How to tell if a wedding speech was written by AI
Twelve specific tells, with the bad version and the fix for each. This is the same list our own generator is instructed to avoid, so it is a product spec as much as a guide.
Nobody at a wedding minds that you had help writing the speech. What they notice is a speech that could have been about anybody. The tells below are what make it feel that way, and most of them are structural rather than a matter of vocabulary, which is why running a draft through a word blocklist does not fix it.
Read your draft aloud with this list beside you. Anything you would not actually say in a room full of people who know you, cut.
01 The em-dash
The single most reliable signal. Large language models reach for the em-dash constantly; most people writing a speech on their phone never type one, because it is awkward to reach on a British keyboard.
Sounds like AI: James is loyal — fiercely, unshakeably loyal — and Sarah knows it.
Better: James is loyal. Fiercely, unshakeably loyal, and Sarah knows it.
02 The rule of three, everywhere
One tricolon in a speech is rhetoric. Four is a machine. AI writing stacks three-part lists because they feel balanced, and the effect across a whole speech is oddly weightless.
Sounds like AI: He is kind, funny, and unfailingly loyal.
Better: He irons his socks. Nobody has ever asked him to.
03 "Not just X, but Y"
The balanced-antithesis construction. Once you notice it you cannot stop noticing it, and it appears in almost every machine-written toast.
Sounds like AI: Sarah did not just gain a husband, she gained a best friend.
Better: Sarah has taken on a man who irons his socks and proposes in monsoons.
04 Summarising the emotion instead of showing it
AI tells the room what it just felt. A person shows the thing that happened and lets the room work it out.
Sounds like AI: It was a truly beautiful and touching moment.
Better: He never once mentioned it. Not to me, not to anyone.
05 Adjective pairs
Warm and welcoming. Kind and generous. Two adjectives doing the work one specific fact would do better.
Sounds like AI: Sarah is warm and welcoming to everyone she meets.
Better: Sarah sends a card for everything, including my cat’s birthday.
06 A neat aphorism closing every paragraph
The instinct to land each section on a tidy universal truth. Real speech is lumpier: some paragraphs just stop.
Sounds like AI: And really, that is what love is all about.
Better: That is the whole story. No punchline. He just did that.
07 Signposting
"Firstly." "What is more." "But here is the thing." Written scaffolding that a speaker would never say out loud.
Sounds like AI: Firstly, let me say what an honour it is to be here today.
Better: Start on the detail. Get in late and skip the throat-clearing.
08 The rhetorical question opener
"What can I say about James?" is the most common opening line in machine-written speeches and it delays the actual start by a sentence.
Sounds like AI: What can I say about James that has not already been said?
Better: The first time I saw James he was wearing a shirt with actual flames on it.
09 Uniform sentence length
Human speech is uneven. Some sentences run on because the speaker is enjoying themselves. Then stop. AI output tends to sit at a steady fifteen to twenty words a sentence throughout.
Sounds like AI: A paragraph where every sentence is the same comfortable length as the one before it.
Better: Vary it deliberately. When a line matters, make it short.
10 The vocabulary list
Certain words appear far more often in machine text than in British speech: tapestry, journey, navigate, testament, delve, whirlwind, better half, partner in crime, tie the knot, happily ever after, words cannot express.
Sounds like AI: Their love story is a testament to the beautiful tapestry of life.
Better: Cut the lot. None of them survives being said aloud in a marquee.
11 American register in a British room
Mom, buddy, bachelor party, gotten. Also the sentimentality: British speeches undercut a feeling immediately after landing it, and AI trained largely on American text tends to escalate instead.
Sounds like AI: I love you, man. You are my rock and my inspiration.
Better: He is not the worst. From a British best man, that is high praise.
12 Invented specifics
The most dangerous tell, and the one people miss. A model short of detail will quietly manufacture one: a conversation that never happened, a relative who was not there, a colour, a date. At a wedding you then say something untrue about your own friends, out loud, in front of them.
Sounds like AI: I only found out because his mum let it slip at Christmas.
Better: If you were not told it, do not say it. Go deeper into a detail you do have instead.
The one that actually matters
Eleven of these are style. The twelfth is not. A model that is short of detail will invent one, and it will do it fluently: a conversation nobody had, a relative who was not in the room, a date, a colour, a count. Style tells make a speech forgettable. An invented detail makes you stand up and state something untrue about your own friends, in front of those friends, on the one day it matters.
If you take one thing from this page, make it that check. Go through your draft and ask of every name, number, quote and place: did I actually supply this? If not, cut it, or hedge it out loud the way a person would.
You can see what that looks like in practice. We publish eight complete speeches, unedited, where every detail traces back to something the speaker typed in.
Questions people ask
How can you tell if a wedding speech was written by AI?
Look for the em-dash first, then the structural habits: three-part lists used repeatedly, the "not just X, but Y" construction, adjective pairs like "kind and generous", and a neat universal truth closing every paragraph. Vocabulary tells such as tapestry, journey, testament and whirlwind are easy to catch. The most serious sign is an invented specific: a quoted conversation, a bystander or a date that the speaker never actually supplied.
Does using AI to write a wedding speech matter?
Nobody minds that you had help. What people notice is a speech that could have been about anybody. The failure is genericness and, worse, invented detail: if the speech states something untrue about the couple or their family, you say it out loud in front of them. Using AI is fine. Reading out a fact you did not know was fabricated is not.
Why do AI speeches use so many em-dashes?
Large language models are trained heavily on edited published prose, where the em-dash is common. Most people writing a speech on a phone or laptop never type one, because it is not on a British keyboard without a shortcut. That mismatch makes it the single most reliable giveaway.
How do I make an AI-written speech sound like me?
Feed it specifics rather than adjectives. One real thing that happened, with names and places, is worth more than a paragraph about how generous somebody is. Then read the result aloud: anything you would not actually say in a room, cut. Short sentences carry, long ones collapse.
Or let something write it that already knows all this.
Tell us one story. We write your opening lines free, before you pay.
Start free preview