Ghost in the Machine, Part 3: The Anatomy of a Scandal

This is the third installment in our six-part series, “Ghost in the Machine,” which explores the hidden use of artificial intelligence in journalism and the media’s growing trust crisis. You can read the Part 1 here and Part 2 here

In the first two parts of this series, we uncovered the shocking scandals at publications like Elle Belgium and established a clear pattern of deception connecting major brands from Sports Illustrated to CNET. These recurring failures are not merely the result of poor decision-making; they are rooted in the fundamental nature of the technology itself.

When publishers misapply generative AI, treating it as a sentient author rather than a probabilistic tool, the results are predictably flawed. The failures manifest in several distinct ways, from fabricating information and plagiarizing sources to producing soulless prose, each stemming from a core mismatch between how AI functions and what journalism requires.

The Hallucination Engine: AI’s Problem with Truth

The most significant technical pitfall of large language models (LLMs) is their tendency to “hallucinate”, a term for when the model generates information that is plausible-sounding but factually incorrect or entirely fabricated. Because these models are designed to predict the next most likely word in a sequence to create convincing text, not to verify truth, they can state falsehoods with unwavering authority.

The CNET case provides a textbook example. Its AI-powered financial articles contained multiple basic errors that no human expert would make, such as the wildly incorrect calculation of compound interest. The AI presented this wrong information confidently, demonstrating how these systems can mislead readers, particularly those with low financial literacy who are the most likely to seek out such basic explainers. The problem is not limited to numbers. In a stark 2024 case in Wyoming, a reporter for the Cody Enterprise admitted to using AI to generate quotes for his stories, attributing fabricated statements to real public officials, including the state’s governor. The officials confirmed they had never spoken to the reporter, revealing that the AI had simply invented believable-sounding quotes. This moves beyond simple error into the realm of outright fabrication, a cardinal sin in journalism.

The Plagiarism Problem: Unoriginal Sin

A second critical failure is plagiarism. Because LLMs are trained on enormous datasets scraped from the internet, which include vast amounts of copyrighted material, their output can closely mirror or directly lift text from existing sources without attribution. This is not plagiarism in the human sense of intentional theft, but an inherent function of a system designed to reproduce patterns it has learned from its training data.

Again, the CNET scandal is illustrative. A follow-up investigation by Futurism found that numerous AI-generated articles on the site bore “deep structural and phrasing similarities” to content previously published by competitors like Forbes and other financial websites. The AI would swap out a few words for synonyms or make minor syntactical changes, but the underlying structure and language were clearly derived from the original source. This practice not only represents a profound ethical breach but also wades into a complex legal battleground over copyright and fair use that is currently being litigated in courts, with major publishers like The New York Times suing AI companies for using their content for training without permission.

The Soulless Prose of the Content Mill

Beyond factual errors and plagiarism, there is a qualitative failure that defines much of the AI-generated content seen in these scandals. The prose is often generic, formulaic, and devoid of the nuance, creativity, and authentic voice that characterize good writing. It is content designed not to engage a human reader, but to satisfy the cold logic of a search engine algorithm.

The travel articles published by BuzzFeed’s “Buzzy the Robot” are a prime example. They were filled with repetitive, clichéd phrases like, “Now, I know what you’re thinking…” and generic recommendations that lacked any real insight or personality. Similarly, the product reviews from Sports Illustrated’s fake authors were described as sounding like they were “written by an alien,” with awkward phrasing and a complete lack of genuine human experience. This soulless quality reveals the ultimate purpose of such content: to occupy digital space, attract clicks through SEO, and generate affiliate revenue. It is the antithesis of journalism meant to inform, enlighten, or entertain.

Embedded Bias and Other Dangers

Finally, AI models carry the risk of perpetuating and amplifying the societal biases present in their vast training data. If the data reflects historical biases against certain demographic groups, the AI’s output will likely reproduce those biases, creating content that is unfair or stereotypical. Beyond bias, the lack of human judgment can lead to glaringly insensitive or inappropriate content. A notorious example occurred in September 2023, when an AI-written obituary for former NBA player Brandon Hunter, published on Microsoft’s MSN, was headlined “Brandon Hunter useless at 42”. This egregious error, likely a misinterpretation of a phrase from a source text, demonstrated a complete lack of the sensitivity and common sense required for such a topic, causing significant public backlash and further eroding trust in AI-generated news. These failures underscore that the core issue is a fundamental mismatch: publishers are using a tool of probabilistic synthesis for the journalistic task of factual reporting, a purpose for which it is dangerously unsuited without rigorous human oversight.

Next up in Part 4: If the technology is so flawed, why are reputable publishers risking their credibility to use it? The answer lies in the brutal economics of modern media. Next, we’ll follow the money.