People remain accountable
At HAI Magazine, AI may support research, planning, writing, editing, translation and media creation. Authors are accountable for the material they submit, and authorized editors for their publication decisions. Using a tool does not transfer responsibility for accuracy, rights or consequences to that tool.
We describe people’s actual contributions. AI tools and editorial AI personas are not human coauthors or interviewees. An AI assessment, independent human review and a publication decision are separate stages. A process disclosure does not establish that any of them took place.
Readers can see how a piece was made
A published piece includes a “How this piece was made” disclosure. The author identifies tools and their contribution, their own work, checks performed and material limitations. Model names and versions are provided when known. A generic “written with AI” does not explain the actual process.
The disclosure concerns a particular edition. We do not retrospectively assign tools, reviews or checks to historical articles without documentation. A missing disclosure means the process is unknown, not that no AI was used. We preserve an older disclosure as recorded; its presence alone does not establish compliance with today’s standard.
Facts, sources and interpretation
We check material claims against appropriate primary sources: documents, studies, recordings, data or product documentation. We read the source rather than rely on an AI-generated summary. We connect evidence to a specific claim and retain its scope, date and limitations. A vendor announcement establishes what the vendor claims; it does not replace an independent measurement.
We distinguish facts from opinions, interpretations, hypotheses and scenarios. A link to an AI Race or AI Regulations fact helps readers find the record and its sources; an Atlas AI profile provides context. None of these links automatically verifies the full argument. Legal, health, financial and safety claims need current sources and appropriate subject-matter review. We state uncertainty explicitly.
Interviews and quotations need evidence
We do not invent interviews, quotations, people, experiences or research findings. Quotations require a verifiable record and faithful context; paraphrases are identified as such. Reconstructions, fictional dialogues and hypothetical examples are labelled before the passage so readers do not mistake them for reporting. AI must not fill gaps in an interviewee’s words as if they were authentic statements.
Privacy, rights and media labels
Before submitting material to an AI tool, we check our authority to use it and the processing conditions. Without that basis, we do not submit confidential documents, private correspondence or personal data. We use only the data needed. The process disclosure is public: it must not contain private prompts, links to working documents or other people’s private data.
We check the provenance and rights of text, images, voices and recordings. AI generation does not establish permission to use material or someone’s likeness. We label AI-generated images, audio and video, and substantial AI alterations, beside the item and describe the scope. Synthetic illustrations must not be presented as documentary records of events. Consent for voice, likeness and public attribution remains a separate matter.
Translations have their own review scope
A translation identifies the source language and edition and the role of AI tools. We describe the checks actually performed on meaning, terminology, quotations, numbers and limitations. Checking selected passages does not mean the whole text was reviewed. A translation is not a new independent source, and language review does not verify facts. A material change to the original calls for checking the translation again.
Relevant relationships must be disclosed
We disclose promotional, financial, professional and organizational relationships relevant to the topic, including sponsorship and product access provided by a supplier. Such a relationship is not evidence and must not be concealed by neutral-sounding AI prose. Independent review and editorial decisions remain subject to the collective’s rules.
Corrections and limits of certainty
Before publication, we recheck changing information that matters to the conclusion: tool versions, availability, prices, rules and deployment status. Event dates are distinct from publication and verification dates. An enduring essay may rely on historical facts; age alone does not make them wrong.
We assess reported errors against the specific passage and evidence. A confirmed correction should explain what changed and why; a new version follows the appropriate editorial process. We do not describe unfinished checks as completed. Readers can report a problem using the feedback button on the piece.
An AI detector does not establish authorship
We do not treat an AI-detector result or writing style as proof of AI use or non-use, plagiarism or authorship. We investigate questions through sources, versions, a documented process and discussion with the author. A detector result alone is not grounds for attributing a tool or making an allegation.
Six points for a process disclosure
In the material form, describe the actual process in “How the piece was made”. This outline helps prepare the account; it is not a statement that work was completed or a new consent to accept.
- Human contribution. Who developed the topic, argument and first-hand observations? Describe the authors’ actual work.
- AI tools and scope. Name the tools, their version if known, and tasks: research, outlining, prose, code or copy-editing. State that no AI was used only when that is known for the whole process.
- Fact and source checks. Explain who checked material claims, quotations and primary sources, to what extent and when. Identify unchecked elements; an AI answer alone is not verification.
- Editing and decisions. State what a person changed, rejected or read in full. Do not claim independent review or approval before it has happened.
- Images, audio and translations. Describe AI use in illustrations, recordings and translation, their labels and human checks. For a translation, identify the source language and edition. State when this does not apply.
- Limitations and relationships. Disclose material gaps, limits to currency, and promotional, financial or organizational relationships. Do not include private data or confidential material.
AI policy assessment checklist
This separate checklist supports a deliberate review of a specific edition. It does not change Mary’s six assessment dimensions or run an automatic assessment. The outline below is a template: without a completed, documented check, the result is “not assessed”.
- Met ·
met - The requirement was checked within the stated scope and evidence recorded. This does not verify the entire piece or other people’s consent.
- Revise ·
revise - A specific gap or error needs correction. Record the passage, necessary change and how to check it again.
- Critical ·
critical - An established serious problem, such as a fabricated quotation or disclosure of confidential data, requires editorial resolution before publication. This is a recommendation, not an automatic system block.
- Not assessed ·
not-assessed - The check was not performed, or necessary evidence or access is missing. An absent result does not mean the requirement was met.
Does the disclosure identify actual human work, tools, AI’s scope and limitations, without invented historical declarations?
Result in the blank template: Not assessed.
Do key claims have checked sources, scope and dates, with interpretations and attributed claims distinguished from facts?
Result in the blank template: Not assessed.
Do quotations, interviews and findings have a real evidential basis, with reconstructions and scenarios clearly labelled?
Result in the blank template: Not assessed.
Are media provenance and alterations labelled, along with AI’s role and the scope of translation review?
Result in the blank template: Not assessed.
Are necessary rights and permissions documented, privacy respected and relevant relationships disclosed? What could the AI review not confirm?
Result in the blank template: Not assessed.
Are the edition, date, completed checks and gaps recorded, separating AI assessment from human decisions and identifying necessary corrections?
Result in the blank template: Not assessed.
For each actual assessment, record the result, passage and evidence, impact, concrete revision, confidence and limits. Identify the date and edition assessed. Unresolved rights or inaccessible sources require human follow-up; an AI detector does not replace evidence.