What is an AI's context?

Status: 🔸 Seed → 🔸 Growing → ✅ Well-formed → 🔸 Fruitful → 🔸 Retired

Transparency label: human-only

When we work with a general-purpose AI like ChatGPT, its behaviour is governed by what is called its context. By this, we mean the information it is given which is additional to the knowledge it learned during its training. We cannot change its learned knowledge, but we can control some of its context. That is how we interact with it. We do this in several ways.

The most immediately obvious is the chat window. Here, you can conduct a conversation with the AI: instructing, guiding and questioning it in real time.

To work jointly on shared text with the AI during a chat, we use a canvas. The AI opens a canvas, alongside the chat window, when we tell it to. The canvas works like a text editor in which both human and AI can edit the content.

We can share files with the AI during a chat by uploading them into the chat window. This lets us refer the AI, in the course of a conversation, to information contained in the files.

The AI stores our chats until we choose to delete them, and we can resume a chat from where we left off. But the AI cannot yet refer to the content of an earlier chat during the course of the currently open chat.

A less obvious way to control the AI is through custom instructions. These are settings which let us configure how the AI behaves across all our chats with it. We can give it information such as:

what to call us;
what we do;
what kind of personality the AI should have;
what traits it should have; and
any other behavioural instructions we want to give it.

To collaborate with the AI on a larger piece of work where we want to share the same files across many chats, we use what ChatGPT calls a project.

All the chats in a project are grouped together for ease of access.
There is a space where we can upload project files.
All chats in the project can read the project files.
I don't think the AI can yet refer to the content of other chats during the conduct of a chat.

Different projects have different purposes, and we might want the AI to behave differently to suit these purposes. In that case, we use project instructions to configure the AI's behaviour within that project. Project instructions:

can contain any behavioural instructions we want to give the AI;
apply only to that project; and
override custom instructions.

Underlying all of the above are the system level instructions coded into the AI by its makers. These configure how it behaves with regard to ethics, bias, safety, and many other important matters. We can't change these, but we can build on top of them through our custom or project instructions.

To summarise: context is the information we give the AI to control its behaviour. We can use it to govern:

all chats through custom instructions;
groups of chats (organised as projects) through project instructions and project files;
individual chats through prompts, the canvas, and file uploads.

We build context. And the better the context, the better the response from the AI. A vague prompt gets a vague answer. A clear prompt, with the right background, gets us something much more useful.