This article summarizes a series of posts I wrote on UX + AI (originally shared on LinkedIn) that explores usability issues which often show up when implementing AI features in existing products.

Pitfall 1: Task Interruption

AI integration is disrupting user flow with poorly timed prompts and suggestions. Companies like LinkedIn and Notion have both implemented intrusive AI features that interrupt users' normal workflows in order to encourage them to use new functionality. Beyond just being annoying, this is disruptive to the users productivity.

Suddenly, digital tools have become a place where every action seems open to feedback by an overly-eager virtual assistant who is VERY available, but needs my expertise to help them help me. Is anyone else having Clippy flashbacks?

Principle 1: Keep users on task, and only break concentration for urgent matters

Most new AI features are not worth breaking task concentration to let users know about.

Pitfall 2: Inviting Dead Ends

Another common issue is a misleading user interface that sends the user to a paywall. Duolingo's "Explain My Answer" feature is an example of this issue – it seems like it'd help the user understand more about the sentence, but it actually functions as an upsell opportunity for an AI-specific subscription.

I love the idea of letting users try these new AI features out. They’re new, and people don’t know what they can do!

But! Once people have tried the functionality out and decided not to purchase, you’ve learned something: the product offering didn’t demonstrate enough value to the user for them to justify a purchase.

Principle 2: Ensure all your in-product buttons are functional

Don't let these inviting dead ends remain littered throughout the product for users to avoid, even after they've decided not to use the AI features. Maintaining these dead ends in the interface erodes user trust and satisfaction over time.

Pitfall 3: Breaking Existing Paradigms

Meta's replacement of Instagram's search feature with "Ask Meta AI" is a great example of how quick AI integration can disrupt established user expectations. The traditional search flow has been transformed into a more complex, AI-first experience - it feels like Meta tried to put AI in front of users as quickly as possible, and didn't consider the existing search flow in the process.

Principle 3: Adhere to existing paradigms unless there's a great reason not to

In this case, Meta could instead include preserving the traditional search experience while offering AI as an additional option, or add AI to the top of the results within Search, like Google has done.

The new experience is entered via a search call to action, but feels very unfamiliar to a user expecting search.

Pitfall 4: Decision Fatigue

The proliferation of AI features means users are constantly bombarded with the option to complete a familiar flow in a brand new way. This can easily lead to decision fatigue and second-guessing of the users own expertise.

For instance, when working in Notion to record meeting notes and thoughts, I've encountered a ton of AI-powered editing suggestions that, ironically, have slowed me down to the point where I've turned them all off.

Principle 4: Have an opinion about how AI might fit into an existing workflow

Of course, your users will surprise you - let them, and build alongside them. But don't give them an experience that feels like an open-ended world in their productivity software.

Final Thoughts

Effective AI integration should enhance user capabilities without creating additional cognitive burden. Ellen Chisa puts this well in her post Enabling Magicians, not Magic: "It won't feel like our AI wrote something. It will feel like we wrote something, and that we're a much better writer than we were just a few years ago."

So, when creating any new AI experience, ask yourself - do my users need to know this is AI? It's not always necessary for a user to be informed that they're interacting with an AI model rather than traditional code. The key is identifying if and when this distinction matters to the user experience.

By steering away from the above pitfalls, and adhering to everything else we've learned from Jakob Neilsen, Brenda Laurel, Steve Krug, and Don Norman over the years, we can create AI-powered features that truly serve user needs and elevate the experience of our products.