The generic AI panel became a sign that nobody designed the feature

Open any product you pay for and look at the top right corner. There's a sparkle. Behind it, a panel slides in from the right, says something like "ask anything" and suggests three prompts nobody asked for. Close it and it comes back in the next update, sometimes under a new name.

The person who described this phenomenon was designer Jon Sabutis, in a piece published on October 25, 2026 in UX Collective. He isn't an AI skeptic: he uses a model every day to write code and generate interface variations. And even so, he spends part of the week looking for the button that turns off the sparkle. That gap between the tool (good) and the product built on top of it (shallow) is the angle I want to pull on here, because it's a design choice, not a side effect of the technology.

The same product, forty times

Sabutis lists where he's seen the identical panel: notes app, spreadsheet, design tool, email client, team chat, video call, browser, photo editor, presentation, calendar, music player, and bank. These aren't similar panels. It's the same panel. Strip away the logo, and nobody on the team that built any one of them could point out which is theirs.

The reason is structural: model vendors sell the same capability, through the same API, with the same demo. A company that buys the demo and ships it didn't design anything, it forwarded it. That's why the feature is named after the technology, "AI Assistant," "Copilot," "Intelligence": there's no user goal to name it after. Nobody ever shipped a "database feature." They shipped search.

This diagnosis isn't new, it was just ignored. In 2023, before most of these panels existed, Amelia Wattenberger published Why Chatbots Are Not the Future showing why a blank text box is the worst way to expose a capable model: it says nothing about what the model can do. That same year, Maggie Appleton released the Language Model Sketchbook, with working alternatives, one of them a model that lives in the document's own right-click menu. The industry read both pieces and shipped the text box anyway.

Forced on, not chosen

What sparks outrage isn't the panel existing. It's it being turned on without asking and being hard to turn off. Microsoft is the clearest case because it's the biggest: Copilot showed up in Notepad, Snipping Tool, Photos, and Windows Widgets, with a fixed button on the toolbar and no simple way to remove it. In March 2026, the company announced it would cut "unnecessary Copilot entry points" starting with these four apps, with the head of Windows publicly acknowledging there had been a lack of intention about where to place the AI.

Meta went further. In spring 2025, a blue and purple ring began appearing in WhatsApp's chat list and search, and there was no way to remove it. The company called the feature optional: you could choose not to tap it. That's the definition of optional a company uses when the feature doesn't survive a real test.

Every designer knows what a feature that people keep turning off is called. It is a bug with a marketing budget.

Jon Sabutis, designer and author of the piece in UX Collective

The reaction doesn't come only from cranky enthusiasts. A Hacker News thread, "I'm drowning in AI features I never asked for and I hate it," passed three hundred points with complaints from people who use AI every day by choice: the problem was the version they couldn't refuse. In a Pew survey from June 2025, half of American adults said the advance of AI in daily life left them more concerned than excited, against one in ten on the opposite side.

What the cycle costs

A toggle that turns itself back on is small, and that's exactly why it does so much damage. A setting is a promise: the product said "you decide" and the update decided again. The click to fix it costs nothing. The cost is what the person learns from it, that the product is going to run over them again.

That lesson doesn't stay confined to one feature. Every update becomes something to audit. The user stops reading the changelog to find out what's new and starts reading it to find out what was done to them. That's real cognitive friction: it isn't the panel itself, it's the vigilance it teaches, attention that should go to the work and instead goes to guarding the interface.

From there, the problem migrates from the product to the brand. In January 2025, Apple paused AI notification summaries for news apps after the feature rewrote a BBC alert falsely claiming a murder suspect had shot himself. The summary came on by default; the error was the model's, but the name on the notification was Apple's, and the complaint went to Apple, not to whoever trained the model.

The next stage is the strangest of all. In August 2024, Procreate's leader recorded a video saying the app would never add generative AI, and artists celebrated. A promise not to ship a roadmap item went viral. That's what the cycle buys in the end: a market where the absence of your feature becomes a competitor's selling point.

The bill behind the sparkle

If the panel doesn't work, why does it keep shipping? Because someone already paid for what's underneath it. UBS estimates the big clouds will spend about US$4.1 trillion in capex between 2026 and 2028, after about US$492 billion in 2025. In 2026, Amazon, Alphabet, and Microsoft together are expected to spend 102% of cloud revenue on capex: every dollar that comes in goes back out on chips and data centers, with a bill left over.

That money has to come back from somewhere, and that somewhere is the subscription you pay. That's why the panel comes on by default: a feature that's off is a feature nobody's being charged to use. At the same time, the measured return is low.

MIT's NANDA group analyzed 300 deployments and interviewed 150 leaders, and found that 95% of companies saw no measurable return on generative AI pilots, published in August 2025. The 5% that did had integrated the tool into a specific workflow and learned from it. Gartner is kinder and still bad: one in five AI initiatives delivers any measurable return.

A feature with AI, not an AI feature

The distinction that underpins my take here is this: there's a feature built with AI and there's an AI feature, and the sparkle panel is almost always the second kind. The difference isn't in the technology underneath, it's in which step of the product process got skipped.

The table below follows the characterization Sabutis makes of these third-party products in the original piece; it isn't a check against the makers' official documentation, but his reading of where each feature lives and how it's triggered.

The Shape of AI catalog, by Emily Campbell, maps the interface patterns that have emerged around models, and the ones that work are almost always invoked from inside the work itself. The sparkle panel isn't a pattern, it's the place where the pattern that hasn't been designed yet gets stored. Jakob Nielsen called this technology the first new interface paradigm in sixty years, because you describe the outcome you want instead of the step-by-step; given that, the docked panel is even stranger, a new paradigm getting the interface of a 1990s support chat.

The counterargument worth taking seriously

The strongest defense of the panel being on by default isn't silly: without it visible, most people would never discover the feature, and adoption data is what justifies next quarter's AI budget in front of a board that's already committed the capex. That's real, and I don't think you can ignore the pressure on whoever needs to show a fast number.

But "panel engagement" isn't the same metric as "the person solved the task," and that's where the argument breaks down. An experiment with real opt-in, the feature off by default, and measurement of who turns it on and keeps it on after a month gives a more honest signal than panel opens, because it measures whether the problem was real. If leadership insists on on-by-default, the concession that's usually possible to negotiate is a one-tap off button that survives updates, which alone avoids the broken-trust cycle described above.

Four questions before it goes on the roadmap

These are common product questions, and that's exactly why they work:

• Who asked for it? Not "people would like," but which person, on which task, was stuck in a way this unsticks. If the answer is "the board," stop.
• What verb is it named after? If the name is the technology, there's no goal inside it yet. If you can't rename it after what it does, it isn't a feature.
• Would it survive off by default? If it only gets used when it's forced, it doesn't earn its space. Whoever wants it will find the button.
• Would it survive at cost price? The token is subsidized today by the trillions in capex. Price the feature as if it weren't. If it dies, it was never worth more than its own marketing.

Most sparkle panels fail all four. The good AI features that exist pass all four and have no sparkle at all, because they were designed the way everything else should be: starting from a person with a goal, working backward.

The sparkle in the corner of the screen isn't an icon chosen with care. It's a confession: the company has a model and doesn't know where it's useful. Before approving the next "AI" item on the roadmap, the question for whoever designs the product isn't whether the model is good. It's whether someone, on a real task, was stuck before it existed.