When GPT-5 launched in August 2025, the reviews agreed: clear progress over the previous generation on factuality, complex reasoning, and admitting its own limits — steadier on open-ended, cross-domain questions too. In everyday use it would still occasionally garble a reply or hang, but as a model, GPT-5 was a very strong system.
The problem is —
a good model does not mean the product experience is right.
Here is how the official introduction described it:
“GPT-5 is a unified system with a smart, efficient model that answers most questions, a deeper reasoning model (GPT-5 thinking) for harder problems, and a real-time router that quickly decides which to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say ‘think hard about this’ in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. In the near future, we plan to integrate these capabilities into a single model.”
Yes, I know you can still switch to Thinking mode manually.
But that doesn’t change the fact: by default, it decides for you when to think hard, instead of leaving that switch in your hand.
This design isn’t “smarter.” It’s “more in charge of you.”
Deleting the model entry points + forcing a real-time router = the root of the problem
The biggest change in the web app was removing the dedicated entry points for different scenarios — reasoning, coding, writing — leaving only GPT-5 and GPT-5 Thinking, with a real-time router deciding in the background which mode you get.
The difference: switching models used to be explicit, and bound to a scenario. The moment you switched, you were actively declaring your intent to the system — “I’m going to write code,” “I need mathematical reasoning.”
Now that step has been cut. Mode selection has become a black-box action buried in system logic, performed on your behalf.
You could argue: “GPT-5 is strong enough that dedicated models are unnecessary.” But the dedicated entry points were never only about capability. They were an explicit act of confirming what you need. They made you choose before the task began — which reinforced the causal chain of intent → tool → result.
Real-time routing does have value. In mass-market scenarios, automatically assigning light or heavy models cuts latency, cost, and cognitive load. But when it becomes a black box — no prompt, no choice, no explanation — users lose the driver’s seat without noticing, and output quality fluctuates in ways that can’t be audited or reproduced.
Using it feels like dealing with an overconfident waiter. You’ve barely sat down when he pats his chest and announces: “I know what you’re going to order — and I’ve decided whether it should be stir-fried or slow-braised.” If you want it slow-braised, you need a code word: appending “think carefully” to the name of the dish.
On the surface, this design saves effort. In substance, it has three consequences.
Consequence 1: it defines your intent for you
Much of the time, you never said “I want deeper reasoning” or “just give me a quick answer” — the system chooses for you. The router switches between the base model and the Thinking model on its own judgment, without asking first.
The automation looks considerate. But over time, it erodes the habit of figuring out what you need before you ask — the system will guess anyway. And the habit of having your needs guessed for you is the first step in the weakening of active thinking.
Consequence 2: it hollows out your right to choose
In the GPT-4o era, most people also used the default — but at least the scenario-based entry points were there. Switching to the stronger coding model was itself a declaration: I’m writing code, so I need the model that codes better. Whatever optimization the system did afterward happened on top of your stated intent.
GPT-5 cut that process out. You can still switch to Thinking manually, but the default flow has no scenario entry and no explicit collection of intent — the system simply shifts modes based on conversation type, complexity, and your history.
The biggest problem: you cannot predict when it will shift gears for you. The trigger is not your explicit intent; it is the system’s inference. Your active chain of “set the goal, then pick the tool” has been inverted into “the tool picks first, and you adapt to its choice.”
Consequence 3: it wears down your agency
Agency comes from defining the goal first, then choosing in service of that goal. When the system defines your intent and picks your model, your participation in the process gets compressed into “accept the result, tweak afterward.”
Keep that up, and you lose your sensitivity to goal-setting and process control — as long as the system’s answer is “close enough,” you take it. The habit quietly replaces the question “what do I want?” with “what does the AI think I should want?”
In other words: you are not taming the tool. The tool is taming you.
A technical upgrade has been traded for an interaction downgrade — what got removed wasn’t steps. It was your opportunities to define and to decide.
When a tool starts deciding your needs, your path, and your pace, you are not far from losing the driver’s seat. The boundary of a good tool is to strengthen the user’s judgment and capacity to act — not to replace them. When a product crosses that line, however advanced the technology, it is in essence a regression.
And this is not just a GPT-5 web-app problem. It is the occupational disease of every “considerate” AI: the better it knows you, the more easily it thinks for you, chooses for you, acts for you.
Which raises a bigger question — in an era when models keep getting stronger and keep “understanding” us better, what actually remains ours? I wrote about that separately, in Why AGI Will Be Impossible — Because of Three Real Human Moats.
For the people building AI products, the reminder is this:
A truly good AI product doesn’t do everything for the user. It leaves the user the chance — at the moments that matter — to define, to choose, to confirm.
AI should know its own boundary: it can compute faster and see wider, but it should not decide “whether to” on a human’s behalf.
And designers should keep asking themselves: is this feature strengthening the user’s perception, context, and agency — or eroding it? If it’s eroding it, then however smart the model underneath, the direction is anti-human.