We took nine brands*, prompted an LLM with their tone of voice guidelines, and asked it to write a short newsletter in the tone of voice covering three things: a price rise, an office move and a staff volunteering day fixing up a school garden.
The great vanilla-fication
With thousands of brands relying on the same three LLMs for their copy, the result is a slide towards sameness. The copy they generate might well be competent, but it’s also totally forgettable. Which is… not ideal.
Vanilla-fication is really happening. Here’s the proof.
Only one successfully followed the tone of voice it was given – and that was Zizzi, the only one we’d created a detailed tone of voice prompt for.
*Experian, Lemonade, Mailchimp, Microsoft, Monzo, UAL, the UK Government, Wise and Zizzi.
What to do about it
There are only two ways to stop your brand sounding like everyone else’s. The first is to stop using AI, hire specialist tone of voice consultants, and go back to the good old Word doc days.
The second is to train your LLM properly, with a brand language system.
1. Make your voice truly distinctive.
People can only hold so much in their heads, so human guidelines have to stay simple. AI never forgets and never cuts corners – so we can make your tone far more intricate, and therefore more distinctive.
2. Wire it into your tools
We build a library of writing examples, each scored against your voice, and use it to train your AI. You can even have your own secure version of Definition AI, with all the best models built in.
3. Keep it fresh
Models drift, so we’ll regularly tweak your prompts and add to the library to keep everything up to date.
One we made earlier
We’ve been working with a big Australian pensions organisation since 2019, helping them sound clear and human in their communications. We built their tone of voice, trained hundreds of their people to use it, then built them a bespoke AI suite complete with a brand language system.
Twice as fast – comms now take half the time to write.
(Nearly) twice as clear – AI writing scored 76 for readability, against 48 for the originals.