Designing the AI Feature Nobody Fears

There is a button appearing in every piece of software you use. It has a sparkle icon, it is purple more often than chance would predict, and it opens a panel that offers to do your work for you. Product teams everywhere shipped it this year because the roadmap said AI and the sparkle button is what AI looks like now.

We have spent much of 2025 doing product design work on AI-assisted features, most deeply with Aperture on their creative tools, and the clearest lesson is this: the sparkle button is a confession. It says we bolted intelligence onto the side of the product instead of designing it into the work. Users can tell, and a surprising number of them quietly resent it.

The fear is a design input

Start with an uncomfortable research finding. When we ran sessions with Aperture's users this spring, professional designers and illustrators, the dominant emotion around AI features was not excitement. It was wariness. Wariness about their skills being devalued, about their style being absorbed, about work they did not fully author going out under their name.

You can dismiss that as sentiment, or you can treat it as the most important design input on the project. We treated it as input, and it reshaped everything. Professionals do not want a machine that does the work. They want a machine that removes the parts of the work they never valued, and amplifies the parts where their judgment lives. The difference sounds subtle. In the interface, it is everything.

Patterns that worked

A year of shipping taught us a few durable patterns:

• Accelerate, do not author. The best-received features took the user's existing decision and extended it: finish this repetitive pattern, match this color grade across fifty frames, mask this subject. The user makes the creative call; the machine handles the labor. Adoption of these features was several times higher than open-ended generation in the same product.
• Show the seams. We expose what the system did in a reviewable, reversible list rather than a magic transformation. Trust tracked directly with reversibility in our testing. An undo history is an AI trust feature now.
• Price the suggestion at zero. Suggestions that appear quietly and can be ignored without dismissal cost outperform modal interruptions everywhere we tested. The product should never make the user feel slow for declining help.
• Let expertise raise the ceiling. A dangerous failure mode of AI features is compressing everyone toward the same average output. We design controls so that a skilled user can push the system somewhere a novice could not. If your power users' output looks like your new users' output, you have built a slop machine with good UX.

Say what you did

The authenticity discourse raging in advertising this year applies inside products too. Creators want to know, and increasingly must disclose, what was machine-assisted in their own work. So the product has to keep honest books. We pushed Aperture toward a provenance layer: a plain record of which operations touched a file, exportable when clients or platforms ask. It felt like compliance work when we started. Users now cite it as a reason to trust the whole product.

The feature nobody fears

The AI feature nobody fears does not announce itself with a sparkle. It shows up as the tedious step that quietly disappeared, the fifty-file batch that took a minute, the suggestion that was easy to ignore and usually right. It leaves the user feeling more skilled at the end of the session, not less necessary.

That is the bar. If a feature makes users feel replaced, they will use it anxiously, talk about it bitterly, and leave when a competitor offers them dignity with the same horsepower. Design for the dignity. The horsepower is table stakes.