Today at 12 I'm running the first Live from the Lab session, on what AI is and what it knows. If there's one idea I'd like people to leave with, it's this one. AI sounds just as sure when it's past the edge of what it knows as when it's well inside it.
Once you understand where its knowledge comes from, you can see where that edge is. Don't worry, this is about as technical as it gets.
01Where the knowledge comes from
Take a huge amount of text: websites, books, articles, code. Compress it all into one very large file that captures the patterns in it. That file is the model, and making it is called training.
When training is done, it's finished. I think of it like baking a cake. Once it's out of the oven, that's the cake. It doesn't keep learning when you chat to it. Every new conversation starts from exactly the same cake.
The key thing for me is that it's not a database. It hasn't stored the text it was trained on. It's stored the patterns. It doesn't know the third word of the fifth Harry Potter book. It knows what that word likely is. When you ask it something, it isn't looking the answer up. It's writing what a likely answer looks like, one word at a time.
02Where it stops
That gives you three edges.
Recent. Training stops on a date, and the model knows nothing after it. Last month's price change, this week's news, the latest version of a tool: none of it is in there.
Local. It has never seen your organisation's documents. Your policies, your processes, your client list, how your team actually does things. If it wasn't public, it wasn't in the training.
Specific to you. Even where something was public, small and specific details leave very little pattern behind. A niche regulation, a small company's history, the exact wording of a contract clause. The rarer the detail, the less it has to go on.
03Why it doesn't warn you
Because it's writing a likely answer, not looking one up, it doesn't notice when it crosses one of those edges. The answer comes out in the same calm, confident tone either way.
The way I put it is that everything that comes out of AI is made up. A lot of it happens to be right. A dangerous amount happens to be right, because we forget it's being made up. Ask about a policy your company changed last month and you might get a confident answer about the old version, or a very plausible one about a policy that never existed.
So the confidence of the answer tells you nothing about whether it's true.
04What about web search?
Most of the main AI tools can now search the web, and that helps a lot with the recent edge. But it doesn't always search, and it doesn't always tell you whether it did. When it does search, the answer is only as good as the pages it found and how well it read them.
So if it matters, look for the sources it used and click a couple. Search moves the edge. It doesn't remove it.
05Three quick checks
Before you rely on an answer, ask yourself three things:
- Is it recent? Anything that might have changed since the model was trained.
- Is it local? Anything about your organisation, your team or your clients.
- Is it specific to me? Exact figures, names, dates, clauses, small details.
If the answer to any of those is yes, give the AI the source rather than asking it to remember. Paste in the policy, attach the document, or ask it to search and show you where it found things. And then check the bits that matter, the way you'd do an important sum on a calculator twice. You own the output, not the AI.
None of this makes AI less useful. It makes it much more useful, because you know when to lean on it and when to hand it the facts first.
06Come and ask
We'll go through all of this in today's Live from the Lab session at 12, a free half hour online. Bring an example of something AI told you with total confidence that turned out to be wrong. They make the best questions.