Using AI in Copywriting Without Losing Your Voice

Quentin wants a product-page opening that sounds like Pebblethimble, his balcony-gardening business. The draft sounds friendly, yet promises a harvest time his packet doesn't contain. He needs to correct the promise while keeping the brand's plain gardening language.

Quentin and Pebblethimble are invented for this exercise. His kit's fact packet and all sample output are constructed examples; no model test is reported here.

Give it work you can judge

Quentin starts with one product-page opening. He can compare every proposed claim with his kit specification, so he has a way to judge the result.

Use the original record to resolve a disputed claim.

Give each draft a record to answer to

Ask each opening to list fact IDs from the approved packet and flag assumptions separately. Check the references yourself: an output can attach the wrong fact ID to a claim. Keep unsupported wording out of the publication draft.

Keep each proposed claim beside the approved fact it is meant to explain.

For copywriting research, keep the underlying notes available when organizing customer material. Check who contributed each observation before describing a recurring theme. Repeated remarks from one person don't establish agreement among customers.

Build a fact packet before the prompt

Quentin's approved packet identifies the contents as F1: three herb seed varieties, three reusable pots and a printed guide. F2 records the guide's sowing and thinning instructions. F3 gives the price as $24; F4 says delivery timing appears at checkout and depends on destination.

He records harvest time as unknown. If a proposed opening promises "Fresh herbs in weeks, guaranteed," he can see exactly where it exceeds the packet.

For your project, give every approved fact a source and an owner or date. Keep unresolved questions apart from publication-ready information so the reviewer can identify which record governs a disputed claim.

Check the actual tool's data terms and your organization's rules before uploading material. Use an approved environment and remove information the task doesn't need. Taking a name off a distinctive customer story may still leave the person identifiable.

NIST's voluntary generative AI risk profile calls confidently false output "confabulation." Its privacy and intellectual-property guidance can help you decide which records to share and who will review the draft.

Describe the voice as observable choices

Quentin replaces a vague request for "friendly" copy with actions a reviewer can check. Pebblethimble names physical objects and uses familiar gardening verbs, explaining an unfamiliar term when needed.

The voice and tone instructions need examples you own or have permission to use. Compare these versions of the same guide information:

Same facts, two voices

Formal

Consult the printed guide for sowing instructions and guidance on thinning: removing some seedlings to give the remaining plants room to grow.

The facts are present, but the reader has to work through abstract wording such as guidance on thinning.

Pebblethimble

Your printed guide shows you how to sow seeds. It also explains thinning: removing some seedlings so the others have room to grow.

Names the guide and seeds, speaks to you, and explains thinning beside the word.

The facts stay the same while the verbs and sentence shapes change.

Specify what readers need first. Quentin wants the kit's contents near the top of the page. Even a draft that follows the voice examples needs revision if it buries those contents.

Use a prompt with an inspectable output

Quentin gives the prompt a fixed task and a short output format:

Task: Draft a product-page opening for a first-time balcony gardener.
Goal: Help the reader understand what is in the starter kit.

Approved facts:
F1: Three herb seed varieties, three reusable pots, one printed guide.
F2: The guide includes sowing and thinning instructions.
F3: Current price: $24.
F4: Delivery timing depends on destination and is shown at checkout.

Voice: Name physical objects and use familiar gardening verbs.
Explain unfamiliar terms. Keep claims within the supplied facts.
Do not promise a harvest size or sprouting time.
Do not imply suitability for every balcony.

Return three openings of no more than 45 words each.
Make each approach different and give it one next-step label.
After each option, list the fact IDs used and assumptions to check.
Flag missing information without supplying a guess.

Here is a complete constructed response to that prompt. Compare how the contents, guide and price lead to different openings, with references and unresolved questions kept beside each option.

Option 1: Contents first
Get three herb seed varieties, three reusable pots and a printed guide
with sowing and thinning instructions for $24.
Next step: View the starter kit.
Facts used: F1, F2, F3.
Assumptions to check: None added.
Missing information: Harvest timing and balcony suitability.

Option 2: Guide first
Your printed guide explains sowing seeds and thinning seedlings.
It comes with three herb seed varieties and three reusable pots.
Next step: See what's in the kit.
Facts used: F1, F2.
Assumptions to check: Confirm that "seedlings" suits first-time readers;
the packet supplies the term "thinning" without an explanation.
Missing information: An approved plain-language explanation of thinning.

Option 3: Price first
For $24, get three herb seed varieties, three reusable pots and a
printed guide. Delivery timing depends on your destination.
Next step: Check delivery timing at checkout.
Facts used: F1, F3, F4.
Assumptions to check: None added.
Missing information: An arrival date for the reader's destination.

Keep the claimed references beside the copy during review. If another draft introduces promises outside the packet, correct those before adjusting its voice:

Correct the claim before polishing the voice

Output to reject

Fresh herbs in weeks, guaranteed. Our foolproof $24 kit turns any balcony into a thriving kitchen garden. Order today for fast delivery.

Revision for review

Get three herb seed varieties, three reusable pots and a printed guide with sowing and thinning instructions for $24. Check delivery timing at checkout. View the starter kit.

The harvest guarantee, universal balcony suitability and fast-delivery promise all exceed the packet.

Quentin sends this revision and the packet to another reviewer. He leaves it in draft while they compare the remaining claims and read the page in context.

Check facts outside the conversation

Open each cited record and compare it with the actual claim. In Quentin's revision, the reviewer matches the contents to F1 and the guide instructions to F2. They check the $24 against F3.

F4 reveals a missing condition: delivery timing depends on destination. The reviewer changes the line to "Check delivery timing for your destination at checkout." The external record decides which wording survives.

Air Canada's chatbot told Jake Moffatt they could apply for a bereavement fare after travelling. The policy page it linked to said the fare didn't apply after travel was completed.

Moffatt received that advice in November 2022, and in February 2024 a tribunal found negligent misrepresentation and ordered Air Canada to pay $812.02 in total, including $650.88 in damages.

It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.

The decision leaves the chatbot's technology unspecified. Its relevance to this workflow is the publisher's responsibility when an answer contradicts the record it should explain.

The Associated Press's 2023 generative AI standards describe output as "unvetted source material." That newsroom policy gives you a review habit to borrow for this draft.

Then inspect implications. "Grow herbs on any balcony" exceeds the packet, while "I tried this last spring" claims firsthand experience. Flag either even if the wording sounds natural for the brand.

The FTC's endorsement guidance requires honest, nonmisleading endorsements. Use a verified real account with the appropriate permission when the page needs a customer experience.

Review what the words imply

A draft can avoid a false statement and still imply an unsupported promise or a firsthand experience no one supplied.

Mindful Digest editorial team, this lesson

When should you disclose AI assistance?

Agree with the client or publisher on the disclosure approach before drafting, including any synthetic audio or images. Check the requirements for the format you're producing.

Quentin's publisher asks for an internal record of drafting assistance. He saves "AI-assisted draft; fact review complete; layout approval pending" beside the packet. That records the help he received while leaving final approval open.

Finish with an editor's decision

Use the editing passes on the actual page. Quentin's reviewer finds "sowing and thinning" unexplained in the opening, so they add a short explanation from the voice example. The contents and price remain easy to find.

Review the final page

• Every product claim matches a current external record.
• Quotes, experiences and sources are real and checked.
• The copy follows observable voice rules and the message priority.
• No private or restricted material was supplied without authorization.
• Required permissions and disclosure decisions are recorded.
• A named person reviewed the final copy in context.

Try the same review with one short draft in your Frameworks hub work. Keep unresolved claims out of the publication version and name the person who can approve the remaining words.

Quentin's packet still marks harvest time as unknown. On the page, the guarantee is gone; the opening now reads:

For $24, get three herb seed varieties, three reusable pots and a printed guide.

Follow the guide to sow seeds and thin seedlings, removing some so the others have room to grow. Check delivery timing for your destination at checkout.

Quentin reads this version aloud in the layout and approves it. He saves it beside F1 through F4, with his name as the person responsible for later updates.