About Using Generative AI

Last Updated: October 7, 2026
Contents

This document presents PFN’s policies on using generative AI to help (a) develop the ideas, thinking, and other intellectual aspects of Content and/or (b) express those intellectual contributions in text, visuals, and other forms.

These policies apply to PFN’s substantive policy content, including Solution Reports, Domain Overviews, notes, articles, blog posts, meeting presentations and discussions, comments, and online discussion posts. They do not apply to the Website’s About pages, governing documents, forms, and similar administrative material.

Many argue that generative AI can improve content ideation, research, brainstorming, discussion, debate, judgments, and decisionmaking. And that it can assist in writing, organizing, editing, and preparing content for others.

And many fear that generative AI use can harm the credibility of and trust among people, organizations, their ideas and content, and their audiences. It can make mistakes and even hallucinate facts and analyses, undercutting many of its benefits. Some also object to generative AI, and the data centers behind it, on the basis of content ownership and copyright issues, environmental harm, energy use, tax expenditures, secretive deals, effects on education and critical thinking, authorship and copyright misappropriation, jobs, socio-economic inequality, misuse by bad actors, and loss of control scenarios.

PFN takes no position on generative AI generally or on these specific perspectives.

Instead, PFN’s general view is that PFN, Network Groups, and PFN Participants can use generative AI in any way to assist in generating intelligence (e.g., ideas and thinking) and/or the expression of intelligence (e.g., writing), but in every case where generative AI is used in significant ways, that use must be disclosed. And disclosure should be made also in some instances of less significant use, such as when readers would be likely to expect disclosure. Also, PFN stresses that disclosing generative AI use does not, in any case, override PFN’s policies on content verification, plagiarism, authorship and copyright ownership, decisions about people, privacy and security, and certain other matters.

For PFN purposes, generative AI means an artificial intelligence system that can make contributions to the ideas, thinking, and other intellectual aspects of Content and/or to the expression and organization of those intellectual contributions, including through text, images, audio, video, code, or other content. The term includes large language models and image, audio, and video generation systems, whatever their underlying architecture, including open and closed systems and general and more specific or targeted instances. Examples include ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), Midjourney, and Llama (Meta).

Capitalized terms that this document does not define have the meanings given in the PFN Glossary. See Related Content below.

General Principles

PFN’s generative AI use policies rest on several key ideas:

  • Disclosure of generative AI use should expand as content shifts from Exploratory Content developed through individual exploration, to Exploratory Content shared within a Gated Audience, to Beta Content, to early and intermediate drafts of Shared Content, and finally to near-final and final versions of Shared Content.
  • In the case of Exploratory Content, disclosure of generative AI use should be made when Gated Audience users would likely find the disclosure important, such as when they are likely to (wrongly) believe that the ideas, thinking, or words used are the author’s or speaker’s.
  • When generative AI is used to develop Shared Content in more a significant way, disclosure should increase from “not required” to “should be considered” to “required” as the use shifts from background, early exploration, and ideation and brainstorming, to early deliberations, analyses, and drafts, to intermediate thinking and drafts, and finally to near-final and final deliberations versions.
  • Disclosure beyond the minimum required is always permitted and is often desirable.
  • PFN and Participants should give complete and candid answers to questions asked by any Participant or user about whether and how generative AI was used.

The table below assigns disclosure levels by several specific activities, and the third principle above describes how disclosure rises through the stages of a project. Where the two point in different directions for a particular piece of work, follow the higher of them.

Disclosure by Content Type

Exploratory Content: individual exploration. A person using generative AI to explore ideas, sources, experts, organizations, or data, and to develop and brainstorm early outlines and versions of content, is not expected to disclose anything while the work remains their own or is shared only with the person’s co-authors, researchers, reviewers, and other contributors.

Exploratory Content: sharing within a Gated Audience. Network Group discussions and content depend on the ability of Network Group Participants to think out loud, try ideas, and trust and build on what they say to each other. Because PFN wants to facilitate Network Group discussions and content that resemble individual exploration but are with a broader Gated Audience of all Network Group Participants in a Gated Venue, a Participant who uses generative AI in a significant way to help generate ideas and thinking or to express content, including spoken and online comments, should consider disclosing the use. More practically, the Participant should consider whether others in the Gated Audience are likely to expect that generative AI use would be disclosed, and whether the failure to disclose would undermine trusting and constructive collaboration within the Gated Audience.

When disclosure is the better approach within a Gated Audience, the Participant may communicate it in a clear and simple way consistent with this Exploratory Content context. In many spoken discussions and informal or early draft writings, disclosure can be as simple as “Claude [or another AI tool or model] says…” or “the following idea came from AI.”

Beta Content. PFN may use generative AI to help produce Beta Content that illustrates possible PFN webpages, Domain Overviews, Solution Reports, comments and discussion forums, and other PFN projects and activities. The use of generative AI in developing Beta Content should be disclosed in the content’s Front Matter and Acknowledgements sections, as if the Beta Content were Shared Content. The Acknowledgements section should also tell users how to submit questions or concerns to PFN about the Beta Content.

Shared Content. For Shared Content, the disclosure of generative AI use depends on the context, including what did the AI do, why was it used, when in the development of the content was it used, and how important was its contribution to the final version. More disclosure than is required is always permitted.

For example, in some cases, it will be appropriate to state something along the lines of the ideas, thinking, and judgements in a Content Item were the responsibility of the named authors, the writing of the Item was the responsibility of the PFN Writer, and AI assisted in the early brainstorming and drafting and in the copyediting of the Item. In other cases, more detail may be appropriate. For example: “For this Content, [AI tool/model] was used to develop [ideas/research/analysis] and/or to [draft/organize/write/edit] in [this/that] section. Nevertheless, every factual claim and citation and the analysis were reviewed and verified by a person. The named authors supplied the ideas, thinking, and other substance.” PFN expects that what is appropriate in particular contexts will evolve.

Images and other visuals. PFN may publish images, infographics, charts, and other visuals produced with generative AI. In some cases, the ideas, thinking, data, and analysis reflected in the visuals may have been produced with generative AI, and in other cases the visuals themselves may have been produced with generative AI. Where generative AI has been used, the use should be disclosed in the Acknowledgements section and the visual itself should carry a short label identifying it as AI-generated.

Table: Disclosure of Generative AI Use

The following table provides more specific guidance about whether generative AI disclosure should be made in various content-related activities relating to the development of Shared Content. In general, disclosure should identify the tool or tools, the activity, and how generative AI was used. For example: “[AI tool/model] assisted in research and first-draft writing for this article. It was also used in copy-editing a near-final version of this article.” Where it would be important to users, the disclosure should also describe the prompts used.

For activities where disclosure is not required, PFN encourages the author to consider disclosure where it is reasonable to believe that the disclosure would be valuable to more than a minimal number of users. Some users expect or prefer disclosure in all or nearly all instances, so a Participant who chooses not to disclose should not be surprised by a later request to disclose.

ActivityGenerative AI use examplesDisclosureNotes
Ideation and brainstorming with authorsSuggesting possible topics, solutions, frameworks, questions, and issues for PFN consideration.Not required 
Research and content collection for authorsIdentifying possible sources; providing citation information; drafting annotations; organizing sources; monitoring new publications; identifying possible proponents, implementers, specialists, and reviewers to contact.Should be consideredCitations are verified against the primary source before they enter Shared Content.
Summarizing content, including legislative and regulatory content, for authorsSummarizing articles, studies, and reports; laws, regulations, bills, and proposed legislation and regulations; hearing, meeting, and presentation transcripts; comments on proposed legislation and regulations.Should be consideredShared analysis follows the drafting and writing row.
Analyzing and interpreting content for authorsMapping current statutes, regulations, and programs; comparative analysis; structuring tradeoffs; separating empirical disputes from normative disputes; testing solutions against long-horizon scenarios; stress tests such as premortems and assumptions checks.Required 
Discussions, deliberations, and debates with authorsDiscussing, comparing, and analyzing different positions; steelmanning or red-teaming topics and solutions; surfacing unexamined assumptions; identifying points of agreement, disagreement, and open questions.Should be consideredWhere the result becomes what the Content says a position or counterargument is, the row below applies instead.
Constructing a position, alternative solution, or counterargument for authorsBuilding the case for a position the authors do not hold; generating the counterarguments to a solution; producing an independent version of a position for the authors to compare with their own; supplying what proponents of a position would concede or refuse.RequiredThis is different from discussing a position with generative AI. Here the output shapes what the Content presents as someone else’s view, and a reader should know that.
Writing or producing content for authors, which the authors use to help write or produce content for end usersProducing first or early drafts of content.Required 
Editing and quality control of near-final draft contentEditing for neutrality, consistency, citations, grammar, style and usage, and other PFN guidelines. Screening for possible plagiarism.Should be considered for light editing; required where the editing changes the expression more than lightlyHeavy editing changes the words, which is the same question as drafting.
Checking source and reference citationsVerifying that cited sources and references exist, match PFN citation guidelines including URLs, and say what the content claims they say.RequiredAutomated checks do not replace verification by PFN or a Participant.
Producing content that is passed through to users, including text, graphs, tables, and other visualizationsWriting and checking statistical code; cleaning datasets; running exploratory analysis; generating charts, tables, and other visualizations for original data work.RequiredConsider more significant contributor acknowledgement. Visuals also carry a label.
Producing documents and files for users, but not the contentGenerating formatted Word, Excel, and PDF documents and files, including written text, graphics, and other visualizations.Not required 
Qualitative research supportDrafting interview protocols and survey instruments; transcribing recordings; coding transcripts and open-ended responses.Should be considered 
Producing translations, accessible versions, and other derivative formatsPreparing versions of PFN content in other languages; plain language and adjusted reading level versions; alternative text for figures; other accessibility and derivative formats.Required on the derivative versionThe original document governs when versions differ.
Public engagement and marketingPreparing podcast and video scripts, social media content, infographics, and interactive explainers; marketing PFN; informing possible Members about PFN’s vision, mission, standards, priorities, and practices.RequiredDescribe what AI drafted.
Fundraising and developmentProposal drafting; donor communications; grant reports.Should be consideredAI-drafted fundraising documents follow the drafting and writing row.
Events and conveningsPreparing event and conference abstracts and agendas; identifying possible speakers and participants; drafting panel questions, moderator briefs, and session summaries.Should be considered 
OperationsMeeting summaries; recurring research digests; correspondence.Should be consideredThis includes routine administrative correspondence. More substantive correspondence follows the drafting and writing row.

What Generative AI Output Is Not

Here are several limits or cautions about using generative AI based on what AI does not produce:

First, generative AI output is not evidence about people. A model can produce the arguments for a position, but they should not be treated as conclusions, but rather as leads requiring further research and verification. Also, AI cannot tell how many people hold a position, how strongly they hold it, or which version of it they hold. Statements in PFN Content about what people believe rest on sources about people, not on AI output.

Second, generative AI narrows the range of ideas across many Content Items. Research on AI-assisted ideas and writing has found that individual pieces are rated more creative while the set of them becomes more similar to each other. The study PFN relies on here is Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances 10, no. 28 (2024). That study was on short fiction written with AI-generated story ideas, not on policy writing, so PFN’s concern about Solution Reports is an inference from it. For PFN that risk can show up in Solution Reports that may tend to present a single flavor of alternative solutions, using similar words, in three different Domains. Working against this tendency requires greater human control and variance over the AI tools and prompts, keeping arguments grounded in what real proponents have stated, and occasionally comparing similar sections of different Solution Reports.

Third, generative AI tends to agree with the person who is using it. An AI model that has followed an argument through a long chat exchange will tend to help build that argument rather than test it. Asking the same AI model, in the same session, to find the weaknesses in what it just helped write is not a real test. Instead, a fresh session or using a different AI model, with the prompt material presented without the author’s framing, is closer to a worthwhile test.

Fourth, generative AI cannot reach all content and information, no matter how large its learning and training base. Paywalled articles, subscription databases, and recent developments may be missing, and an AI model rarely says so as it can’t really know what it doesn’t know. Thus, what an AI model generates may look complete when it is not.

Fifth, generative AI cannot be accountable: it cannot stand behind a claim, disclose a competing interest, answer a reader’s question, or say what the people who hold a position would actually accept. Accountability sits with the named authors. On fabricated and mismatched citations, see Content Verification below.

Content Verification

Disclosing that generative AI has been used to help produce statements, citations, and other content does not reduce the need for human verification. Verification, not disclosure, is what makes content reliable regardless of how it was made.

Thus, PFN requires that each statement and citation in Shared Content be verified by an author, reviewer, or contributor, no matter how the statement and citation were produced. Also, AI tools that perform detection, integrity, verification, and plagiarism checks may supplement human verification, but they do not replace it. See About PFN Content and Statement Types for the general verification standards for each PFN Content Type.

Generative AI can produce a citation to a source that does not exist, or to a real source that says something different from what is claimed. A citation that has come from generative AI should be checked against the primary source before it appears in any PFN content, including Exploratory Content shared within a Gated Audience (unless it is explicitly communicated that it has not been verified).

Plagiarism

Disclosing generative AI use does not allow or excuse intentional or unintentional plagiarism, or near-plagiarism such as close paraphrasing or patchwriting, of generative AI’s ideas or words. Just as an author or copyright owner of PFN content should not use another person’s ideas or words as if they were the author’s or copyright owner’s own without proper quotation, credit, and attribution, an author and copyright owner are expected not to use ideas or words generated by AI without proper quotation, credit, and attribution. This is so even where the Content properly discloses the use of generative AI (e.g., in the Acknowledgements section of the Content) and even though some may disagree with the notion that ideas and/or words should be quoted or attributed when they are generated by an AI tool (instead of by a human).

PFN authorship is byline credit, and it is based on the intellectual contributions (e.g., ideas and thinking) to the Content. Copyright ownership is a legal question about the words, images, and other visuals used to express and arrange the intellectual contributions. See the PFN Glossary.

It is PFN’s policy that if generative AI has made a significant intellectual contribution to a Content Item, that use should be disclosed in the Acknowledgements section, naming the tool or tools and describing what each contributed. If the contribution was substantial enough that a reader should know about it before reading, PFN expects that a Content Item will include a short line directly after the byline, linking to the Acknowledgements section. For example:

By Jane Smith, Robert Chen, and Maria Delgado

Generative AI contributed substantially to the analysis in this Report. See Acknowledgements.

As noted above, the authorship byline itself lists the Participants who made intellectual contributions and have accountability for the Content, and at least one human author is listed. But if every author has opted out of attribution, the byline will credit the responsible Network Group. The byline authors answer for the Content’s accuracy, disclose their competing interests, and stand behind its analysis, whether or not they agree with the Content’s ideas and conclusions. A generative AI system is not accountable in any way, so it is disclosed as a contributor rather than listed as an author.

Also, with respect to copyright, the law extends protection only to humans who express and arrange the intellectual substance of Content, so text that generative AI produces does not have a copyright owner. Indeed, the human who prompts an AI tool, however extensively, does not give the human copyright ownership in what the tool produces. Thus, what PFN owns or licenses in a Content Item for copyright purposes is the human-written expression in the Item, together with the human selection, arrangement, and editing of the material. This means that AI-generated text, visuals, and other expression do not have any copyright protection and thus are free for anyone to use whether or not PFN’s license says so.

PFN is not troubled by this outcome, as PFN has decided to publish its Content under a Creative Commons license that permits copying and adaptation anyway, so in most instances the practical difference is small. For copyright-type protection, PFN relies instead on its name, its standards, and its disclosures rather than on the scope of its technical copyright.

For additional discussions about PFN authorship and copyright ownership, see the PFN Content License, the Participant Agreement, and About Solution Reports.

People and Business Decisions

The PFN Leader, sometimes together with PFN Staff or other PFN Participants, makes all PFN Participant decisions, for both Members and Collaborators, at every stage of the process. Generative AI does not make any Participant decisions. This includes decisions about a Participant’s continued status, which may depend, for example, on whether the Participant has violated the PFN Code of Conduct or another PFN policy. Similarly, the PFN Leader and other PFN Participants, not generative AI, make all PFN decisions about arrangements with contractors and the hiring of employees. PFN may use generative AI to support those processes administratively.

Privacy and Security

PFN and its Participants are expected not to input private, sensitive, identifying, or contact information about Participants, staff, sources, or audience members into any generative AI tool. Participants should use the same care with generative AI tools that PFN’s confidentiality commitments generally require. Before sharing confidential or not-yet-shared PFN materials with a generative AI tool, including near-final drafts under review and other Exploratory Content, a Participant should confirm that the tool does not use the inputs to train its models and will keep the input information confidential.

For example, Participants are expected not to input Exploratory Content into any generative AI tool if the content attributes statements to identifiable Participants within a Gated Audience. And if PFN uses data and generative AI to personalize audience experiences on the PFN website, PFN will disclose how the data is collected and used, consistent with the Privacy Policy.

Policy Feedback and Governance

PFN Participants and others may give feedback on these generative AI policies, and on PFN or Participant activities that may raise questions under them, to the PFN Leader at policy.futures@policyfutures.net. The PFN Leader will coordinate with other PFN Participants as appropriate to review the feedback, and will periodically review PFN’s practices and these policies to consider updates. Violations are generally handled under the PFN Code of Conduct, which provides for a warning, suspension, or termination of Participant status after the Participant has had a reasonable opportunity to respond.

Also, PFN and its Participants should give complete and candid answers to questions about whether and how generative AI was used. The exact answers will depend on the relevant contexts.

PFN Content. About Beta Content · About Constructive Collaboration · About PFN Content and Statement Types · About Solution Reports · PFN Code of Conduct · PFN Content License · PFN Glossary · Privacy Policy · Privacy Policy · Terms of Use.

Acknowledgements

Contributors. Harry Conaway (PFN Leader) is the author and writer of this content and is solely responsible for its substance.

Disclosures. Claude (Opus 5.5, Anthropic) was used to assist in editing this content. All substantive and writing decisions were made by the author and writer noted in the Contributors section above.

PFN Content Type. This content is classified as PFN about and governing content, and thus is not subject to PFN’s Content and Statement Types policy for substantive policy content.

Version History

Current Version. To be added