Should a Renewal Email Say AI Helped? A Disclosure Sheet Built From Five Experiments

An association’s communications lead is asked whether the dues reminder and the spring appeal should say that AI helped write them. We read five experiments on how audiences react when they are told. As far as we could tell from what we read, they did not involve an association, its members or its donors.

The results do not agree, so the research cannot give a plain yes or no. Four of the five report a drop or a lost benefit after a disclosure: lower perceived accuracy and sharing for headlines, lower credibility for an advertisement, lower trust but not accuracy for news articles, and a fundraising gain that disappeared for letters. One large study of policy messages found no effect on attitudes, accuracy judgments or sharing, and several results turn on what the disclosure said. None of the five proposes the sheet below; it is our translation of their findings into questions.

A disclosure sheet for one kind of message at a time

Fill in one row per kind of message (the dues reminder, the newsletter, the appeal) and settle the middle column before drafting.

Decision	Settle first	Then do this
What the line says	What did the AI do, and which named person checked and approved the result?	Describe both in plain words rather than a bare “AI-generated” tag.
Whether money is asked	Does the message ask for a gift, a dues payment or a registration fee?	Treat it as the riskiest kind and try the line on a small slice of the list first.
What the message rests on	Can a reader see where the facts and figures came from?	Put the sources, or a page that lists them, next to the line.
Who reads it	Which readers trust the organization most, and which are older?	Look at the results for those groups separately.
How you will know	Can the list be split at random into two halves that differ only by the line?	Send both versions and compare replies, unsubscribes and gifts.

Say what the AI did, and who checked it

A PNAS Nexus article by University of Zurich political scientists reports two preregistered online experiments with 4,976 participants from the United States and United Kingdom. Its abstract says that labeling a headline as AI-generated lowered perceived accuracy and willingness to share, whether the headline was true or false and whether a human or an AI had made it, though the drop was three times smaller than a “false” label caused. The authors link the drop to readers assuming the AI worked with no human supervision.

That explanation is why the first row asks about the human. In the second experiment, the article told some participants that the AI had only improved a headline’s clarity and style, or had written a first draft, and the label then had no statistically significant effect. The participants were online panel members judging headlines, not members reading an association’s email, so this is a reason to test wording and not a promise that a plain description works.

Appeals for money are the riskiest rows

The abstract of a Scandinavian Journal of Economics article by Shandong University and Nanjing University of Finance and Economics researchers reports a charitable fundraising experiment in which participants received AI-generated or human-written letters, with or without a note of their origin. It says the letters raised donations by similar amounts, and that the effect disappeared when AI involvement was disclosed, whether the AI had generated or only edited the letter. The abstract does not say who took part.

A study in the International Journal of Advertising by researchers in South Korea and New Zealand found that 226 online panel members in the United States rated an advertisement with an AI-made image as less credible when it carried an AI disclosure than when it carried none. Its third experiment found smaller donation amounts after an AI credit than after a human photographer’s credit, but it compared two credits and not a credit with none, the donation was a stated choice about a possible $2 bonus that only 36 randomly chosen people received, and the authors name their online panels and AI-made images as limits.

Both studies suggest that a disclosure can cost a gift or its benefit, though nothing we read shows either tested an association’s donors. That is why the second row calls for a small first trial: a wrong guess costs the gift.

Show what the message rests on

The abstract of an International Journal of Press/Politics article by University of Minnesota and University of Oxford researchers reports a survey experiment with audiences in the United States that used real AI-generated news articles. It says articles labeled as AI-generated were rated less trustworthy but not less accurate or fair, that the effect was largely concentrated among people whose existing trust in news was higher, and that it was largely counteracted when an article disclosed the list of sources used to generate it.

News readers are not members, and a source list on a news article is not a data page linked from a renewal notice. We read the result as support for the third row, and as a reason for the fourth row’s worry that readers who already trust the organization most may be the ones who notice.

A label that changed nothing, and its limits

A PNAS Nexus article by Stanford University and Northeastern University researchers reports a survey experiment with 1,601 Americans who read an AI-generated message about public policies, labeled as from an expert AI model, labeled as from a human policy expert, or not labeled. According to the abstract, 92 percent of participants who saw a label believed it, yet the labels did not significantly change attitude change, judgments of accuracy or intention to share the message. In the full text, the authors report one subgroup result, that older participants reacted more negatively to the AI label than to the human one.

The same article lists reasons for caution. Its authors say the messages were deliberately evidence-based, that a small real difference may have gone unnoticed in their sample, and that the findings fit unattributed messages better than traditional institutional communication, which is closer to what an association sends. The study keeps the sheet from claiming that a disclosure always hurts, and it keeps anyone from claiming that it never does.

What a split of your own list can and cannot show

The last row is the only one that produces evidence about your members: pick one kind of message, split the list at random, change nothing but the line, and compare what readers do. A short list will not reveal a small difference, and the Stanford and Northeastern authors say even their 1,601 participants may have missed one, so a quiet result means no large effect was found, not that there is none.

The studies were published from 2024 to 2026, nothing we read says where in an email a line works best, and attitudes toward AI are moving, as the Stanford and Northeastern authors say of their 2024 data. Whether a law or a platform requires a notice is a separate matter from what readers prefer; our article on state AI-disclosure laws covers the legal side, and our guide to AI-drafted member emails covers the drafting side.

Bring the filled sheet to the board member who asked, and name who approves the wording and who reads the result.