When AI-Written Messages Backfire: A 2026 Guide
New research shows AI-written emotional messages can cut brand recommendation by 24.6%. When to use AI in writing + when to keep it human.
One of the more useful research findings of 2026 is a study showing that audiences punish brands more than people expected when they discover AI wrote an emotional message. The size of the effect is large (~25% drop in brand recommendation), consistent across six different experiments, and concentrated in exactly the contexts where marketers most want to use AI: apologies, founder notes, announcements + condolences.
Below: what the research actually found, why the effect happens, and a practical rule for when to use AI in writing + when not to.
What the research found
The six experiments asked participants to read various kinds of marketing communications + rate how likely they were to recommend the brand. Some participants were told the message was AI-written; others were told it was human-written. Three findings stand out:
- Emotional messages take the biggest hit. Apologies, announcements (especially layoffs / restructures), founder reflections, condolences. Brand recommendation dropped by up to 24.6% when these were labelled AI-written.
- The reaction isn't just disappointment - it's moral. Self-reported 'moral disgust' went up by 58.4% for AI-written emotional content. Readers don't just feel let down; they feel actively misled.
- The deception framing matters most. Presenting AI-written copy as if it came from a real person is what triggers the strongest negative reaction. Acknowledging AI use up front softens it considerably.
For utility content (product descriptions, technical docs, troubleshooting articles, summaries), the effect essentially disappears. Readers don't expect emotional labour from a 'how to reset your password' guide - they want it to be accurate + clear, and the source isn't a moral issue.
Why the reaction is so strong
Two related mechanisms drive it:
- The implicit promise of emotional content is personal effort. When a founder writes a heartfelt note about a difficult decision, the implied value is 'I cared enough to think this through myself'. Outsourcing that to a model breaks the promise - whether or not the words themselves are good.
- Discovery feels like being conned. Readers can't reliably tell AI-written text from human-written text in advance, but once they're told, the trust isn't recoverable in that moment. The brand has essentially admitted that the warmth they were trying to project wasn't real.
The research authors frame this as 'authenticity is a scarce resource' - readers have learned to expect that emotional communication takes someone's actual time + thought, and AI shortcuts that expectation in a way that lands worse than no message at all.
A practical rule for using AI in writing
The takeaway is NOT 'don't use AI for writing'. It's 'match the tool to the type of writing'. A working rule for 2026:
- Utility content - use AI freely. Email drafts you'll heavily edit, technical documentation, product descriptions, internal summaries, translation/localisation, code-related writing. AI is the right tool + audiences don't care.
- Relationship content - use AI lightly, if at all. Customer apologies, founder notes, public announcements about people-affecting decisions (layoffs, policy changes), condolence messages, personal thank-yous. If you use AI here, treat it as a research assistant for outlines or word-finding - never paste the AI output verbatim.
- Mixed cases - acknowledge AI involvement. If you genuinely used AI to draft a piece and the topic is borderline emotional, mention it up front ('drafted with ChatGPT, edited by me'). The research shows transparency softens the negative reaction substantially.
What about AI as an editor?
One real challenge: AI is genuinely useful as a polishing tool. It can tighten prose, fix typos, suggest clearer phrasings, vary sentence length. A second piece of writing research from 2026 ('AI Writes Brilliantly') points out a different problem: when you ask AI to suggest multiple strong versions of a passage, you get many fluent options - but combining them into a final draft tends to flatten the writer's distinct voice, even when each individual edit was an improvement.
The practical countermeasure if you want AI as an editor without losing voice: edit your own draft first, then ask the AI for specific localised improvements ('tighten this paragraph', 'find a better word here'), not broad rewrites. The more the AI is involved in the voice-defining decisions, the more your final piece drifts toward generic AI prose - which audiences are increasingly good at recognising.
The bottom line
AI is great for writing that needs to be clear and correct. It's bad for writing that needs to feel personal or earned. Brand-trust damage from getting that line wrong is measurable + large in 2026 - the 24.6% recommendation drop is the kind of number that should make any marketer second-guess running a customer apology through ChatGPT verbatim.
The simple test: if someone asked you 'who wrote this and how long did it take', would the honest answer make you uncomfortable? If yes, that piece probably shouldn't be AI-written. Most emails, summaries, and technical writing pass the test fine. Apologies + founder notes + emotional announcements don't.