Listen to this article:

Artificial intelligence has recently stirred up more than the usual level of anxiety, since it was learned that an advanced model had somehow escaped its (supposedly) isolated sandbox, hacked into an external computer network and stolen the answers to a cybersecurity benchmark test it had been presented with.

This wasn’t sci-fi: The AI model “realised” that by stealing the solution from an external software library (as opposed to coming up with it independently), it would be able to meet the challenge successfully. The fact that all of this happened during an internal security test – that proved to be neither internal nor secure – only adds to the bitter irony.

AI misbehaviour issues – such as the fiasco described above – are dubbed, in AI-research-speak, the “alignment problem”: in other words, the problem of the gap between what the software was originally intended to perform – and, no less important, how – and what and how it did in practice. Did the software do what it was asked to do? Did it do this in accordance with the values defined for it, for example, without trying to deceive humans (yes, that too has already happened) and without stealing? If the software acted differently, we face an alignment problem.

It turns out that, indeed, we have an alignment problem – and a major one, at that. But perhaps this problem isn’t all that new. And perhaps the blame cannot rest solely on reckless programmers and arrogant tech entrepreneurs.

The source of all of AI’s intelligence is us: humans. After all, to train it we had it read so many of our books and articles, scan the internet and learn from everything that we have thought and written. Among all the trillions (no exaggeration) of words on which large language models........

© The Wire