Two pieces of information collided in my head this week. One came through a book I read during our phone-free vacation: Nudge. Improving Decisions about Health, Wealth, and Happiness. Final Edition, by Richard H. Thaler and Cass R. Sunstein. The other was the result of running the same prompt (word by word) through the same LLM, first with the default setting, then adapting those with the CounterDefaults tool.
The theory (from the book) explains that default settings are one of the easiest ways to shape behaviour, with people “choosing” the default option (as in: not doing anything) 80-90% of the time. We are busy after all, and we assume that the default option is the “best” option.
Then you get 150 notifications per hour from your phone, because by default all notifications are on. And you get 15 additional newsletters per week, because the website by default signs you up to ALL topics. Or you end up in the default pension plan, only realising upon retirement that you could have had a much better return on investment.
Because that’s the thing: some default settings influence your subsequent choices without you (ever) realising that there’s a difference.
So here’s a little exercise for you:
Choose a question you wanted to talk through with your AI tool of choice. Write the prompt down in a separate document for easy copy-paste.
Paste the prompt into your AI chat, and have the conversation.
Now, go to https://counterdefaults.aixdesign.co/ and play around with the settings to create one profile that is very far away from the defaults, and another one that is a mix. Copy both into your document.
Paste the instructions for the first set of instructions in an incognito chat, followed by the exact same prompt you used before, and have the conversation.
Now, do the same with the other set of instructions, again followed by the exact same prompt/question you started with.
The results were fascinating. It wasn’t just the tone that changed. Depending on the settings, the AI focused on different details within the question, and in one instance it interpreted the problem completely different and went down a different rabbit hole.
I also realised that I do care for tone, even with an AI. I just couldn’t stomach the confrontational persona, and abandoned the chat pretty quickly. Apparently, constant reassurance also puts me on edge. I guess I love long meandering conversations of a very specific flavour, and curiosity and kindness are necessary ingredients for both human and artificial counterparts.
For me personally, that means I can now refine a lot more the type of communication I want to have with my AI instances, including where I want it to go. Hyper-logical and scientific is great for product research, and not so great for discovering the emotional landscape of your target market. Neither means I have to put up with a caustic and unfriendly AI, because that’s also a setting, and it’s independent of scientific rigour. You can be exact and kind, as a human and as an AI.
Special thanks to Abi Awomosu for constantly challenging my work with AI, and for throwing so many resources and thought challenges at me through her online presence.
