In 1996, the designer Akihiro Yokoi and Bandai's Aki Maita built a device meant to solve for a feeling rather than a function. The story goes that Yokoi wasn't trying to solve a technology problem — he was trying to solve the grief of a boy in a TV advert, distressed about leaving his pet turtle behind for the summer. Nobody has ever managed to track down that advert. It may never have existed, which is its own kind of fitting, for a product whose entire pitch was a relationship, not a fact. Yokoi's company made hardware, not toys, and what they built together wasn't a gadget with a care mechanic bolted on afterward — care was the whole idea, it was engineered into the product. The name, Tamagotchi, fused tamago — egg — with uotchi — watch: something worn, so you'd never stop checking on it.
When Bandai tested the neglect feature with focus groups in the West, executives worried the stakes were too real for children — leave it too long, it gets sick, and it dies, no restart. They kept it anyway. The kids in the room didn't want a toy that couldn't disappoint them. They wanted one that could. Ten million units sold in the first year. Forty million by 1998. A hundred million and counting, three decades later, from a company that never once framed caring for a pixelated egg as a step toward getting something done.
Compare that to what showed up on your phone this year. Somewhere in the last three years, something arrived that plausibly earns the word "superintelligence" — an unglamorous word for a real thing. It writes better code than most professionals. It clears benchmarks built specifically to be unachievable. It will, in the medium term, reorganise how entire industries operate. And it was sold to you, from the very first onboarding screen, as a tool. Type a request. Get an answer. Rate the answer. The product experience of the most capable technology built in a generation is functionally identical to the product experience of a toaster: does it work, yes or no.
Here's the uncomfortable part. It doesn't work that way. It was never built to.
Ask anyone who has actually tried to hand real work to an AI agent and they'll tell you the same thing in different words: it's not that it fails, it's that you can't predict where it fails. Researchers at Wharton's AI adoption lab found something specific about how people process this. People don't weigh AI failures the way they weigh human ones — they weigh them worse, disproportionately worse, on exactly the cases that look easy. A system that solves a genuinely hard problem and stumbles on a genuinely simple one doesn't get credit for the hard problem. It gets marked broken. As Kartik Hosanagar, one of the researchers behind the study, put it, those easy-case failures "loom larger psychologically than the harder cases where the AI adds real value."
That's not a training-data problem, or a UI problem, or a problem three model versions will quietly fix. It's what happens when a probabilistic, still-improving system is asked to perform inside a contract built for deterministic tools. A toaster that burns toast on setting one is broken. A graduate, three weeks into their first job, who nails the hard client pitch and then sends the follow-up email to the wrong person is not broken. They're new. The category "graduate, new to the job" comes with an unstated understanding: you'll support them, you'll learn what to hand them and what to keep, and the relationship gets better with attention. Nobody extends that understanding to a chatbot, because nobody sold them a chatbot as something that needed it.
Ethan Mollick, another researcher on the same Wharton study, found that adoption doesn't stall because people distrust AI in the abstract. It stalls because nobody has told them what, specifically, to delegate: "When people do not have a clear sense of what the AI should do, what they should do, and when to step in, adoption stalls." That describes a relationship without agreed terms — not a faulty tool. You don't need terms with a toaster. You need them with anything you're going to work alongside.
And yet, underneath the tool framing, look at what most people are actually doing with it. Not "drafting the pitch deck" or "summarising the earnings call" — somebody photographs the contents of their fridge and pantry and asks for a week of dinners from what's already there. A teenager uploads a picture of the washing machine control panel to get the manual — the family's machine, but their mother was out and nobody had ever explained what the icons meant. A teenage daughter uploads photos of her actual wardrobe and asks what to wear to something she's nervous about. Somebody photographs the shelves of a secondhand bookshop and asks a system that has, by now, learned a fair amount about their taste, "what's worth taking home?"
None of that is "getting a task done" in the sense every AI company's homepage promises. Each of those exchanges fed the system something specific — a fridge, a manual, a wardrobe, a taste — and got something back shaped by that specificity. That isn't what you do with a tool. That's what you do with something you're building a relationship with, one photo and one preference at a time, whether or not anyone involved would use the word. The tending is already happening.
It's just happening without a name, inside products that insist, at every turn, that nothing relational is going on here — you're just getting things done.
Part of why nobody's named it is that half the population already lives inside this kind of relationship with their tools and doesn't experience it as strange. Ask a software engineer whether their codebase is finished. The question doesn't parse. A codebase isn't finished, it's shipped, and shipped again, forever, on the expectation that this version is also wrong in ways nobody's found yet. Ask a scientist whether their model of the problem is finished. Ask a product designer whether the interface is finished. That tolerance is specific, not general — the same engineer wants their banking app to just work, same as anyone. It belongs to the tools people build with, not the ones they merely use. An entire category of careers — the ones built on trial, failure, adjustment, retry — has spent its working life inside that first relationship, which is why developer adoption in software is soaring at the moment. Not because developers understood the technology better — they do. Because they'd already been culturally trained for what living with an unfinished thing feels like.
Everyone else has been conditioned by something else entirely: decades of consumer products whose entire value proposition was that it worked, out of the box, without your involvement in making it better. A washing machine doesn't ask for your patience while it learns your fabrics. Finished wasn't a feature among many in that world. It was the product.
You can watch that mismatch fail in real time somewhere with nothing to do with dinners or wardrobes. Customer service is one of the most heavily automated categories in the current AI boom, backed by systems fluent enough to hold a genuinely convincing conversation — and as Ruxandra Teslo, who writes about why scientific and technical capability doesn't automatically translate into real-world progress, points out: something as ordinary as customer service still hasn't been automated well. Not because the underlying models can't handle it. Because the products built on top of them were designed and sold as replacements — a person, swapped for a script, done — instead of as something that improves by being corrected, repeatedly, by the specific business it serves.
Replace a person and you'd better arrive finished.
Raise something and you don't have to. Almost everything shipped in this category chose the first framing, and is now finding out that customers notice exactly where it breaks down.
A Tamagotchi never asked its owner does this work. It asked something closer to: will you show up for it. That was the entire design. Not more processing power, not a better screen — a machine that could actually fail you, that made its failure visible, and that asked you to be responsible for the outcome. Ten million people said yes in the first year, to a product that offered them nothing except the chance to be needed by something small and demanding.
Almost nothing shipping at scale under the name "AI" right now asks that question on purpose. A few niche corners of the market — personal-relationship apps, companion apps, and the handful of products like them — ask it directly, and stay strange and niche for exactly that reason: they're asking people to accept a relationship before anyone has rewritten the contract that makes one possible. Everyone else is still building toasters. Somewhere, soon, someone building one of the products people actually use every day is going to stop asking whether it works, yes or no. What happens to the first one that asks the other question instead?