CNET’s AI-written articles

From November 2022, the technology news site CNET used an internally designed AI engine to help write personal finance explainers for its CNET Money section, with titles such as “What Is Compound Interest?”. Editors produced outlines, and then expanded and edited the AI drafts before publication. The stories carried the byline “CNET Money Staff”.¹

The experiment attracted little attention until January 2023, when the website Futurism reported that CNET had been quietly publishing AI-written articles for months without a formal announcement.²

What went wrong

Readers and journalists quickly found errors. The compound interest article, for example, suggested that a saver depositing $10,000 would earn $10,300 after a year. The correction made clear that the saver would earn $300 on top of the original $10,000.¹

CNET carried out an internal review of all 77 stories. It issued corrections on 41 of them.² Its editor-in-chief, Connie Guglielmo, said the audit found further stories needing correction, “with a small number requiring substantial correction”, and that some corrections addressed phrases that “were not entirely original”, raising concerns about plagiarism.¹ She added: “We’ve paused and will restart using the AI tool when we feel confident” that the tool and editorial processes would prevent errors.¹

CNET also changed the byline from “CNET Money Staff” to “CNET Money”, after critics said the original did not make the role of AI clear.¹

The lasting cost

The damage did not end with the corrections. In early 2023, Wikipedia’s volunteer editors reviewed CNET as a source. As Futurism reported in February 2024, they concluded that its content from November 2022 to January 2023 should be considered “generally unreliable”, citing the AI-written articles, and that content after its 2020 change of ownership should be treated with caution.³ For a publisher whose value depends on being trusted and cited, that is a significant loss.

Why it failed

Three things combined.

The tool was used where it is weakest. Generative AI produces fluent, confident text. It is not a reliable source of facts or arithmetic. Personal finance explainers depend on exactly those things.

Review did not scale with output. CNET’s process included human editing. However, an editor reviewing fluent copy at volume is prone to miss the subtle error, and the errors that got through were not subtle to readers who checked.

It was done quietly. The use of AI was not clearly disclosed. When it came to light, the story became not just “AI made mistakes” but “a trusted publisher did not tell its readers”, which is far harder to recover from.

Lessons for leaders

1. Match the tool to the task. Use generative AI for work where fluency matters more than precision, or build verification into the workflow for anything factual, numerical or regulated.
2. Size review to the risk, not the schedule. If AI multiplies output, the checking effort must grow with it. Otherwise human review becomes a formality.
3. Disclose before someone else does. Being open about how AI is used costs little. Being found out costs trust that took years to build.

Questions to ask in your organisation

1. Where are we already publishing or sending AI-assisted content to customers, citizens or investors, and would they know?
2. For that content, what specifically does our human review check, and how much time does it actually get per item?
3. If a journalist reported on our use of AI tomorrow, would we be comfortable with how it was described?