00:00Can AI make mistakes?
00:03Of course. Google's AI recommended one user top their pizza with non-toxic glue
00:09to stop the cheese from slipping off.
00:11Not exactly intelligent. Among other unusual answers to user questions,
00:16the dogs have played in American basketball and football leagues
00:20and that Barack Obama was the first Muslim president of the US.
00:23All fake news. But how does AI come up with such nonsense?
00:28Google used summarised information in what it calls AI overviews
00:32to provide quick answers.
00:34But the AI was unable to distinguish serious facts from jokes or satire.
00:40The tip to add glue to pizza likely came from a joke on Reddit.
00:45And however diligently AI learns, things can still go wrong.
00:49Artificial intelligence used in a motorboat racing game
00:52was supposed to optimise itself.
00:55But the AI made the boat spin in circles, catch fire
00:59and crash into docks and other boats.
01:03While the boat scored lots of points, it never crossed the finish line.
01:08In this case, the goal was not formulated precisely enough for chap GPT.
01:14Mistakes like these can have drastic consequences in real life.
01:18There were plans to use AI to categorise patients
01:20depending on how urgently they need hospital treatment.
01:25But the AI concluded that patients whose previous treatment had cost more
01:28were sicker and therefore needed help more quickly.
01:32Big mistake.
01:33Many poorer people only incur lower treatment costs
01:36because they cannot afford to see a doctor, even if they're just as sick.
01:40If AI sends them to the back of the queue,
01:42they're doubly discriminated against.
01:46What is clear is that AI is only ever as good as the data it's fed
01:50and how precise its instructions from humans are.
Comments