WSJ’s AI Key Points Outgrew the Test—Editors Still Check Them
The Wall Street Journal’s AI-generated Key Points are no longer merely the limited experiment first seen in 2024. The publication now identifies article summarization as an active newsroom use, and its official AI disclosure states that a journalist reviews work containing AI input before publication.
The clearest public update arrived on June 5, 2025: Nieman Journalism Lab’s account of the rollout established that WSJ had moved from an A/B test to including Key Points on core news articles. The feature therefore progressed beyond its original trial, although WSJ has not published the test results or established that every article receives a summary.
How Key Points moved beyond the original experiment
WSJ began developing the summary workflow in early 2024 for Dow Jones Newswires customers seeking the essential information from an article. Once that workflow was integrated into the content-management system, the same infrastructure could be used on the Journal’s consumer site.
The public experiment presented selected users with three AI-generated bullets above an article. The test examined whether Key Points changed subscriber engagement—including reading volume, session length and time spent on the summarized article—and whether the feature affected conversion among nonsubscribers.
After the experiment, WSJ began applying the summaries to core news coverage. The public account identifies Google Gemini as the technology powering Key Points, while the Journal’s broader disclosure separately says it uses Google’s Vertex AI platform for open-source investigations and certain other newsroom tools. Those are distinct statements about different uses, so the broader Vertex AI disclosure should not be treated as a more precise description of the summary system.
The expansion does not amount to a universal summary on everything WSJ publishes. Editors retain discretion to add or remove Key Points, and the format is considered most suitable for core news stories whose facts can be stated clearly.
The summary follows the article through the newsroom
Key Points are generated from the underlying article rather than from an open-ended search across outside material. They move through the Journal’s content-management system with the story, where the newsroom reviews them for accuracy, clarity and house style before publication.
That workflow keeps the model’s role narrow. The system compresses reporting already produced by journalists; it does not conduct interviews, assess source credibility or decide what evidence belongs in the original story. Human review is therefore part of the publishing process rather than an optional check after the summary has appeared.
An editor can also reject a summary even when every bullet is technically supported by the article. Opinion, interpretation and stories built around uncertainty may lose essential meaning when reduced to three points. A concise statement can be factually defensible yet still give disproportionate weight to one number, omit a qualification or make a developing situation appear settled.
The workflow requires maintenance after launch. When an underlying model is retired or changed, the newsroom must update prompts and evaluate the resulting output again. Previous performance does not guarantee that a new model-and-prompt configuration will preserve the same level of accuracy.
Disclosure tells readers who remains responsible
Readers see a module labeled “Key Points.” Its “What’s this?” control explains that an artificial-intelligence tool created the summary from the article and that an editor checked it.
The disclosure matters because it assigns responsibility to the publication. The bullets are not presented as a separate autonomous news product: they are an AI-assisted layer attached to a WSJ article and released through an editorial workflow.
The official policy also places summarization within a broader set of newsroom applications. WSJ and Dow Jones Newswires use AI or machine learning for work that includes data-driven investigations, translation, audio versions of articles and alt text for visuals, while applying stated standards concerning accuracy, accountability, originality, quality, transparency and intellectual property.
Human review reduces risk but cannot eliminate it
A three-bullet summary creates a different editorial risk from a fully fabricated article. Its claims may all appear in the source text while their selection or emphasis still changes the reader’s understanding. Because the module sits before the full story, that framing can influence whether a reader continues and what they take away.
The risk is observable elsewhere in the industry. In February 2026, Associated Press coverage of newsroom AI governance noted that Bloomberg had issued several corrections during the preceding year for mistakes in AI-generated news summaries. That does not establish any comparable failure at WSJ, but it shows why editorial review, disclosure and correction procedures remain material safeguards.
Sentence-level verification is only one part of the task. An editor must also judge whether the chosen bullets represent the article’s central findings and preserve qualifications that affect their meaning. This is especially important for investigations, policy disputes and financial coverage, where uncertainty may be as consequential as the headline fact.
What remains undisclosed
WSJ has not publicly released the engagement or conversion figures from its A/B test. The expansion shows that the organization found sufficient reason to continue using Key Points, but it does not prove that the summaries increased subscriptions, reading time, comprehension or any other particular metric.
The publication has also not provided a public error rate or a fixed eligibility standard covering every section and article type. Readers should therefore expect editorial variation rather than assume that the absence of Key Points reflects a technical failure or that their presence certifies a story as more important.
The current status is nonetheless substantially clearer than it was when the feature surfaced as a test in 2024. AI-generated Key Points became part of the workflow for core news, while editor review, disclosure and the ability to remove a summary remained built into the process. The result is an editor-checked shortcut into WSJ reporting—not evidence that the underlying article was written by AI and not a substitute for its full context.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.