CNET let an AI draft its finance advice under a staff byline

An in-house engine under a staff byline

In November 2022, the money desk at CNET began publishing short personal-finance explainers written by an in-house tool the company called an AI engine. It was not ChatGPT. An editor sketched an outline, the engine drafted the article, and an editor was meant to expand and check it before it ran.¹ Over the following weeks the desk published 77 of these stories, about 1 percent of everything CNET ran in the period.¹

The stories covered the basics a reader might search for: what compound interest is, how a certificate of deposit works, what an APR means. Each one carried the byline "CNET Money Staff." A reader who hovered over that byline found a line saying the article had been generated with automation technology and edited by a human on the staff.² Nobody outside the building had reason to look closer. The stories read like every other explainer on the site.

The error a reader could check

In January 2023, the technology site Futurism read the stories closely and found that several of them were wrong.³ One explained compound interest with a worked example. It told readers that $10,000 deposited at 3 percent annual interest would earn "$10,300 at the end of the first year."³ The number is wrong. At 3 percent, $10,000 earns $300 in a year. The $10,300 is the deposit and the interest added together, not the interest earned. A reader who trusted the sentence would overstate a year of savings by more than thirtyfold.

Its neighbors carried errors of the same kind. On a $25,000 car loan at 4 percent, one story told readers they would pay "a flat $1,000 in interest per year," when the interest falls each year as the balance is paid down.³ Another said a one-year certificate of deposit compounds only once, at maturity, when banks typically compound such a deposit every day or every month.³

CNET later appended a correction to the compound-interest article. It said an earlier version "was assisted by an AI engine and it mischaracterized some aspects of CDs, savings accounts and loan payments," that the points were corrected, and that a staff writer had substantially rewritten the piece.⁴ The article now carries a named human byline.

The corrections, counted from outside

CNET's editor-in-chief, Connie Guglielmo, published an editor's note on January 25, 2023. She wrote that the desk had reviewed the stories and found some that needed fixing: "a small number requiring substantial correction and several stories with minor issues" such as incomplete company names or transposed numbers.¹ She gave no total. The figure repeated everywhere since, that more than half the stories were corrected, did not come from CNET. The Verge counted the correction notes appended to the live articles and reported errors in more than half of them.⁵ Engadget put the count at 41 of the 77.⁶ That number is an outside tally of the published articles, not a figure CNET announced.

Futurism found that many of the AI's sentences closely tracked text already published elsewhere, including on Bankrate, a personal-finance site owned by Red Ventures, the same company that owns CNET.⁷ Where Forbes Advisor had written that "overdraft and NSF fees need not be the norm," the CNET story read "overdraft and NSF fees don't have to be a common consequence."⁷ The wording was changed just enough to cover the source.

CNET paused the tool. It changed the byline from "CNET Money Staff" to "CNET Money," moved the disclosure out of the byline hover so a reader could see it, and added a human editor's name to the pieces.¹

1. Nov 2022

   CNET's money desk starts publishing AI-drafted explainers¹

   The stories run under the byline "CNET Money Staff."

2. Jan 2023

   Futurism reveals the program and its hidden disclosure²

   Later reporting documents the factual errors and the borrowed phrasing.

3. Jan 25, 2023

   CNET's editor publishes a note, pauses the tool, and makes the disclosure visible¹

   Outside outlets tally corrections on 41 of the 77 stories.

4. May 16, 2023

   About 100 CNET workers unionize with the Writers Guild of America, East⁸

   Their organizing statement names automation as a threat.

In May 2023, about 100 CNET writers, editors, and producers formed a union with the Writers Guild of America, East, and their organizing statement said that "automated technology threatens our jobs and reputations."⁸ Reporting at the time placed the AI episode among the reasons they organized, alongside layoffs under Red Ventures.⁹

Why the wrong number sounded right

An editor missed the interest error because it did not look like an error. A language model of this kind is trained to do one thing: predict the next word from the words before it, over and over, choosing what is statistically most likely to come next.¹⁰ Nothing in that process checks a figure against arithmetic or a claim against the world.

So the model wrote "$10,300" because that string was a plausible continuation, not because anything computed the interest. And a plausible continuation reads exactly as fluent when it is wrong as when it is right. The confidence in the sentence is not evidence about the number in it. A 2025 paper by researchers at OpenAI and Georgia Tech, "Why Language Models Hallucinate," made the general point plainly. Large language models "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty," the authors write. They argue the models do this because training and grading reward a confident guess over an admission of doubt.¹¹ That paper came out two years after the CNET stories and says nothing about them. It describes the class of failure, not CNET's particular tool, whose design the company never published. The course covers this mechanism in its own lesson on why models hallucinate.

There is a second effect, and it is the one that turned a software limit into a publishing failure. A rough or hesitant draft invites scrutiny. Clean, confident prose does not. The polish that made the stories publishable is the same polish that made their errors easy to wave through.

This is what separates the CNET case from the AI failures that draw more attention. When a company's chatbot gives one customer a wrong answer, the mistake is a single exchange, and the reader knows a machine produced it. CNET published the machine's output as its own reviewed work, under a human-sounding byline, to everyone who searched the topic. That a model produced a false statement was ordinary. The failure was that a newsroom published it as vetted advice without the one step that would have caught it.