What Studies Say About AI Written Email, Trust and Engagement

The study research on AI generated email marketing, engagement and trust comes down to five original sources we could open and read. One field experiment tracked purchases from AI-written email that customers did not appear to notice. Two sets of controlled experiments measured what happens once people believe, or are told, that AI was involved. Two vendor surveys asked consumers how they think they react. Our reading is that the penalty follows the reader's belief about who wrote the message and the kind of message it is, and that we found no published number for how much AI is too much.

Who was studied, when, and what was measured?

Studies that bear on generative AI in email marketing, measuring consumer trust or engagement with AI generated content in emails or in marketing messages generally, use three designs, and the design decides how far each result travels. The field trial measured purchases on a live list, the controlled experiments isolate a cause, and the surveys record what people say they do.

What happened to sales when customers noticed nothing unusual?

The strongest behavioral evidence is a set of field trials at online wine retailer Wine Access, written up by Chicago Booth Review on July 7, 2026. Jean-Pierre Dubé and Ariel Xu ran three randomized controlled trials over a year, each lasting two weeks and drawing on the 27,500 customers subscribed to the company's twice-daily newsletter. Customers were assigned at random to no newsletter, emails from the three staff writers, emails generated by AI, or emails drafted by AI and edited by a person. Two trials used a custom-trained model, and the third used Claude with a simple prompt.

The researchers tracked purchase probability, bottles sold, revenue and gross profit. Per the Booth write-up, profits in the three emailed groups far outpaced the control group and were "generally indistinguishable from one another," and customers submitted no complaints or inquiries. In the third trial, where the researchers prompted Claude, the AI copy "significantly edged out staff writers on gross profits." The article puts the three writers' combined salary and benefits at $375,000 a year, useful context for what email marketing costs.

Two limits matter. This is one company's recurring newsletter spotlighting a single bottle, which we read as closer to factual product copy than to a personal letter. The write-up reports purchases only, and we found no open rate, click rate, unsubscribe rate or trust measure in it.

What changes once readers believe, or are told, that AI wrote it?

On disclosure, the studies that compare AI generated and human written content for consumer trust and engagement point one way, though neither abstract names marketing emails as the test material. Colleen Kirk and Julian Givi's Journal of Business Research paper is the nearest to marketing. Its abstract on IDEAS/RePEc reports seven preregistered experiments showing that when consumers believe emotional marketing communications were written by an AI, positive word of mouth and customer loyalty fall. The effect weakens for factual messages, when an AI only edits, when the message is signed directly by an AI, and when consumers believe most marketing communications are AI-written.

Oliver Schilke and Martin Reimann tested disclosure itself. The University of Arizona describes 13 experiments with more than 5,000 participants in which revealing AI use lowered trust; investors, for one, trusted firms 18% less when ads disclosed AI use. Gentler wording, such as saying AI was only used for proofreading or that a human reviewed the output, still lowered trust. The paper's abstract, also on IDEAS/RePEc, adds that disclosing cost less than being exposed by a third party.

The papers pull apart on one point: Kirk and Givi find a smaller effect when AI only edits, while Schilke and Reimann find a trust cost even when the disclosure says AI only proofread. They measured different outcomes in different settings, and our combined reading is that a lighter AI role helps while a label still carries a price.

What do the consumer surveys add?

Both surveys are self-reported. Adobe Express surveyed 1,007 US consumers in December 2025: 30% were not confident they could tell an AI-written email from a human one, 18% said they had unsubscribed from a marketing email because they suspected AI wrote it, and 46% said they would be more likely to unsubscribe if they knew an email was clearly written by AI. Against that, 37% said they didn't mind an AI-sounding email as long as the content was useful and relevant.

Validity, in a post dated September 5, 2025, says it asked over 1,000 consumers about marketing emails they knew were written by AI, and two in five were "somewhat less" or "much less" likely to trust them.

With that many readers unsure they can tell the difference, the Adobe Express figure for past unsubscribes cannot say whether the emails those readers left were machine-written. Treat these numbers like any other email marketing statistics: good for sizing a concern, weak for predicting your own list.

Has any study published a threshold?

We found no study that publishes a threshold for AI generated email content, meaning a share of machine-written copy above which engagement or trust falls. None of the five sources varies the proportion of AI text inside an email and reports the response at each level. The Wine Access trials come nearest, with three levels of involvement (staff-written, AI-drafted then human-edited, fully AI-generated) and profits that were generally indistinguishable across them, apart from the Claude trial, where AI copy edged out staff copy on gross profit. Until one is published, an article that quotes a safe percentage is offering an opinion.

How much AI-generated content can an email carry before engagement and trust drop?

How much AI-generated content an email can carry before engagement and trust drop is a question the published work answers with conditions in place of a percentage:

• What the reader believes. AI copy that customers did not appear to notice matched or beat staff copy on profit at Wine Access, while believed or disclosed AI authorship lowered loyalty and trust in the controlled experiments.
• What kind of message it is. Kirk and Givi tie their effect to emotional messages and report that it weakens for factual ones.
• What the AI did. The same abstract reports a weaker effect when an AI only edits.

Spam placement is a separate worry. Validity wrote on November 14, 2024: "There's no evidence to suggest that AI-generated content is more likely to be marked as spam by filters or by recipients." The post names a bad sender reputation and missing authentication as far likelier causes. This is one vendor's view from a dated post, so unsubscribe and complaint rates from any AI copy test still belong beside the basics in our explainer on email deliverability.

How much AI in email is too much for your own list?

How much AI in email is too much for trust and engagement is, on this evidence, something only your own list can settle, because the published results turn on message type and reader belief more than on volume. Our position is to keep a named human writing the emotional, personal-voice sends, use AI freely on factual ones, and test everything in between.

A worked example, with assumed numbers throughout. A retailer with 40,000 engaged subscribers borrows the Wine Access design and splits the list at random into four cells of 10,000: a holdout that gets no campaign email, staff-written copy, AI-drafted copy that a person edits, and AI copy sent as generated. Subject line, offer, sender name and send time are identical across the emailed cells. The test runs two weeks, the length Dubé and Xu used, and the outcomes are written down before the first send: revenue per recipient, click rate, and the share of each cell that unsubscribes or reports spam.

Then check that the list is big enough. Assume 1.0% of the staff-copy cell unsubscribes over the two weeks and the team wants to detect a rise to 1.5%. A standard two-proportion sample-size calculation at 95% confidence and 80% power is 7.84 times (0.0099 + 0.014775), divided by 0.000025, or about 7,700 recipients per cell, so 10,000 is enough. Detecting a rise from 1.0% to 1.2% would take about 42,600 per cell, more than this list holds.

A team that plans to label AI-written email can add a fifth cell carrying the same AI copy plus the disclosure line; five cells of 8,000 would still clear the 7,700 needed here. Repeat the test on a second campaign type before setting policy, because one company's wine newsletter tells us about wine newsletters. Legal disclosure duties are a separate question that this page does not cover.

Our email deliverability checker and lifecycle email playbook are further resources for the groundwork around a test like this, and our lifecycle and demand generation team is there for teams that want help designing one.

Sources

• Chicago Booth Review: AI Is Coming for Marketing Copywriters, July 7, 2026
• IDEAS/RePEc: Kirk and Givi, The AI-authorship effect, 2025
• IDEAS/RePEc: Schilke and Reimann, The transparency dilemma, 2025
• University of Arizona: Disclosing AI use can backfire, research shows, May 22, 2025
• Adobe Express: Why a "human touch" still matters in AI-written emails, February 19, 2026
• Validity: Smarter, Not Sneakier, September 5, 2025
• Validity: The Future of AI in Email, November 14, 2024