Professor Richard Buckland

The only way of stopping and making mistakes is spending more money and often to reduce mistakes by a lot costs a huge amount more.

So usually we solve mistakes by shifting them from one type of mistake to the other.

So there's type one and type two mistakes that this software can make.

One is where it wrongfully identifies an innocent person as being a bad person.

And the other mistake is where a bad person doesn't get properly identified.

And really there's a knob almost on the software.

You can choose where you set it.

If you set it too far, if you're too determined to catch all the criminals and not let anyone through, then it will wrongfully identify a range of people.

Yeah.

Most of the data sets that things are being trained on are white and middle-class white males and high-profile people in disadvantaged groups, minority groups in America, people with perhaps disabilities.

You can just imagine, I haven't trained it on everyone in the world.

If your training data isn't good enough,

It's making a guess based on what it's seeing.

So yeah, absolutely.

There's all these sort of biases in the software.

Even facial recognition with human makes mistakes.

It's not a very good way of identifying people.

Yeah, I mean, it is.

Can I just say that's not a brilliant argument.

My children try that one.