From Bloomberg's decade-old earnings-report bot to Sports Illustrated's fabricated AI bylines, AI in newsrooms splits cleanly into what works and what blew up — while AI answer engines quietly absorb the referral traffic that used to fund it all.
In January 2023, CNET quietly published dozens of AI-generated financial explainer articles — pieces like "What Is Compound Interest?" — under a byline that read "CNET Money Staff." Nobody outside the company noticed until Futurism started fact-checking them and found the AI had gotten basic financial math wrong in more than half the articles it reviewed: miscalculated interest, garbled definitions, plagiarized phrasing lifted close enough to other outlets' wording to draw plagiarism accusations. CNET paused the program, corrected dozens of pieces, and added disclosure labels. It's a useful starting point because it captures the whole shape of AI in journalism in one incident: a real cost-saving use case (routine explainer content is expensive to staff), a real failure mode (a model confidently wrong about facts nobody double-checked), and a business under enough margin pressure to try the shortcut anyway.
That combination — genuine utility, genuine failure risk, and a news industry with collapsing margins — is why AI adoption in journalism looks so uneven. Some applications are boring, well-established, and essentially invisible to readers. Others have blown up publicly and expensively. And the biggest disruption isn't inside the newsroom at all — it's happening to the traffic that used to fund it.
| Application area | Maturity | Representative example |
|---|---|---|
| Automated reporting on structured data | Mature, production since ~2014 | Associated Press corporate earnings stories via Automated Insights; Bloomberg's Cyborg system for market reports |
| Reporter tooling (transcription, research, summarization) | Widely adopted | Otter.ai-style transcription; internal newsroom copilots at large outlets |
| Fully AI-drafted narrative journalism | Early, high failure rate | CNET explainers, Sports Illustrated/AdVon bylines, Gannett's LedeAI sports recaps |
| AI answer engines replacing referral traffic | Disruptive, unresolved | Google AI Overviews, ChatGPT search, Perplexity |
| Licensing AI training/retrieval rights to labs | Emerging revenue line | AP, Axel Springer, News Corp deals with OpenAI; Perplexity's Publisher Program |
The oldest and least controversial use of AI in journalism isn't generative AI at all — it's template-based natural language generation applied to structured data feeds. The Associated Press has used Automated Insights' Wordsmith engine since 2014 to turn corporate earnings filings into short business stories automatically, a move AP said let it scale corporate earnings coverage roughly tenfold without adding reporters, freeing business desk staff to chase stories the software couldn't write. Bloomberg's internal system, known as Cyborg, has done something similar for market-moving financial reports for close to a decade, generating routine stories from data the moment it's released — faster than any reporter could type.
This works because the domain is narrow and verifiable: the input is a structured number (quarterly revenue, a stock price move, a box score), the output is a templated sentence, and there's no interpretive judgment for a model to get wrong. It's closer to a sophisticated mail-merge than to journalism in the reportorial sense, and newsrooms treat it that way — it replaced work nobody wanted to do by hand, not work that required judgment.
The moment outlets tried to extend the same idea to less structured domains, it broke. Gannett rolled out AI-generated local high school sports recaps in 2023 using a tool called LedeAI, and the results went viral for the wrong reasons — stories with robotic, repetitive phrasing like describing a loss as "a tough contest" regardless of what actually happened in the game. Gannett paused the program within weeks. The lesson newsrooms drew wasn't "don't automate," it was "automate only where the data is genuinely structured and the story genuinely is the data."
The most widely adopted AI use in journalism today barely makes news because it doesn't change what gets published — it changes how reporters get there. Transcription tools have replaced manual interview transcription across most large newsrooms. Several major outlets, including the Washington Post and the New York Times, have built internal AI assistants for reporters and editors that summarize documents, search archives, and suggest headlines, deliberately kept separate from anything that ends up as bylined copy without human rewriting. The distinction newsrooms draw — AI as research assistant versus AI as author — maps directly onto where the trust failures have and haven't happened.
None of the newsroom-side use cases above are what's actually reshaping the business of journalism. That's happening on the distribution side, where AI answer engines are increasingly answering readers' questions without sending them to the publisher at all. Google's AI Overviews, rolled out broadly in 2024, place an AI-generated summary above traditional search results — often synthesizing information from multiple articles into an answer that satisfies the query without a click. Multiple industry analyses from outlets like Press Gazette and Similarweb have reported measurable declines in referral traffic to publishers correlated with Overviews' rollout, and the pattern extends to chatbot search generally: when ChatGPT or Perplexity answers a question directly, the publisher whose reporting fed that answer typically gets no visit and no ad impression.
This is the actual existential pressure driving the deals you've probably seen in the news: AP signed a licensing agreement with OpenAI in 2023 granting access to its news archive; Axel Springer (Politico, Business Insider, Bild) struck a multi-year content deal with OpenAI in December 2023; News Corp followed in May 2024 with a deal widely reported to be worth more than $250 million over five years. Perplexity launched a Publisher Program that shares advertising revenue with outlets like Time, Fortune, and Der Spiegel when their content anchors an AI-generated answer. These aren't goodwill gestures — they're publishers trying to convert a traffic source that's disappearing into a licensing revenue line that survives the disappearance.
The technical countermeasure showing up alongside the legal one is crawler gating. Cloudflare began blocking AI crawlers by default for new domains in 2025 and introduced a "pay per crawl" mechanism letting sites charge AI bots per request rather than allow free scraping — a direct response to publishers wanting the same leverage over bots that they never had over Google's original search crawler, which indexed for free in exchange for referral traffic that AI answers no longer reliably deliver. Whether that leverage sticks depends on how much of the open web ends up gated the same way, and on litigation like the New York Times' copyright suit against OpenAI and Microsoft — filed in December 2023 and still working through discovery years later — which will help determine whether training on news content without a license is infringement at all, or whether licensing deals are commercially necessary rather than legally required.
Every major AI-in-journalism scandal so far shares a structure: a model or an AI-content vendor produced narrative text presented as reported journalism, without a human meaningfully checking it before publication. Sports Illustrated's 2023 scandal is the starkest example — Futurism found the outlet had published AI-generated product reviews under fabricated author bylines complete with AI-generated headshots and fake biographies. Sports Illustrated's publisher, Arena Group, said the content came from a third-party vendor, AdVon Commerce, and pulled it, but the reputational damage — for a 70-year-old sports institution — was immediate and lasting. CNET's incident and Gannett's sports-recap pause follow the same pattern at smaller scale: the failure wasn't that AI wrote something, it was that AI wrote something nobody with domain expertise reviewed before it went out under the outlet's name.
That's why AP's editorial standards, published after its OpenAI deal, explicitly restrict generative AI from writing publishable copy and require any AI-assisted material to go through the same editorial process as human-written work — the policy exists precisely because the failure mode is well understood, not hypothetical. Outlets that have avoided public AI scandals aren't the ones that avoided AI; they're the ones that kept it upstream of a human who's accountable for what goes out.
AI in journalism is mature exactly where the task is templated and the underlying data is already verified — earnings reports, box scores — and unreliable exactly where it's asked to exercise judgment or report facts nobody checks afterward. The bigger story for the industry isn't what AI can write; it's that AI answer engines are absorbing the search traffic publishers depended on to fund reporting, and the licensing deals, paywalled crawlers, and pay-per-crawl schemes now emerging are publishers' first real attempt to get paid for the raw material instead of the click.