AI and automation policy

This policy is the group minimum for automated systems, including machine-learning tools. Publications may be stricter. The explanatory page is Technology with accountability.

The policy distinguishes kinds of use because they do not carry the same risk.

Categories of use

• Assisted research. A person asks a tool to help find or organize material. The person checks anything that will be published.
• Search and discovery. Tools may surface documents, data, or prior coverage. A hit is a lead, not a fact.
• Transcription. A transcript is a draft. Names and quotations used in public are checked against the recording.
• Translation. A translation used in public is reviewed by someone who can judge the source language, or it is labeled under the disclosure rule when no person reviewed it before publication.
• Summarization. A summary is an aid. It is not published as the publication’s account unless an editor has compared it with the source, or the item is labeled because no person reviewed it before publication.
• Classification. Labels inside a newsroom, such as topic or priority, may be suggested by a system. A person can override them. Classification is not a published judgment about a person unless the standards for that judgment are met.
• Automated data processing. Counting, matching, and cleaning data are allowed. The method is described when the result is a published figure.
• Generation. Drafting sentences, headlines, or images with a generative system is a restricted use, below.
• Synthetic media. Audio, video, or images that depict events or people who were not recorded that way are a restricted use, below.
• Publication automation. Scheduling, formatting, link checks, and template fills may run without a person watching each step. They may not invent facts in order to fill a template.

Three separate acts

The policy uses three names that are not interchangeable.

• Preparation. A system may gather material, draft, classify, compare sources, or propose a headline. Preparation is not publication and it is not a review.
• Review before publication. A natural person examines the substance, including the factual claims and the sources, and can change or stop the item. A second model, a spell-check, or a glance at the headline is not this review.
• Oversight after publication. A person in an editorial role watches the live record, corrects errors in public, and can stop a workflow. A later review does not change the fact that an earlier item went out without review before publication.

A publication may use one path or more than one. This website does not state which path a given story on a publication site used. That label belongs on the item, under the publication’s own rules.

Human accountability

A person in an editorial role is accountable for what a publication publishes. A person in a publisher role is accountable for what this website publishes.

Accountability means a person in an editorial role owns the public result and the correction path. It is not a claim that a person reviewed every item before publication. It is not a claim that a person re-reads every log line or every automated formatting step.

The presence of people in the organization is not, by itself, review of a particular item. Review by another automated system is not human review.

Disclosure

A publication discloses the use of generative systems when a reader could otherwise think a generated passage, image, or voice was a direct record or a fully human draft of the published wording.

Where an item is published without the review defined above, the publication says so on the article. It does not describe that item as human-reviewed. A general sentence on this website is not that label.

Perspicora’s published rule is the same: when an article is prepared without human editorial review, the article says so. The label is on the article. The publication’s account of those labels is AI and editorial systems.

Ordinary tools — spell-check, transcription as a notebook, search — do not each require a badge. Generation of the published wording, and any synthetic media, do.

This website’s own pages are edited by people. They are not presented as the output of a model.

Verification

Anything a system produces is unverified until it is checked against sources. The check may be the human review defined above, or a workflow the publication discloses. A model is not a source. It is not cited as evidence. It does not satisfy the accuracy standard on its own. A check by another model is not a human review.

Restricted uses

Generation of draft copy is allowed as a draft. Publishing it is allowed only when a responsible editorial role owns the result under this policy. If a natural person has not reviewed the substance before publication, the item is not described as human-reviewed. Synthetic media is allowed only when it is labeled and is not offered as evidence of a real event.

Automated publication of data is allowed when the data source, the transformation, and the update rhythm can be explained, and when a person can stop it.

Prohibited uses

The group and its publications do not:

• present model output as a quotation from a real person
• invent sources, documents, or interviews
• generate a picture or clip of a real event that did not happen and offer it as coverage
• let a system decide not to publish a story in order to please an advertiser, an owner, or a government
• hide that a public passage was generated when the disclosure rule above applies
• use unpublished personal data, unpublished legal advice, or secrets as training or prompt material for a third-party system
• publish journalism from this domain

Corrections

Errors introduced with the help of a tool are corrected the same way as any other error, under the corrections procedure. The correction does not blame the tool and leave the record unchanged.

Auditability

A publication keeps enough of a record to explain, later, whether a tool materially shaped a published story: which class of tool, whether a natural person reviewed the substance before publication, and which role answers for the result. It does not have to publish prompts. It does have to be able to answer an editor or a complaint.

This policy does not claim a completed technical audit log on day one. Where that log does not exist yet, the publication does not use generation for published wording.

Review

This document is reviewed at least once a year, or sooner if a publication starts a new class of use. The review date on this page is the record.