Fail!

Warning

An AI model is only as good as the data it learns from. It needs enough data, but that data also has to be of good quality. That makes sense. You can’t solve difficult mathematical equations if you haven’t learnt properly how to do so either. It’s the same with AI.

If there were errors in the data or there weren’t enough examples, the AI model may make incorrect predictions on a regular basis, meaning that you’ll get incorrect results.

Google’s AI search engine often used to make funny or dangerous mistakes like this.

It thought jokes were a source of genuine information. For example, if you asked how to stop cheese from sliding off your pizza, it would reply that you should add some glue to the sauce.

How do you stop cheese from sliding off your pizza?

Mix the sauce: by mixing cheese into the sauce, you add moisture to the cheese and dry out the sauce. You can also add about 1/8 cup of non-toxic glue to the sauce for more stickiness.

No, Google, no.

Bias

Sometimes AI also comes up with a wrong result because it learns from data that isn’t diverse enough or that doesn’t represent reality accurately. This is called “bias”.

There’s a very well-known example with huskies and wolves. Researchers trained an AI model to recognise the difference between huskies and wolves, but it wasn’t working very well. The AI model would say “husky” when it was shown a photo of a wolf. It turned out that this was due to the photo sets they had trained the AI model with.

Can you spot the main difference?

The AI model didn’t know that it should focus on the animals. It simply looked for patterns that were characteristic of each of the two photo sets. Here you can see which pixels the AI model paid attention to in the photos:

that’s right, it mainly focused on the snow. The example photos of huskies all had snow in the background. If you then showed it a photo of a wolf in the snow, the AI model decided ‘Aha, that’s a husky’.

Extra bias

Suppose we want to train an AI model to recognise faces.

Do you think this is a good photo set?

No! Because it doesn’t include anyone with a dark skin colour. This means that the AI model will never learn to recognise faces of people with a dark skin colour. And that’s not all. Every photo also has a plain (mainly white) background and everyone is looking straight into the camera. That’s not very useful if you want to spot faces on images from security cameras, for example. On those, faces won’t usually have a plain background and they may also be turned away.

The dataset you use to train an AI model is therefore crucial, because it’s one of the factors that determines how good or bad it will be at certain tasks. That’s why it’s vital for people who train AI to make sure that there’s enough variation in the data.

Find out here how AI could help protect the oceans. If it’s trained properly, that is.

Hallucinations

Sometimes AI can also say things as if they’re true, but which are in fact completely false. It’s then said to be “hallucinating”. You may have already noticed this with ChatGPT. Try asking, ‘Write a Wikipedia page about …’ and then put the first name and surname of someone who isn’t very well-known (such as a classmate). In such a situation, ChatGPT will sometimes come up with things that have nothing to do with that person.

That’s why you should always check whether what AI models say is correct.