How AI learns, and when to trust it

Artificial intelligence, or AI, is a computer program that does tasks we usually think need a human brain, like recognising a face, translating a sentence or answering a question. Most AI today uses machine learning. Instead of a programmer writing every rule, the computer is shown thousands of labelled examples and finds the patterns itself.

Fish, volunteers and a learning machine

The Department of Conservation films fish in marine reserves to check whether the reserves are working. Watching thousands of hours of video takes people a very long time. So DOC and a charity called Wildlife.AI started Spyfish Aotearoa. Volunteers watch short clips online and count the fish. Their counts become labelled examples that train an AI to recognise snapper, blue cod and other species on its own.

Rubbish in, rubbish out

An AI can only learn from the data it is given. If it is shown fish only in clear water, it may fail in murky water. If a face-recognition AI is trained mostly on photos of one group of people, it makes more mistakes with everyone else. This is called bias. The AI is not being mean. It is copying the gaps and unfairness in its data.

AI can sound sure and still be wrong

Chatbots write smooth, confident sentences, but they predict likely words rather than checking facts. They can invent sources, dates and quotes. Always cross-check anything important with a trusted source, and remember that deepfake photos and videos can be made to look completely real.

Is AI right for this job?

Who owns the data?

In Te Hiku, in the Far North, Te Hiku Media collected hundreds of hours of recordings of people speaking te reo Māori to build speech-recognition tools. They created the Kaitiakitanga Licence so the data is used for the benefit of Māori and not simply taken by big companies. It is a strong example of a community deciding how its own data is used.