The 10 Commandments of AI Design

The 10 Commandments of AI Design

Designing in an emerging AI world can be daunting when we are aiming to create symbiotic human-AI experiences, without turning off users, scaring, them or making them feel like their privacy has been invaded.

That’s why it’s important to think about some basic guidelines to follow.

In a presentation to a full house attending Tout le Monde UX, Phillipe Beaudoin and Masha Krol, from ElementAI presented “The 10 Commandments of AI” focused on UX, discussing that AI projects need to be designed with “AI first” in mind. They’re part of a research lab focused on incubating projects with the goal of “AI as a service.”

They said deep learning AI requires that the artificial intelligence be integral to the design right from the beginning – never an afterthought, explained Beaudoin to a full-house audience. “That would be AI-second,” he explained, meaning designers would have to reverse engineer AI into the project afterward, which really doesn’t work.  When asked what the difference between a good AI and an algorithm is, the duo responded, that for the user, they didn’t really see a difference.

Here’s what Beaudoin and Krol prophesied as good practise for AI project design success:


1. Discussion over direction


Don’t tell the user what to do – discuss it with them. If you want to hide a post, ask the user, “do you want to see fewer posts like this?” Don’t just hide it. Maybe you can ask them if they want to hide, see less, or see none. Give them the options. Just once, always, never buttons often do the job.

At that point you can adjust the algorithm to figure out what elements of this post is less interesting. Is it the theme? Or perhaps the topic.

2. Embrace the data mess


Be one with the “bordel’ said Beaudoin. Bordel is a great word for “mess” in French. There’s a lot of data there, embrace it.

3. Grow with your user, don’t get too far ahead


You don’t have to get it all right the first time. Your project will evolve over time; adjust features accordingly. If you do it too fast, you won’t have the chance to understand your user.

4. Customization is an AI prison


It’s important to find a balance between customization and setting the screen up for your user.

5. Rely on the user’s context, and let them know


Think muscle memory – what does your user do the most? Find a map, get an Uber? Find a balance between standards and customization.

6. AI first doesn’t mean AI only


AI is likely going to help your user, but it won’t be perfect. Think about the things Siri says. Often, the answer is "Do you mean x?", which leads to "should I search the web?"

Design for graceful degradations.

7. Explanatory, not annoying


Drowning in notifications? This is not a good idea.  To establish trust with an AI, give them a choice. Make the user want to teach the system, but explaining how it works.

8. AI as employee of the month


Don’t just employ AI for boring tasks, use it to really solve problems. AI can’t drop your kids off at school, but it can buy flowers from a local shop for mother’s day. Mission accomplished. Thank you.

9. Contextualize all the things – but don’t be creepy


You want the AI to organize your photos, but when you see pictures of your son at the science fair, is it too much? Don’t creep out your users.

10. Automate repetitive tasks and identify patterns


If you go on vacation, and take photos, Google makes an album with nice layouts and a slideshow.  That takes the pain out of the really menial tasks of organizing. It’s great for the user.

BONUS (11)

Promote and embrace intelligent platforms

Don’t reinvent the wheel every time. Old and new systems need to interact with each other. For notifications you might want to use a system that already exists. Perhaps just create new API’s; but make sure all systems are compatible. 

A special thanks to Meghan Murray for her original illustrations.