There were more than one system proven to “cheat” through biased training materials.
One model used to tell duck and chicken apart because it was trained with pictures of ducks in the water and chicken on a sandy ground, if I remember correctly.
Since multiple medical image recognition systems are in development, I can’t imagine they’re all this faulty trained with unsuitable materials.
They are not ‘faulty’, they have been fed wrong training data.
This is the most important aspect of any AI - it’s only as good as the training dataset is. If you don’t know the dataset, you know nothing about the AI.
That’s why every claim of ‘super efficient AI’ need to be investigated deeper. But that goes against line-goes-up principle. So don’t expect that to happen a lot.
