ChatGPT's UI Changes Too Often

Open ChatGPT today and there is a decent chance something has moved since you last used it. Perhaps the model picker has been renamed, or hidden behind a menu, or replaced with an automatic selector that decides for you. Perhaps a feature you relied on is now somewhere else, or gone, or behind a different tier. The product is under permanent renovation, and you are living in it while the builders work.

Change is not the same as improvement

Software teams love to ship. Shipping is visible, measurable, and rewarded. A redesign generates a changelog, a blog post, a sense of momentum. What it does not generate — and what almost nothing in the incentive structure measures — is the cost paid by the person who had finally learned where everything was.

That cost is real. The entire value of a tool you use daily is that it becomes automatic. Your hands know where to go. You stop thinking about the interface and start thinking about the work. Every redesign resets that clock. The task you did without looking now needs looking. Multiply that by a large user base and a frequent release cadence, and you have taxed millions of small habits to produce a screenshot for a launch.

The A/B test you did not agree to join

Some of the churn is not even a decision so much as a permanent experiment. Large consumer products are continuously running tests — different layouts, different defaults, different wording — shown to different slices of users at once. This is why two people comparing notes will swear the app looks different for each of them, and why a feature you found yesterday has vanished today: you were quietly moved between buckets. The interface is not a fixed thing being occasionally revised; it is a rolling trial, and you are an unpaid participant who never saw a consent form.

Optimising a product this way is not inherently sinister — measuring what works is better than guessing. But it treats the interface as a dial to be tuned for aggregate metrics rather than a place a person has to live and work. The metric being optimised is rarely “did this help the user finish their task” and often “did this increase engagement, or upgrades, or time in the app” — goals that are not always aligned with your convenience and are sometimes directly opposed to it. When the layout is perpetually in flux in service of numbers you cannot see, “we're improving the experience” starts to mean “we're improving our experience, and you are the test surface.”

The model picker as a case study in churn

Nothing captures the problem like the saga of choosing which model you are talking to. Over successive iterations this has been a simple dropdown, then a longer dropdown with cryptic names, then a set of names that meant nothing to anyone outside the company, then an “it picks for you” approach that removed the choice, then — after complaints — a partial return of the choice. Users were, at various points, asked to understand the difference between a confusing roster of near-identically-named options, and then told not to worry about it, and then told to worry about it again.

Each of these was defensible in isolation. Together they are whiplash. The one thing a person picking a model wants is stability: learn it once, keep the knowledge. Instead the knowledge has a short shelf life, and the reward for paying attention is that your understanding expires.

Who redesigns are actually for

It is worth being honest about the audience for a big redesign. It is partly users, yes. But it is also the press, the launch narrative, the internal sense that a product is advancing, and the metrics dashboard that lights up when something changes. A stable interface that quietly keeps working is a hard thing to write a triumphant blog post about, even though it is exactly what a mature tool should aspire to be.

There is a survivorship problem too. The people who love the new design are visible and enthusiastic. The people whose workflow just broke mostly do not write in; they just quietly grumble, adapt, and trust the product a little less. Their absence from the feedback looks like approval. It is not.

The pro-user position

Nobody is asking for the interface to be frozen in amber. Genuine improvements are welcome, and some redesigns are real progress. The ask is narrower and entirely reasonable:

• Do not move things that work without a concrete reason that benefits the user, not the launch calendar.
• Let people opt into big changes before they are forced on everyone, and let them keep the old layout for a grace period.
• Stop renaming things. A feature that gets a new name every quarter is a feature nobody can recommend to a friend, because the recommendation is out of date by the time it is read.
• Treat familiarity as a feature — because for a daily tool, it is the most valuable one you have, and the only one you can destroy by accident.

The documentation can never keep up

A hidden cost of constant redesign is that nobody can reliably explain how to use the thing. Every tutorial, help article, screenshot, forum answer and passed-down team convention has a shelf life measured in weeks. Search for how to do something and you find three sets of instructions from three different eras, none of which quite matches the interface in front of you, all written with total confidence. The knowledge base does not so much inform you as offer you a small archaeology exercise: work out which layer you are living in.

This quietly undermines one of software's most valuable properties — that expertise accumulates and can be shared. When the interface holds still, a power user's hard-won knowledge becomes a resource: they teach colleagues, write guides, build shortcuts. When it churns, that knowledge decays as fast as it is created, and the community's collective competence never compounds. Everyone is perpetually a beginner again, not because the tool got harder, but because it got different, which for a daily workflow amounts to the same tax paid over and over.

Change as a substitute for direction

There is an uncomfortable possibility behind the churn: that constant visible change is what a product does when it is unsure what it should be. A confident, mature tool can afford to sit still, because it knows its job and does it. A product still searching for its identity — or under pressure to look like it is advancing as fast as the models beneath it — reaches for redesign as evidence of momentum. Motion becomes a proxy for progress, and the interface becomes the place that motion is performed.

The trouble is that users cannot tell purposeful evolution from restless fiddling; they only feel the ground move. And the more often it moves without clearly making their life better, the more they suspect the changes are for the company's benefit rather than theirs — a suspicion the frequent migration of features up the paid tiers does nothing to dispel. Stability, in the end, is a form of respect: it says the company values your accumulated familiarity more than it values a fresh screenshot. Its absence says the reverse, whatever the launch post claims.

Muscle memory is a feature you can only destroy

The most undervalued thing a daily tool possesses is the ability to be used without thinking about it. When the interface holds still long enough, your hands learn it; the tool recedes and the work comes forward. That state — fluent, automatic, effortless — is the entire payoff of familiarity, and it is worth an enormous amount precisely because it is invisible when it is working. You only notice it when it is taken away, which is exactly what every redesign does. The buttons move, and suddenly you are looking at the interface again instead of through it, thinking about the tool instead of the task.

The cruel asymmetry is that muscle memory is slow to build and instant to destroy. Weeks of daily use go into learning where everything is; a single release erases it in an afternoon and resets the clock. A company can add this kind of value only patiently, by leaving things alone, and can subtract it in one deploy. That should make stability feel precious to anyone designing a daily tool — a hard-won asset held in trust for the user. Instead the industry treats the interface as a canvas for demonstrating progress, repainting it often enough that the fluency never has time to form. The result is millions of people permanently re-learning tools they had already learned, paying in attention for changes that were mostly for someone else's benefit.

The underlying product is genuinely capable. That is what makes the churn frustrating rather than trivial: people want to build a durable working relationship with it, and the interface keeps filing for a fresh start. Consistency is not the enemy of progress. For the person who uses your product every day, it is most of what progress feels like.