Why AI Wearables Fail When They Add a Device but Remove No Behavior

AI hardware obituaries are easy to write. The smartphone won. The Humane AI Pin died. The Rabbit R1 stumbled out of the gate. Case closed. Everybody go home and put the weird orange gadget in a drawer.

It is a satisfying argument. It is also incomplete.

Humane really did reach a hard ending. HP acquired its key AI technology, patents, and talent in 2025, and the service that made the AI Pin useful shut down. Rabbit is a different story. Its launch was rough, but the company is still shipping software. rabbitOS 2.3 arrived in July 2026 with Hermes Agent, BYOK model support, and more agent tooling.

That does not retroactively make the R1 launch good. It does mean “they all failed” is lazy product analysis.

The more useful product question is what a new piece of AI hardware removes from my life.

Every new object pays a carry tax

My phone already owns a pocket. It has a camera, radios, a bright screen, my accounts, payment cards, apps, contacts, and years of muscle memory. A new device does not compete with an empty space. It competes with all of that while asking for another charger.

That is the carry tax.

A smartwatch gets around it by taking over a place many people already use and making a few phone interactions faster. Headphones let me take a call or listen to something without holding the phone. Smart glasses can capture a photo while my hands are occupied. None of these examples replaces the smartphone as a platform. They remove the need to touch it in a specific moment.

The weaker AI gadgets add a step instead. I take out the new device, ask it to do something, wait for a cloud model, and reach for my phone when the action fails or needs confirmation. The gadget did not replace the phone. It became an opening act for the phone.

This is why “familiar form factor” is not enough as an explanation. Plenty of terrible products have been shaped like familiar objects. What matters is whether the hardware removes enough behavior to justify its physical and mental overhead.

The R1 is a useful example because it is a genuinely charming object. Teenage Engineering helped make something that looks playful and specific in a category full of smooth gray rectangles. I get why people wanted it. I wanted the idea of it.

But Rabbit launched it as a standalone route to an app-free experience. Its original announcement described a Large Action Model that could operate apps for you. That set the bar much higher than “a fun pocket computer for experimenting with AI.” The hardware could earn attention, but only the action layer could earn the extra pocket.

Rabbit later admitted there was a large gap between expectations and the original software experience when it announced rabbitOS 2. I respect the directness of that. The product they are improving now may end up being good at a different job than the one used to launch it.

That happens. Sometimes the thing users keep is not the thesis on the pitch deck.

The hardware is the shortest part of the interaction

As a product designer, the industrial design is the fun part to stare at. The actual interaction is much longer.

The device has to notice that I am speaking, capture clean input, wake its radio, reach a model, interpret my request, call another service, wait for that service, and tell me what happened. If the task changes something important, it also needs confirmation and a way to recover.

Every part can be “working” while the experience still feels broken.

A second of microphone delay, a few seconds of model latency, and a slow third-party integration can turn a simple request into standing around while a small object thinks at me. If the answer is probabilistic, the device also needs enough interface to show uncertainty. Did it hear me? Is it waiting? Did it act? Is it asking permission? Can I undo it?

Phones are very good at this repair work. They have space for progress, errors, account recovery, permissions, and dense information. A wearable trying to disappear has less room to explain itself at the exact moment AI software needs more explanation.

That tension is more interesting to me than whether the object has a screen. Minimal hardware and probabilistic software naturally pull in opposite directions. The cleaner the physical interface becomes, the more disciplined the system has to be about state.

One glowing light cannot mean listening, thinking, acting, waiting, failing, and “DoorDash changed its API again.”

What is left when the server goes away?

The Humane AI Pin exposed the harshest version of cloud dependence. Once the service ended, the object lost the system that justified it. The shutdown did not merely remove a feature. It reached backward into hardware people already owned.

This is the question I would ask much earlier in a hardware review now: what survives?

Ordinary glasses remain glasses. A smartwatch usually remains a clock, timer, and local sensor if one online service is having a bad day. Ray-Ban Meta glasses lose most of their intelligence without Meta's services, but they were designed around an existing object and work with a companion app instead of pretending the phone no longer exists.

Meta has leaned further into that relationship. Its Meta AI app became the companion app for the glasses, with some continuity between conversations on the glasses, phone, and web. That is a much more honest division of labor. Quick capture and audio happen on the face. Setup, history, recovery, and detail happen on a screen.

The glasses still raise difficult questions about recording, privacy, battery life, and whether people around the wearer know a camera is active. Reusing a familiar object lowers the carry tax. It does not make the social cost disappear.

The brief I would use now

I am less interested in hearing that a device provides “AI anywhere.” My phone has already made that phrase fairly meaningless.

I want one specific moment. What am I doing, why is the phone awkward, and what can this hardware finish before I would have pulled the phone out anyway? Then show me what happens when the network is slow, the model is unsure, or the partner service rejects the action.

That is where the product is.

The smartphone did not end hardware innovation. It just made vague hardware much harder to justify. A new AI object needs to be unusually good in one context, clear about its state, and graceful when the cloud misbehaves. It should use the phone for phone-shaped work without acting embarrassed about it.

I still want strange new gadgets. Honestly, I want more of them. I just want the strangeness to live in the interaction, not only in the object I now have to remember to charge.