Wedding Vows? Nope. Here's the AI-Written Message That Hurts the Most.

You might think your wedding vows will be the most important thing you ever write. There's an audience, someone is probably crying, and sometimes there's an alarming amount of money riding on them. But when it comes to words people most resent being handed over to AI, something that comes a bit later in marriage takes the cake.

In a survey of 2,000 US adults, 71.3% would feel genuinely betrayed by an AI-written apology, compared with 36.1% for wedding vows. Apparently, getting help with "I do" is more accepted than getting help with "I'm sorry."

Being offended by an AI-written apology turns out to be easier than recognizing one. While 65% feel confident they could spot AI behind a heartfelt message, accuracy across five pairs of human and AI writing reaches just 54.8%, barely better than a coin flip.

Key Takeaways:

• Serious apologies are the thing 71.3% of people say they'd feel betrayed by if AI wrote it, more than any other type of writing
• 65% are confident they can spot AI writing. But when tested, people averaged just 55% accuracy.
• Majorities reject AI-written eulogies, love letters, claimed fiction, dating openers, and apologies, even first drafts.
• Some 12.0% report discovering for certain that a personal message they received was AI-generated.
• 23.5% have secretly used AI for personal writing, while 8.3% were wrongly accused of it.

An Apology Has More to Prove Than a Wedding Vow

An apology asks someone to believe that the person who hurt them has given it some thought. Discovering that AI handled the wording leaves an uncomfortable question hanging over even the most eloquent expression of remorse: how much thought went into it before the prompt?

Among the 16 kinds of writing considered, serious apologies provoke the strongest sense of betrayal, at 71.3%. A novel someone claimed to have written themselves follows at 63.6%, ahead of a love letter at 59.9% and a eulogy at 54%. A performance review joins them at 53.2%, bringing a distinctly office-shaped grievance into otherwise intimate company. A manager may expect employees to put care into their work; employees have expectations about the written assessment of it, too.

Wedding vows have all the ingredients for a first-place finish: lifelong promises, an audience, and relatives who will remember the wording for decades. Yet at 36.1%, they finish ninth, below a best-man or maid-of-honor toast at 39.8% and even a thank-you note at 38.6%. The promise to cherish someone forever gets more latitude than the note thanking them for a casserole. Holiday cards, which are usually generically written, attract less scrutiny still, at 22%.

People are more forgiving when the writing has a practical job to do. An AI-written complaint letter would leave 18.4% feeling genuinely betrayed, with recommendation letters at 18.3% and dating profiles at 17.9%.

Some Messages Should Start With a Blank Page

There is quite a bit of distance between "that would hurt" and "AI has no business writing it in the first place." A separate question drew the line in the sand, asking which kinds of writing should remain entirely human, even through the awkward and sometimes perplexing first draft, when there was no polished opening to borrow and no rough version to tidy up: just the writer and an ominously blinking cursor.

Five kinds of writing cross that line for a majority: a eulogy at 58.9%, a love letter at 57%, a story or novel claimed as someone's own at 55.4%, a first dating message at 54.6%, and a serious apology at 53.4%. Even a draft that would later be edited is too much assistance for these respondents.

Only 16.1% say nothing is off-limits. For everyone else, there is at least one occasion when finding the right words is an inescapable part of the job.

Everyone's an AI Detective Until the Test Starts

To see how careful would-be wordsmiths need to be about passing off AI's prose as their own, the survey included a test of whether readers can tell the difference. Five matched pairs place a human-written passage alongside an AI-written version, with examples ranging from an apology to a wedding vow. In each pair, the goal is to identify which passage is written by a human.

In the results, the average person got 2.7 correct answers out of five, with overall accuracy of 54.8%, barely above the 50% expected from a coin flip. Readers may have firm rules about when AI should stay out of a message, though catching it in the act proves considerably harder. Readers can test their own AI Detective skills with the quiz below.

Confidence among respondents was plentiful, but correct answers were harder to come by. Some 65% felt confident they could spot AI in a heartfelt message, including 37.1% who are very confident. Yet among those very confident respondents, 39.6% scored at or below chance.

The apology fooled 47.3%, making it the toughest pair of passages to tell apart. Wedding vows offered slightly better odds, but still fooled 43.3%. The message people most insist should come from the heart turns out to be particularly good at concealing where it actually came from.

Not Even Humans Are Above Suspicion

In the responses, 12% report discovering for certain that a personal message they received was AI-generated. Another 19.5% strongly suspect it, though suspicion and confirmation remain two separate things.

Among confirmed cases, 32.2% say they kept quiet because it really did not bother them. For 24.7%, it mattered enough to pull back from or end the relationship. This means AI-assisted messages can save writing time, but that comes at the considerable expense of any future correspondence.

On the other side of the spectrum, 23.5% say they have used AI for personal writing without disclosing it. Meanwhile, 8.3% report being wrongly accused of using AI, and 20.3% worry their own authentic writing sounds like AI wrote it.

Spellcheck to One Person, Betrayal to Another

There is no agreed etiquette for letting AI help with the heartfelt stuff. For 22.1%, it is a tool like spellcheck, a little assistance getting the words right. Another 18.6% consider it dishonest regardless of the circumstances, while 22.5% accept it only if the writer discloses the help. A sender can therefore regard a message as carefully polished while the recipient regards it as grounds for a very different conversation.

The sensible approach is to consider what the other person expects from the message. AI might help untangle a clumsy sentence, but if discovering its involvement would undo the goodwill the words are meant to create, that is a reason to leave it out or be open about the help. A useful test is whether explaining the writing process would require an apology. Sometimes the awkward sentence is worth keeping if it sounds like the person who actually means it.

Methodology

In September 2026, Machined.ai conducted a survey of 2,000 US adults, census-balanced on age, gender, and region. Measures include a 16-item betrayal matrix, a separate never-acceptable measure, and a five-pair forced-choice detection test using matched human and AI passages. The findings from the survey form the basis of this article.

Fair Use

Want to share what you've found here? Go for it. You can use our data and graphics for non-commercial projects; just add a link to this page as attribution when you do so.