← Back to the island

The chat tower · the atrium

AI chatbot: the tower, floor by floor

This is the entrance to the ai chatbot tower — the assistants themselves, what each one is actually good at, the ways all of them fail, and what it takes to run one on hardware you own.

The tower is built around a distinction that most coverage skips. There is the question of which assistant to use, which changes every few months and is the subject of endless argument. And there is the question of how to use any of them well, which barely changes at all and determines almost everything about whether they are useful to you. The second wing is the larger one here, deliberately.

The assistant floors are written to be dated rather than definitive. What each model can do moves quickly, a page claiming a ranking is out of date within a season, and the useful thing a floor can offer is a description of what a service says it does, what it costs, what it does with your conversations, and when that was last checked.

The failure wing exists because these systems fail in a specific and consistent way: fluently. A wrong answer arrives in the same confident register as a right one, with no signal distinguishing them, and that is a property of the technology rather than a bug awaiting a fix. The floors there are about recognising where that risk is highest and building the habit of checking.

And there is a wing about privacy, which is the question people ask last and should ask first. What a service does with what you type into it varies enormously between products and between the free and paid tiers of the same product, it is written down, and almost nobody reads it.

Three levels, and you are on the first. This atrium lists the floors; a floor covers one subject completely and holds the shorter guides underneath it. Nothing here is more than three clicks from the front door.

Where to start

Seven ways in. The second is the one that changes how useful all of this is to you.

“Which assistant should I use?”
Start at the assistants
“I use one and the answers are mediocre.”
That is using them well
“How much of what it tells me is wrong?”
Go to where they fail
“Is my conversation being used for training?”
That is what happens to what you type
“Is the paid version worth it?”
There is a floor for that in the assistants
“Can I run one without the internet?”
That is on your own machine
“I want it to stop sounding like a press release.”
Read using them well

The assistants

One floor per assistant, all answering the same questions in the same order, all carrying the date they were last checked: what it is good at, what it costs, what it keeps.

ChatGPTWhat the free and paid tiers include, and what the company says about your data.Being built
ClaudeWhat it is used for most, and what the published terms say about conversations.Being built
GeminiHow it sits inside the rest of Google, and what that means in practice.Being built
CopilotThe AI assistant built into software people already have open all day.Being built
PerplexityAnswering with sources attached, and how to check the sources it attaches.Being built
The open modelsWhat the openly published models are, and why they matter even if you never run one.Being built
Free or paidWhat a free AI chatbot leaves out, what the money buys, and who genuinely needs it.Being built

Using them well

The wing that transfers between every assistant and does not go out of date. If you only read one part of this tower, read this one.

Asking properlyThe handful of changes that reliably improve answers, and the folklore that does not.Being built
Giving it your materialWorking from your own documents rather than from the model’s memory.Being built
Long conversationsWhy quality drifts over a long session, and what to do about it.Being built
Making it sound like youGetting past the default register that everybody now recognises.Being built
Tasks they are good atThe work where these tools are genuinely strong, listed honestly.Being built
Tasks to avoidWhere the failure rate is high enough that the tool costs you time.Being built

Where they fail

Not a criticism but an operating manual. These failures are structural, they are predictable, and knowing the shape of them is what makes the tools safe to rely on.

Confident wrong answersWhy fluency and accuracy are separate, and where the risk concentrates.Being built
Checking what it saysA workable habit for verifying, rather than an instruction to be careful.Being built
Numbers, dates and quotesThe three categories worth checking every single time.Being built
Bias and gapsWhat the training leaves out, and where that shows up in answers.Being built
When it agrees with youThe tendency to follow your framing, and how to test against it.Being built

What happens to what you type

The privacy wing. Each floor quotes the service’s own published position with the date, because these settings change and the defaults differ between tiers.

Training on conversationsWhich services use your chats, which do not, and where the setting lives.Being built
Work and confidential materialWhat should never be pasted into a consumer assistant, and why.Being built
Deleting your historyWhat deletion actually removes, and what is kept regardless.Being built

On your own machine

Running a model locally: no connection, no account, nothing leaving the computer. A short wing about what that costs in quality and in hardware.

Why run one locallyThe cases where it is worth the trouble, which are fewer than enthusiasts suggest.Being built
Getting startedThe simplest working setup, and what your machine needs to manage it.Being built
What you give upAn honest comparison with the hosted services, without loyalty to either.Being built

What this tower will not do

It will not crown a best assistant. That answer changes every few months, it depends heavily on what you do, and any page claiming otherwise is either out of date or selling something.

It will not pretend these tools are reliable narrators. The failure wing is not a disclaimer bolted on at the end — it is a quarter of the building, because using them well means using them with that in mind.

It will not tell you what a service does with your conversations without quoting the service. Those policies change, they differ between free and paid, and second-hand summaries of them are frequently a year behind.

And it will not carry affiliate links, which in this category means declining the fastest-growing referral programmes on the internet.

One closing thought about how to use this tower. If you already have an assistant and are vaguely disappointed by it, the wing to read is not the one about which product to switch to. Switching rarely fixes a disappointing experience, because the difference between a poor answer and a good one is usually on your side of the conversation rather than theirs. What holds instead is simple: every assistant floor carries the date it was checked, and no floor carries an affiliate link.

Written by Alberto Gulotta

Founder and editor of AI Tools Primer, writing from Palermo, Italy. Thirty-five years of taking computers apart, starting with a Commodore 64 — the long version is on the about page.

Something wrong on this page? Write to aitoolsprimer@gmail.com and it gets fixed.

Independence and limits

No affiliate links and no paid placements anywhere on this site. Nobody pays to appear here, and no company has seen this page before you did.

This is general information, not professional advice. Where a page touches money, health, safety or the law, it names its source and the date it was read — and your situation may still differ. See the privacy page and the cookie policy.