Blog · 2026-10-15 · baryreview editors

The published AI content mistakes of 2025–2026, and what each one teaches

Why keep a list

Most AI generated content mistakes never make the news. A small business publishes a blog post with a made-up statistic, a customer notices or doesn’t, and the page stays up. The cases below are different only because the publisher was big enough that someone wrote about it. The failure patterns are the same ones we see from solo founders and small marketing teams.

We have limited this list to incidents documented by the publisher’s own correction or by reporting we could verify. Where a publisher explained what happened, we quote them. For each case: what happened, the failure pattern, and the check that would have caught it.

1. CNET’s AI explainers (January 2023)

The oldest case here, and still the clearest. CNET had been publishing explainers drafted by an AI tool. After errors were found, it reviewed the series and, according to the Press Gazette tracker of AI journalism mistakes, issued corrections on 41 of 77 stories and paused the program.

Pattern: volume without a per-article reviewer. The tool produced fluent, plausible text, and nobody with subject knowledge read each piece before publication.

The check: a human reads every article and recalculates every number the article derives. An explainer that says “you would earn $X” is a claim with a formula behind it. Run the formula.

2. Sports Illustrated’s invented authors (November 2023)

In November 2023, product reviews on Sports Illustrated’s site were found to carry bylines of writers who did not appear to exist, with apparently AI-generated headshots. The Arena Group, which operated the site, said the content came from a third-party vendor, AdVon Commerce, and that “we are removing the content while our internal investigation continues,” per the Press Gazette tracker.

Pattern: outsourced content with no verification of who wrote it or how. The whole apparatus around the text (author, credentials, photo) was fabricated, and the publisher hadn’t looked.

The check: if a byline is on your site, you should be able to say who that person is. For vendor content, ask in writing how it is produced and who is accountable.

3. The summer reading list with books that don’t exist (May 2025)

On Sunday, May 18, 2025, the Chicago Sun-Times ran a 64-page syndicated supplement called “Heat Index.” Its summer reading list recommended titles such as “Tidewater Dreams” by Isabel Allende and “The Last Algorithm” by Andy Weir. Those books do not exist. Several other real authors were credited with invented titles. The Philadelphia Inquirer ran the same section. Both papers had received it from King Features, a Hearst syndication service, according to reporting carried by Boston.com.

The freelance writer, Marco Buscaglia, said he had used AI and not checked the result: there was “no excuse,” and “I totally missed it.” The Sun-Times said the section “was inserted into our paper without review from our editorial team,” per Press Gazette. The Inquirer’s publisher, Lisa Hughes, called the use of AI to produce content “a violation of our own internal policies and a serious breach.”

Pattern: hallucinated specifics. The AI was asked for a reading list and produced one that looked exactly like a reading list. Real author names made the fake titles credible. Syndication then carried the error into two major papers whose editors never saw it.

The check: every proper noun gets looked up. A book title takes fifteen seconds to confirm in a library catalog. This is the cheapest check in editing, and the one most often skipped because the output looks so finished.

4. The freelancer who was “an AI fabrication” (May–August 2025)

In 2025, Wired removed a feature by a contributor named Margaux Blanchard after, in Wired’s words, a closer look “made it clear to us that the story had been an AI fabrication.” Business Insider later removed essays under the same name; Press Gazette reports the count reached 38. The articles had used case studies of people that Press Gazette could not verify, per the tracker.

Pattern: invented examples presented as real. Human-interest anecdotes are exactly the detail an editor takes on trust because they’re hard to check and seem harmless.

The check: for any named person, quoted or described, ask the writer for contact details and confirm at least one. If a case study can’t be verified, cut it or label it a composite. This applies with full force to business blogs that run “customer stories”: if you didn’t talk to the customer, don’t publish the story.

5. The Deloitte Australia report with fabricated citations (October 2025)

Deloitte Australia produced a 237-page report for the Department of Employment and Workplace Relations on an automated welfare-compliance system, under a contract worth about AU$440,000. A University of Sydney researcher, Chris Rudge, found that it cited academic papers that do not exist and quoted a federal court judgment that said no such thing, according to AP reporting. The revised version disclosed that Azure OpenAI had been used in preparing the report. Deloitte agreed to repay the final installment of its fee and said the “matter has been resolved directly with the client”; the department said the report’s findings and recommendations were unchanged. The Conversation covers the episode in a broader piece on liability.

Pattern: fake sources. A citation is a claim that a document exists and says a specific thing. Language models produce citation-shaped text with great confidence, and a long report with many footnotes tempts a reviewer to spot-check.

The check: open every citation. Not a sample. If a reference is a paper, find the passage. If it is a court judgment, read the paragraph. The citation check would have cost a day.

6. Ars Technica’s fabricated quotes (February 2026)

On February 13, 2026, Ars Technica published a story about an AI agent that had written a critical blog post about a software maintainer, Scott Shambaugh. The Ars story quoted Shambaugh saying things he had not said; he wrote on his own blog that he had not spoken to Ars. Ars pulled the article the same day. Its editor-in-chief, Ken Fisher, wrote that the article contained “fabricated quotations generated by an AI tool,” called it “a serious failure of our standards,” and said that “direct quotations must always reflect what a source actually said,” according to 404 Media. The reporter said he had accidentally used a ChatGPT-paraphrased version of Shambaugh’s blog instead of the original text. Nieman Lab reported in March 2026 that the reporter was no longer with the publication.

Pattern: paraphrase presented as quotation. The tool was used to pull material from a source document. The output was a summary in the model’s words, and it was pasted inside quotation marks. Nothing was “invented” in the usual sense; the words were simply not the source’s.

The check: every direct quote is compared against the original. If a quote came through any tool, the tool’s output is not the source; the original document is.

7. The New York Times quote that was an AI summary (April–May 2026)

An April 14, 2026 Times article attributed a remark to Canadian Conservative leader Pierre Poilievre. It turned out to be an AI tool’s summary of his views, rendered as a direct quotation. The Times corrected the article and published an editors’ note that said “the reporter should have checked the accuracy of what the AI tool returned,” per Press Gazette and the note as reproduced by Simon Willison.

Pattern: the same as Ars, two months later. A tool that compresses a speech produces a sentence that sounds like the speaker. The reporter trusted it.

The check: the same one. A quote must match a transcript, recording, or document you have yourself.

8. The Telegraph’s leftover chatbot text (May 2026)

On May 13, 2026, a Telegraph story about Trump and Xi contained a paragraph that read as a chatbot’s offer to help edit the piece. It was removed shortly after publication; the Telegraph declined to comment, per Press Gazette. The tracker lists similar 2026 cases at Marie Claire Australia (a reporter’s prompt left in an article) and Bristol Live (a chatbot-style reply at the end of a social post).

Pattern: mechanical. No fact was wrong. The paste from an AI tool included the tool’s conversational framing, and no one read the final paragraph before publishing.

The check: read the whole thing once, top to bottom, after the last edit. Not the parts you changed. All of it. This error gets through when the only reader was the person who was rushing.

9. The Financial Times column condensed by AI (August 2026)

In August 2026 the FT published an editors’ note on a column by an outside academic: “It has come to our attention that AI was used to condense a longer draft of this column.” The FT’s code of conduct does not allow AI-written article text, per Press Gazette.

Pattern: a disclosure failure rather than a factual one. The column’s argument was the author’s; the sentences were partly not, and the FT’s readers and policy both assumed otherwise.

The check: ask contributors, in writing, how the text was produced, and set the rule before the piece arrives. If you run guest posts or freelance content, your policy should say what AI use is allowed and what must be disclosed.

The courtroom version: over 2,000 decisions

The pattern outside journalism is the same, with higher stakes. The database maintained by legal researcher Damien Charlotin tracks court decisions worldwide in which a court found, or implied, that a party relied on AI-fabricated content, mostly citations to cases that do not exist. When we checked it on October 15, 2026, it listed 2,184 cases identified so far. Charlotin told Scientific American in May 2026 that the rate had settled at “around 350, 400 decisions a quarter.” That article describes an Alabama Supreme Court sanction in April 2026 against an attorney whose briefs cited cases that did not exist.

Pattern: fake sources again, in a setting where every source is checkable and the reader (a judge) checks.

The check: the same as for Deloitte. Open the citation.

What the pattern looks like

Nine cases, four failure modes:

None of the checks are sophisticated. None require software. Every one is something an editor does by habit, which is the point: the organizations above had editors, and the failures happened where the editor was taken out of the loop by syndication, a vendor, speed, or trust in a tool’s output as if it were the source.

If you publish AI-assisted content without an editor, you are in the position those outlets were in for the piece that went wrong. The list of what we check is at /what-we-check; the first two items (facts and hallucinated specifics) cover most of this page.

Practical takeaway

Before anything AI-assisted goes live, look up every name, title, and citation; compare every quote to its original; and read the final version top to bottom once. Those three steps would have prevented every incident on this list, and they take less time than writing the correction.