Familiarity Is a Floor, Not a Ceiling

Open any AI product and you meet the same room: one input field, a blinking cursor, and a line of grey placeholder text inviting you to ask anything. No sidebar. No categories. No sign of what the software can actually do until you have already guessed at it and typed the guess.

Download

Search finds pages. Chat finds answers. Those are different jobs, a case I made in Search Finds Pages. Chat Finds Answers.. One box looks cleaner in a screenshot. Cleaner is not the same as clearer.

When someone challenges that box in a design review, the usual defence is familiarity: users already know how to type into a message field. Designers call that Jakob's Law.

Jakob Nielsen wrote it in 2000: "Users spend most of their time on other websites, so they expect your site to work like all the other sites they already know." jakobs-law Correct on its own terms: ignore it and users meet an unfamiliar pattern, click the wrong thing, and leave. It is a good law. That is why rooms that have already decided what to ship like citing it.

Here is the distinction that gets lost. A pattern is a single interaction element: a search box, a scroll bar, a cart icon, a chat window. A container is the larger decision that holds those elements together, what surfaces exist, what each one is for, how they hand off to each other. Jakob's Law, cited correctly, governs patterns. Familiarity is a floor: what users already bring with them from other products. It is not a ceiling. The law tells you what deviation costs. It does not tell you what a design should reach. I am not arguing that a chat interface is always the wrong pattern. I am arguing that the law is being asked a question about containers, and it was written to answer a question about buttons.

The Law Is Right. That Is What Makes It Useful to People Who Want to Stop Thinking.

Trevor Calabro has spent twenty years in UX research watching this happen. A team hides a dense table's controls to make the interface look cleaner. Usability testing finds nobody can locate them. Someone produces a screenshot from an unrelated niche analytics tool that hides its controls too, and the room relaxes. Calabro named this weaponising the law in January 2026: one cherry-picked example does not build a mental model, and a mental model is the only thing the law actually claims to describe. calabro-weaponizing

The law survives that misuse because the underlying claim is true. Nielsen found that centring a logo instead of placing it top left produced a six-fold jump in users who failed to reach the homepage in one click, four per cent against twenty-four. nng-centered-logos His formula for the trade-off is blunt. A design ten per cent better than the standard still loses if early meetings cost users twenty per cent in aggravation. jakobs-law Everyone with a launch date benefits from that formula. A chat box has no information architecture to test, no empty states to design, and no navigation model to defend in a review. It demos in four seconds, and the demo is the whole interface. The law says deviation is expensive. It has never once said the convention is good.

A Law About Buttons Is Being Used to Settle a Question About Architecture

The same essay contains the half nobody quotes: "Gradually, more and more design patterns reach the elevated peak of user expectations and become the new standard. Then, according to Jakob's Law, you should implement such an innovation." Nielsen wrote that about scroll bars and cart icons. He then extended the law itself to voice interfaces, virtual reality, and what he called AI-driven interactions. He argued separately that the new AI paradigm "completely reverses the locus of control" between person and machine. nielsen-ai-paradigm In the same breath he narrowed the extension: honour the law "for those interaction elements that have crossed over from traditional technologies." Reading that as elements versus architecture is my inference, not his phrase. It is the reading that makes both halves of his own essay agree with each other. A scroll bar crossing into a new product is a pattern. Whether the product is a chat window, a search box, or three separate surfaces is the container the pattern sits inside.

Calabro's fix for weaponised citation is a three-step test. Name the task in plain language, collect examples from at least five comparable products, then look for what actually repeats across them. calabro-weaponizing Jakob's Law applies to tasks, not to brand names. That rules out copying one company because it happens to be large. Nielsen Norman Group's own diary study sorted 425 real conversations into six distinct task shapes: search queries, funnelling, exploring, chiselling, expanding, and pinpointing. nng-ai-conversation-types The researchers' conclusion: different task types need different interfaces, not one box that treats them all the same. Run that test on a general-purpose AI product. The comparison set will not agree on a single text box. Search engines, help desks, spreadsheets, and sat-navs solve wildly different tasks through wildly different interfaces.

The Labs Reopened the Panel

Anthropic answered its own chat box's limits in June 2024 with Artifacts: a second panel beside the conversation that holds the document, the code, or the design the model just produced, so a person can point at a paragraph instead of re-describing it in prose. anthropic-artifacts OpenAI opened the same kind of panel four months later, in October 2024, and called it Canvas, built for the same reason. openai-canvas Editing a highlighted section is a click. Re-explaining that section through a prompt is a rewritten essay. Two labs that could have kept building only the box that made their own product simplest to demo chose, within a single year, to add a second surface with visible state instead. That is not nostalgia for buttons. It is the market re-running Calabro's task-comparison test on itself and finding that one text field could not hold every task its own users brought to it.

There is a line painted along the tiles of most swimming pools, the same kind of unmarked drop a blank input field hides under its blinking cursor: the point where the bottom drops away. Nobody explains it to you. You feel the difference the moment your foot reaches down mid-stroke and finds nothing, long before you could describe the gradient in numbers. Anyone can practise a stroke in water shallow enough to stand in: the arms look right, the breathing looks right, and the only thing untested is what happens once the water no longer meets you halfway. A chat box is the shallow end of a product. Almost anyone can type a sentence into it on the first try, which is exactly why so many products never had to find out whether the rest of what they built could hold weight once a user crossed the line.

A Menu Shows What the Software Can Do

One participant in Nielsen Norman Group's 2023 diary study typed a question about Independence Day foods into Bing Chat and got back a paragraph of prose. She clicked thumbs down, turned to Google instead, and found pictures, recipes, and party favours in about the time it took to read the chatbot's answer. She had not asked the wrong question. She had asked a search-shaped question, one of the six task types named above. The box could not tell it apart from a five-turn conversation someone else was having about something else entirely. Another participant in the same study asked Bing Chat for free events in Nashville. Unsatisfied, she tried the interface's own suggested follow-up question and got results nearly identical to the ones she had already read. Neither failure came from the model guessing wrong. Both came from a container that offered one shape of answer no matter which of six tasks a person actually brought to it. nng-ai-conversation-types

Nielsen has made the underlying numbers explicit himself. Citing the OECD's PIAAC literacy survey of roughly a quarter of a million adults across twenty rich countries, he puts half the adult population of countries like the United States and Germany in the lower literacy bands. He estimates, by his own admission a best guess rather than a measured figure, that fewer than 20 per cent, likely closer to 10 per cent, can write prompts articulate enough for advanced use of a chat interface. His own explanation for why is worth quoting directly: "GUIs have superior usability because they show people what can be done rather than requiring them to articulate what they want." nielsen-articulation-barrier Collapsing a menu into a blank field does not remove the cost of choosing. It moves the cost onto the user, and it lands heaviest on whoever has the least literacy to spare. That argument has a limit worth admitting on my own site too. The June essay split search from Ask at roughly five hundred sessions a month, because "the discipline scales before the traffic does." Nobody was watching closely enough to make me pay if I had been wrong. A team with a launch date and a growth target this quarter does not get that luxury. The roadmap decides for them, and it reaches for the box for the same reason a review room does.

The Strongest Case for the Empty Box, Properly Stated

Give the counterargument its full weight. A blank field is not a failure of design. It is arguably the most powerful interface built since the command line, because it does not pretend to know your request before you arrive. Every button is a guess about what you might want, written by someone who has never met you, and a menu of forty options is forty guesses, most of them wrong for you specifically. Buttons age the moment a company adds a forty-first feature nobody built a home for. A sentence never runs out of room. If the true cost of a menu is every task its designer failed to imagine, the box is not evasive. It is the only honest admission that software cannot predict what you need.

That argument survives contact with everything except the data on what actually happens once someone starts typing. The same Nielsen Norman Group diary study that logged 425 conversations in 2023 found no correlation at all between how long a conversation ran and how helpful or trustworthy the person rated it. Infinite room to be exact should reward whoever uses more of that room. It did not. Length tracked how well the first sentence was written, not how well the box served the task. The honesty the objection prizes belongs to the person typing, not to the interface receiving it, and least of all to the roughly one in ten Nielsen estimates can write that sentence well.

The Blinking Cursor

Jakob's Law prices familiarity for the parts a product borrows from older technology, not for the architecture built around them. Until a design review can name that difference out loud, citing the law is just the room's earlier decision wearing a citation. Somewhere this week, a new user opens a product built like that room decided, one field, no menu, and closes the tab holding their question.