# Write a speech, toast or eulogy with AI that still sounds like you

Gather your own stories, ask for structure rather than a finished speech, have the assistant flag clichés and invented details, then check every name and read it aloud against the clock.

Workflows · v1.0 · draft · https://hayzlabs.com/workflows/write-a-speech-with-ai/

In short:
- Gather your own stories, names and dates first, ideally by talking, then ask for structure and options, not a finished speech.
- Give it a sample of how you talk, ask it to cut clichés and flag any detail it added, and check every name and story yourself.
- Read it aloud and time it. For a eulogy, paste only what the family would be comfortable with, and let them read it first.

**Use it when** you have to speak about a person: a wedding toast, a retirement or leaving speech, a birthday or a eulogy. You have the stories. You want help with order, length and the hard parts.

**Not when** you want the assistant to supply the memories. It never met the person, and stories it writes may not be true.

**You need** an assistant, a phone for voice memos, a timer and about 90 minutes for a five-minute speech. The steps, prompts and times are our suggestion. For a written piece with facts in it, such as a work memo, use [Writing: brief, draft, check](https://hayzlabs.com/workflows/writing-brief-draft-check/). This page is the personal speech case.

## What the research says

In three experiments on a workplace apology, where "AI-MC" means AI writing help, "the choice to use AI-MC tools diminished perceptions of the apology's authenticity", disclosing it "did not mitigate this effect", and limited use "had no negative impact on the perceived authenticity" ([Glikson and Asscher, Computers in Human Behavior](https://doi.org/10.1016/j.chb.2022.107592), 2023). In an experiment with 293 writers, AI story ideas made stories rated more creative, but "GenAI-enabled stories are more similar to each other than stories by humans alone" ([Doshi and Hauser, Science Advances](https://www.science.org/doi/10.1126/sciadv.adn5290), 2024).

## Steps

| # | Step | Who | Time | Check |
|---|---|---|---|---|
| 1 | Talk for ten minutes into a voice memo about the person: stories, what they always say, what you wish them. List names, places and dates you are sure of. | Person | 25 min | You have three stories only you could tell. |
| 2 | Find a sample of how you talk: a message to a friend, or a transcript of you telling one story aloud. | Person | 5 min | It sounds like you, not like a card. |
| 3 | Ask for structure and options, not a speech (the prompt below is the text you paste into the chat). | AI | 5 min | Two or three outlines and some openings, built from your stories only. |
| 4 | Pick an outline and write or dictate the speech yourself. Ask for help only where you are stuck. | Person | 20 min | Most sentences came from you. |
| 5 | Ask it to list clichés and anything not in your notes (prompt below). | AI | 5 min | Two lists. **If it fails:** if it rewrote the speech, ask again for lists only. |
| 6 | Check every name, date, place and story against your notes, or with someone who was there. | Person | 10 min | You can vouch for every detail. Cut what you cannot. |
| 7 | Read it aloud with a timer, twice. Mark any line you trip over or would not say. | Person | 15 min | It fits your slot. **If it fails:** cut a story rather than speak faster. |

For step 1, [Talk, then edit: dictation and voice mode](https://hayzlabs.com/guides/talk-then-edit/) covers getting a clean transcript.

```text title="Structure and options · v1.0"
I am giving <OCCASION> for <PERSON>. I am their <RELATIONSHIP>.
The audience is <AUDIENCE>. I have <MINUTES> minutes.
Use only my notes below. Do not invent stories, quotes, names, dates
or details. Do not write the full speech.
1. Suggest two or three structures, each a short outline naming which
   of my stories goes where.
2. Give three opening lines and three closing lines in the style of my
   sample. Plain words, no clichés.
3. Tell me which story to cut if I run long, and why.
MY NOTES:
<NOTES>
HOW I TALK:
<SAMPLE>
```

Replace the following:
- `<OCCASION>`: such as `a wedding toast` or `a eulogy`.
- `<PERSON>` and `<RELATIONSHIP>`: such as `my sister Ana` and `older brother`. A first name is enough.
- `<AUDIENCE>`: such as `80 guests, mostly family`.
- `<MINUTES>`: the time you were given.
- `<NOTES>` and `<SAMPLE>`: your text from steps 1 and 2.

```text title="List clichés and inventions · v1.0"
Do not rewrite my speech. Quoting exactly, list:
1. Every phrase that sounds like a greeting card or a stock speech, with
   a plainer suggestion in the style of my sample.
2. Every name, date, place, quote, number or event in the draft that is
   not in my notes.
3. Any sentence that sounds unlike my sample.
If a list is empty, write "none".
MY NOTES:
<NOTES>
MY SAMPLE:
<SAMPLE>
DRAFT:
<DRAFT>
```

Replace `<DRAFT>` with your speech from step 4, and `<NOTES>` and `<SAMPLE>` as before.

> [!CHECK]
> An assistant can add a detail that sounds right: a year, a town, a nickname, a line the person "always said". Check each one against your notes or with someone who was there. A wrong detail at a funeral or a wedding is hard to take back.

## For a eulogy

- **Paste gently.** Health details, family rifts and money are other people's private matters. Use first names, and leave out what the person would not have wanted said. See [What not to paste](https://hayzlabs.com/start/what-not-to-paste/).
- **Let the family read it first.** Ask the closest relative, or whoever is organising the service, to read the final text. They may know a story that should not be told.

> [!OUR RULE]
> The stories, names and feeling come from you. The assistant helps with order, length and the odd stuck sentence. If "Did you write this?" would make you uncomfortable, there is too much of the assistant in it.

**Should you say AI helped?** We found no study of speeches. In the apology study, saying so did not help and limited use did not hurt. Our suggestion: use it lightly, and if asked, answer honestly, such as "I used it to sort my notes."

## How to tell it worked

- Every story and name came from you or someone who was there.
- Read aloud, it fits your slot without rushing.
- For a eulogy, the family has read it and is comfortable with it.
- It has failed if you are reading lines you could not have written, or a story you first heard from the assistant.

> [!LIMIT]
> The studies cover a written apology and short stories. Neither tested a speech given in a room, or a speech mostly yours and tidied by AI. These checks catch invented details and stock phrases, not a lack of sincerity.

## Sources

1. Glikson E and Asscher O. AI-written apologies, authenticity and forgiveness. Computers in Human Behavior, 2023. Scenario experiments (peer-reviewed). [Link](https://doi.org/10.1016/j.chb.2022.107592)
2. Doshi AR and Hauser OP. AI ideas, creativity and similarity of stories. Science Advances, 2024. Randomised experiment (peer-reviewed). [Link](https://www.science.org/doi/10.1126/sciadv.adn5290)