What is truth? It sounds like a question for a philosophy class. It has become a practical one, because a lot of people now take their answers from a machine that has no idea.
I want to explain what that machine actually is, because once you see it the responsibility lands somewhere obvious.
A probability engine
A large language model is a brain that can only ever surface an answer based on probability. A guess. It runs some maths against a huge body of information to work out what most likely answers the question you asked.
Ask it who the king of England is and it will say Charles. Now ask who the king of England was in 1901. The whole calculation changes. It has to go looking through everything it holds for the answer that best fits the new question, and it will hand you whatever comes out on top.
It gives you the most likely response. It has no way of checking whether that response is true.
A hallucination is a lie
The industry calls the misses hallucinations. I think that word is too kind.
Think of the model as a brain. If I try to recall what happened a week ago, and I don't remember exactly, and I tell you a version anyway, I'm lying to you. I might not mean to. You still walk away with something that isn't what happened. That is what a model does when it fills a gap.
People have these tools everywhere now. In the messaging app, in the browser, in the office software. And a lot of them assume that every time they ask, they get the truth, every time. A few simple tests will show you that isn't the case.
So the job of knowing what is true sits with us, the humans in this. I think plenty of people are handing that job to the tools as well.
What outsourcing your thinking looks like
I shared a story recently about a university professor who set his class a test to do at home. Most of them scored between 90 and 100 per cent. A few weeks later he gave them the same test in person, unannounced. Most scored 50 per cent or below on exactly the same questions. Only a handful landed close to their first result.
His point was that the students had started outsourcing the whole thinking process. Whether the tool was right or wrong on any given question almost doesn't matter. They weren't retaining anything. They weren't being challenged by a perspective they met in a book. They weren't learning to be curious.
I think this is happening faster, and at a larger scale, than most people are willing to admit.
The fakes are about to get very good
Here is the other half of the truth problem. Right now I could take a single video of myself and use one tool to create a visual clone of me. I could take the audio and use another to clone my voice. Combine the two, point a model at everything I've ever written and said, and you have something that looks like me and sounds like me, talking about whatever you like.
People send me examples every week asking if they are real. At the moment I can usually tell. In 18 months I have no idea. The scams are going to be ridiculous. I've written about the voice cloning version of this and what a business can do about it.
Your responsibility as a user
I love this technology. Used with curiosity, and with a firm grip on what is right and wrong, it lets us do phenomenal things. That is the whole idea behind amplifying intelligence. The risks are real all the same, and they land on each of us as individuals.
So here is what I would ask of anyone using these tools.
- Know what it is. A probability engine. Treat every answer as a strong first guess.
- Check anything that matters. If you would be embarrassed, out of pocket or in trouble when it turns out to be wrong, a person verifies it first.
- Keep doing your own thinking. Use the tool to go further. If it is doing the thinking for you, you are losing the skill you will need to catch it out.
- Test the tools yourself. You can't judge what is possible, or what is fake, from the sidelines.
- Tell other people. Show them the misses. Awareness spreads the same way complacency does.
There is one more side to this. The answer to these risks can't be that a small few get to hold the models and decide what the rest of us may use. Staying in control means all of us understanding the thing, with access to it. Models that run locally, on your own phone or computer, are a good sign on that front.
It is an interesting time to be a human. If we stop thinking, stop having these conversations and stop holding on to what we know to be true, we hand that over as well.