The AI-authorship effect.
When people learn an emotional marketing message was written by AI, they react with moral disgust.
The same response is triggered by counterfeits, plagiarism, and dishonesty.
Emotional marketing message is communication that expresses feelings, moods, or emotions: gratitude, sympathy, happiness, pride, inspiration, care, or similar "we feel this" language.
Why should you care about it?
Because it causes a drop in positive word-of-mouth (PWOM) and customer loyalty (aka some people may not come back to buy from the brand again).
PWOM fell by 1.34 points on a 7-point scale, from 5.45 to 4.11, which is about a 24.6% relative drop. Customer loyalty fell by 0.83 points, from 5.85 to 5.02, which is about a 14.2% relative drop.
Kirk & Givi, 2025. The AI-authorship effect.
Please note that it is true only for an emotional marketing message, NOT a factual one.
When the penalty does not fully apply
Factual content (delivery confirmations, policy updates, schedules): no penalty.
AI editing human-written copy: much smaller penalty than AI writing it from scratch.
AI signing the message as itself ("From the brand's AI assistant..."): smaller penalty than a human signing AI text.
Templated content people already expect to be recycled: the penalty actually flips (so, the effect is positive). The human author takes the bigger authenticity hit. The AI looks more honest by default.
What to do about it
Keep humans writing anything emotional or identity-laden. Gratitude posts. Condolences. Anniversary emails. Brand stories. Founder messages.
Use AI for factual operations. Order updates, scheduling, FAQs, internal automation.
If AI has to touch emotional copy, let it edit a human draft. Not the other way around.
If AI authorship of emotional content has to be disclosed, sign as the AI itself. Do not let a human take the credit line.
Do not assume audiences can reliably spot AI content visually. They cannot. Disclosure, not detection, is what drives the response. Plan for the disclosure.
❗️Be aware
People cannot reliably tell which content is AI-written by looking at it.
The reaction appears when AI authorship becomes known, through disclosure, labels, leaks, or suspicion. It is not mainly driven by people spotting AI artefacts in the content.
The risk is not only whether the message was actually written by AI. The risk is whether people believe it was. For brands that rely on emotional value, craft, taste, intimacy, or founder voice, this is a serious risk. The message has to read as genuinely human, and ideally, it should be genuinely human.
Examples
Source: MrBeast, "Ages 1 - 100 Race For $250,000!", YouTube, 2026. Screenshots used for commentary and analysis.
Brüns and Meißner (2024) use Mr. Beast in their paper as the canonical case of authenticity built on visible, sustained human dedication. He has spent years publicly framing himself as someone who thinks about content production constantly, every day. The audience follows partly because of that signal. Extreme effort put into the production is the brand.
The thought experiment is simple: if he switched to AI-generated scripts, removed most of the human side of it, and that switch became known, followers would update their authenticity perception even if the views held up at first.
The case generalizes to any creator whose brand is built on visible human obsession. Solo founders. Coaches. Newsletter writers. The format with the highest disclosure risk is the one with the most personal voice.
Why it works 💡
Emotional content is, structurally, a claim about an inner state.
"I'm so grateful." "It is with a heavy heart." "We're proud to announce." The reader is meant to take it as a window onto how the speaker actually feels.
People believe AI does not have an inner state.
It cannot actually feel, care, grieve, or admire. Mind-perception research has shown this consistently for years. Even people who use AI tools daily hold this intuition. They like the output and still believe nothing is feeling on the other side.
So an AI-written emotional message is, by construction, a misrepresentation.
The brand is putting feeling-words in the mouth of something that has no feelings. The issue is sincerity. Quality stays the same. The text can be excellent and still fail this test.
Misrepresentation triggers moral disgust.
The same emotion attached to fakes, frauds, and lies. Kirk and Givi measure it directly with a moral-disgust scale and observe it spike across seven studies.
Moral disgust drives avoidance.
Less recommendation. Less loyalty. More distance from the brand. Kirk and Givi back this with a behavioural measure on top of the self-report. Participants in the AI-author condition spent fewer real incentive points to promote the brand. The effect shows up in what people do, in addition to what they say.
The core step is the link from 2 to 3. This is also why factual AI content gets a pass. A delivery confirmation does not claim feeling, so there is no misrepresentation, so there is no disgust.
The brand-level finding has similar logic. Following a brand is identity-driven consumption. Brand authenticity is built on signals of passion, continuity, and craft. GenAI adoption breaks all three at once. It signals the brand has stopped putting in the human effort the audience originally followed it for.
Watch Out 🚨
The most important caveat is that the ground is moving fast. AI adoption is getting into... everywhere. Both papers were run in 2023 and early 2024. As AI authorship becomes the default in marketing, some of what these studies measure will shift.
What probably will fade
The novelty layer. Right now, AI in brand content is unfamiliar, and unfamiliarity makes the violation feel sharper. Kirk and Givi Study (Web Appendix Study 1) tests a version of this: when participants are primed to believe most marketing is AI-written, the headline penalty attenuates. The shock layer will not be permanent. If the same studies are repeated in 2027 or 2028, the AI-authorship penalty will probably be smaller.
What should not fade
The structural layer. The mechanism rests on a mismatch between what emotional language claims and what AI actually is. Emotional language claims an inner state. AI does not have one. That mismatch is not a 2024 phenomenon. It is a permanent feature of what emotional language is and what AI is. Universal AI adoption does not change the structural fact that "I'm so grateful" is a claim about feeling.
Research context 📇
▼Click to reveal a study-by-study breakdown
Two papers that explore the effects.
Kirk and Givi (2025), Journal of Business Research – message level
Seven preregistered experiments. Roughly 2,300 US and UK Prolific participants. Five emotions tested (inspiration, happiness, gratitude, sympathy, pride). Both individual and mass communications.
What each study did:
-
Study 1 (n = 300). An inspirational follow-up email from a salesperson. AI authorship dropped positive word-of-mouth from 5.45 to 4.11 on a 7-point scale. Loyalty dropped from 5.85 to 5.02. Moral disgust roughly doubled (1.85 to 2.93). Behavioural measure on top of the self-report: a custom review website where participants could spend incentive points to promote a positive review. The AI condition bought fewer promotional votes.
-
Study 2 (n = 400). Same setup. Two versions: emotional and factual. Penalty appeared only for emotional content. Factual content showed no significant difference between AI and human authorship.
-
Study 3 (n = 400). A condolences email. Four conditions: human-written, AI-written, AI-written then human-edited, human-written then AI-edited. AI editing human copy: safe path. AI writing copy that a human signs: same penalty as straight AI authorship.
-
Study 4 (n = 300). Three signing conditions: human signs human-written; human signs AI-written; AI signs AI-written. AI-signed messages take a smaller hit than AI-written messages signed by a human. The deception layer amplifies the disgust.
-
Study 5 (n = 200). Confirms the full causal chain explicitly: authorship → perceived authenticity → moral disgust → word-of-mouth and loyalty.
-
Study 6 (n = 600). The reversal. When emotional content is openly copied or reused, the human author takes a bigger authenticity penalty than the AI. People attribute intent to humans (a deliberate reluctance to write something fresh). They do not attribute intent to AI in the same way.
-
Study WA1. When people are primed to believe most marketing is AI-written, the penalty largely disappears. The headline finding is partly a function of current expectations.
Brüns and Meißner (2024), Journal of Retailing and Consumer Services – brand level
Three experiments. Roughly 420 Prolific participants. Brand level rather than message level.
What each study did:
-
Study 1 (n = 118). Participants imagined their favourite real brand on social media announcing either full GenAI automation or no GenAI use. Automation dropped perceived brand authenticity from 5.82 to 4.42. Post credibility, EWOM (electronic word-of-mouth), and brand loyalty dropped proportionally. Brand authenticity mediated all three.
-
Pre-study to Study 2 (n = 131). A diagnostic check, not a hypothesis test. Participants saw either a real fashion-brand photoshoot or an AI replication built with ChatGPT and Midjourney. They could not distinguish them on perceived passion, professionalism, or brand fit. 67% thought the real image was AI. 58% thought the AI image was AI. The penalty is triggered by disclosure, not by spotting AI artefacts.
-
Study 2 (n = 111). A fictitious brand. Concrete AI-generated Instagram post with disclosure manipulated. Same pattern: disclosure drops authenticity, credibility, and brand attitudes. EWOM did not move – likely because there is no prior relationship with a fictitious brand.
-
Study 3 (n = 190). Three conditions on participants' favourite real influencers: no adoption, AI assists human, AI fully automates. Clean linear pattern: no adoption > assistance > automation. Assistance still hurts. Just less.
The two papers agree on three things, which is what makes the combined claim worth acting on:
→ Perceived authenticity is the mediator.
→ Disclosure, not visual detection, triggers the response.
→ AI assists humans is genuinely different from AI replaces humans.
Thoughts from the Author 💭
The EU AI Act changes the disclosure question.
Article 50 of the EU AI Act becomes fully enforceable on 02.08.2026.
Providers of generative AI systems must ensure outputs are marked in a machine-readable format.
Deployers, that includes brands using AI in marketing, must disclose AI-generated or manipulated image, audio, and video that could be mistaken for authentic content (the "deepfake" category).
Public-interest text published without human editorial oversight also falls in scope.
Carve-outs exist but are narrowing. AI used for minor editing, spelling and grammar that does not substantially alter the input does not require disclosure. Content under documented human editorial responsibility may be exempt for the public-interest text rule. The European Commission's draft Code of Practice (first draft published December 2025, final expected around June 2026) narrows the human-review exception by requiring an actual documented editorial workflow with identified responsible persons. The mere claim that a human looked at it will not qualify.
What this means for marketing in practice. Even if a brand decides not to label its AI-generated content voluntarily, the watermark is at the model-provider level. Platforms and detection tools will flag it. The audience sees the label. The brand ends up disclosed whether it wants to or not.
For the nerds 🤓
Figure. Mediation analysis from Brüns and Meißner (2024), Study 1, redrawn with the indirect-effect math worked through.
▼Click to reveal the nerdy deep-dive
Brüns and Meißner make a specific causal claim with the Study 1 mediation analysis – GenAI adoption does not directly hurt follower reactions in any meaningful way. It hurts perceived brand authenticity, and authenticity is what hurts follower reactions (PWOM, EWOM, loyalty, brand credibility, etc.).
Reading the paths
Path A. GenAI Adoption (Automation vs. No Adoption) → affects Brand Authenticity. Coefficient −1.40, p < .001. The size of the authenticity hit when a brand switches from no GenAI to full automation.
Path B. Brand Authenticity → affects Follower Reactions. Coefficients 0.82 (Post Credibility), 0.77 (EWOM Intentions), 0.73 (Brand Loyalty). All p < .001. How much each follower outcome moves per unit of authenticity.
Path C (direct). GenAI Adoption → affects Follower Reactions, controlling for authenticity. Coefficients −0.24 (p < .1), −0.44 (p < .01), −0.12 (n.s.). Once you control for authenticity, GenAI adoption has very little independent impact. Almost all of the action is going through Brand Authenticity.
Why the mediator chain matters in practice
Authenticity (indirect effect) is doing nearly all the work between adopting GenAI and the metrics a brand actually cares about. The direct effects are almost-zero. If a brand could keep authenticity perception intact while using GenAI, the negative follower reactions would mostly disappear. That is exactly what Study 3 shows when an "AI assists humans" condition is added – authenticity holds up better, and credibility, EWOM, and loyalty hold up with it. The intervention point is the authenticity perception itself.
Sample size for the path coefficients above: n = 118 (Brüns and Meißner Study 1). Bootstrapped 95% confidence intervals on the indirect effects: Post Credibility [−1.59, −0.69], EWOM Intentions [−1.55, −0.63], Brand Loyalty [−1.50, −0.58]. None cross zero. The mediation is robust.Share with your network
Sources
Academic papers
-
Brüns, J. D., & Meißner, M. (2024). Do you create your content yourself? Using generative artificial intelligence for social media content creation diminishes perceived brand authenticity. Journal of Retailing and Consumer Services, 79, 103790.
-
Kirk, C. P., & Givi, J. (2025). The AI-authorship effect: Understanding authenticity, moral disgust, and consumer responses to AI-generated marketing communications. Journal of Business Research, 186, 114984.
-
Castelo, N., Boegershausen, J., Hildebrand, C., & Henkel, A. P. (2023). Understanding and improving consumer reactions to service bots. Journal of Consumer Research, 50(4), 848–863.
-
Castelo, N., Bos, M. W., & Lehmann, D. R. (2019). Task-dependent algorithm aversion. Journal of Marketing Research, 56(5), 809–825.
-
Leung, E., Paolacci, G., & Puntoni, S. (2018). Man versus machine: Resisting automation in identity-based consumer behavior. Journal of Marketing Research, 55(6), 818–831.
Regulations