AI Experiments 6 min read

AI is a mirror, not a genie

Every time a PM asks me to 'add AI' I ask what behavior we're reflecting back at the user. Most answers involve a magic wand. Most good answers involve a mirror.

The wish

A PM walks up with a Jira ticket titled 'Add AI'. What they mean, most of the time, is: make this thing feel smarter. What they hope for is a small deity that lives inside the product and just knows.

That's not the offer. The offer is a mirror — something that reflects your intent back at you, faster and sometimes more articulately than you could alone. The design question is what you choose to reflect.

The mirror

A mirror is useful because it's honest about what's in front of it. LLMs are the same: they reflect the quality of the prompt, the data, the context. Garbage in, a confidently-written paragraph of garbage out.

The design job is picking what to mirror — which part of the user's intent deserves to be amplified, which part is noise. That's taste work. It's the same job we've always done.

What to do instead

Ask: what signal is this user already producing that we can reflect back more clearly? That reframe kills the magic-wand briefs and starts producing AI features that feel less like spectacle and more like help.