remain generally reliable overall while their entries state “There is consensus that AI-generated articles are generally unreliable”, identified by bot bylines introduced in July 2023.
Is disclosing AI use enough to protect a publication's standing?
Not necessarily. Business Insider's 2025 request for comment determined that “any LLM-generated writing disclosed as written using AI is generally unreliable”. Disclosure prevents the breach of trust that concealment causes, but the labelled content is still judged on its own merits.
Why was CNET downgraded?
Its entry cites an AI tool deployed from November 2022 generating articles “riddled with factual inaccuracies and affiliate links, with the purpose of increasing SEO rankings”, published under a staff-implying byline, with over half receiving corrections — and CNET “never formally disclosed its use of AI” until journalists exposed it.
Can a parent company's AI policy affect an outlet I buy coverage in?
Yes. Red Ventures carries its own entry rated generally unreliable over a policy of using “AI-authored content on its properties in a non-transparent and unreliable manner”, and that reasoning is applied to ZDNet and cited in the deprecation of The Points Guy.
Can a rating recover after an ownership change?
The entries allow for it. Both CNET and ZDNet note purchases by Ziff Davis in August 2024, with editors observing this “may mean that the reasons for considering it unreliable may no longer apply”. Ratings track ownership and policy and can move in both directions.
Sources
- official Wikipedia. Reliable sources/Perennial sources: the entries for CNET, Business Insider, Newsweek, Red Ventures, ZDNet, The Points Guy, Sports Illustrated, Gizmodo, The A.V. Club, VentureBeat, The Times of India and ScienceDirect topic pages.
- vendor research Media Morpheus, September 2026. Original parse of all 497 classified entries on the live perennial sources list, searched for AI and LLM reasoning, with the Amnesty International abbreviation match excluded manually. Method and limitations stated above.
- official Google Search Central. Spam policies, for the separate question of how search treats automated content produced primarily to manipulate ranking.