Neue Forschungsergebnisse zeigen, dass 90% der AI -Chatbot -Antworten über Nachrichten einige Ungenauigkeiten enthalten, und 51% enthalten „signifikante“ Ungenauigkeiten.

Submission Statement

AI is at its most impressive when the answers to the questions it seeks are in its training data. It’s why it can get almost 100% in law and medical exams. The questions have been discussed so often on the internet, that all the answers are in training data scrapped from the internet. This can make AI very useful for narrow tasks, say detecting breast cancer in x-rays, but it’s much less useful when it has to deal with new information that doesn’t come from extensive training data.

For obvious reasons, it does not enjoy those advantages when it comes to news and current affairs. The great drawback of current AI is that it lacks reasoning ability, so frequently makes simple errors when it encounters new combinations of information that aren’t in its training data.

All the big tech companies developing AI are collectively pouring hundreds of billions of dollars into the efforts. To varying degrees, they are under huge pressure to justify this to investors. Hence, there is a rush to integrate AI into everything.

Perhaps the hope is that fundamental problems with reasoning will be quickly solved along the way. But they haven’t been, and so we see ridiculous outcomes like this.

Language models should not be used as knowledge repositories. They should be used to interpret language that they derive the facts from.