Welcome to The AI Brief.

Every week, you get clear, honest breakdown of what’s really happening in AI. No recycled press releases pretending to be insight. Just the developments that actually matter.

Most people who use ChatGPT daily have never spent more than 30 seconds in its settings menu. They signed up, typed their first prompt, and have been running on defaults ever since — paying $20 a month for a tool they have never actually configured.

That is exactly how OpenAI designed it. Not because the settings are unimportant, but because buried controls do not create confusion, do not generate support tickets, and do not scare new users away before they convert.

This issue goes through the settings most ChatGPT users have never properly opened: what is actually in there, what each control does, and what OpenAI is quietly opting you into by leaving everything on its defaults.

Let's get into it.

Agents are desperate for ephemeral databases.

They spin up projects, fork environments, test ideas, and tear them down. Over and over. But every database on the market was designed for humans who provision once and stick around. Agents don't work that way.

Ghost is a database built for agents. Unlimited databases, unlimited forks, 1 TB of storage, and 100 compute hours per month. All free. Try it here.

Install the ghost MCP

Open ChatGPT's settings and go past the first screen. Keep going.

Most people stop at appearance or notification toggles and close the menu. The controls that actually matter are several layers deeper, and they change how ChatGPT works in ways the default experience never will.

The most consequential is Custom Instructions.

Two text fields — one asking what ChatGPT should know about you, one asking how you want it to respond. Most users have left both blank since the day they signed up, which means every conversation starts from zero: no context about who you are, what you do, or how you want information presented.

OpenAI buried these fields in a submenu rather than making them part of onboarding, and the result is that the majority of paying users have never touched the feature that would most improve their daily experience.

Below that is Memory. Unless you turned it off — and most people did not know they could — ChatGPT has been storing facts, preferences, and patterns inferred from your conversations since the feature launched. You can view all of it under Settings, then Personalization, then Manage Memory.

Everything it thinks it knows about you is listed there. Some of it is accurate. Some of it is wrong in ways that have been subtly shaping every answer you have received for months. You can edit individual entries, delete specific ones, or turn the whole system off. Most users have never seen this screen.

Then there are Data Controls.

Under this tab sits a toggle called "Improve the model for everyone," which defaults to on. With it on, OpenAI can use your conversations to train future models. Turning it off takes about four seconds, but the button is buried, the label is vague, and the default was chosen in OpenAI's interest rather than yours.

Temporary Chat is something else entirely. It is an incognito mode for ChatGPT — conversations are not saved, memory is not updated, nothing goes to training. It is available from the sidebar on desktop and most people have never clicked it because nothing in the interface points you there.

Finally, model switching. The default experience routes you to whatever model OpenAI considers appropriate, which is not always the right one. You can override it manually: o3 for complex reasoning, o4-mini for speed, GPT-4o for general use. Most Plus and Pro subscribers have never switched models intentionally, and the quality difference between them is not small.

None of these features are hidden in any technical sense. They are all in the settings menu. The problem is that the menu was designed around compliance requirements, not around helping you get more from the product. OpenAI built these controls because regulation and user pressure required them to.

The placement is a separate choice entirely.

OpenAI added a setting that lets you stop ChatGPT from using your conversations to train its models. It defaults to on. The option to turn it off is three clicks deep in Settings and is labeled "Improve the model for everyone," which is a remarkably cheerful way to describe taking something that belongs to you.

Right.

The controls are real and they work. Memory management, data opt-outs, custom instructions, temporary chat — all of it is functional, and all of it was placed where it is on purpose.

Defaults are decisions. When data training defaults to on, OpenAI benefits. When custom instructions are buried in a submenu rather than surfaced during onboarding, more users skip them and more conversations stay cold, which reduces friction during the free-to-paid conversion. When memory management takes three clicks to find, fewer users engage with it.

None of this makes ChatGPT a bad product. It makes it a product whose defaults were chosen in the company's interest rather than the user's, which is true of most consumer software and particularly true when the entry price is low. The difference here is scale — hundreds of millions of people running on settings they never chose, contributing data they did not consciously agree to share, and missing features they are already paying to access.

Knowing where the controls are is the first step to using this product on your own terms.

Open your ChatGPT settings today and spend five minutes in there. Check what memory has stored about you, look at your custom instructions, and find the Data Controls tab. You do not have to change anything. You should at least know what is there.

If you found something unexpected — a memory that was wrong, a setting you had no idea existed — reply and tell me what it was. Those are always the most interesting replies.