Raimondo is now talking about regulations arriving next week that will require US cloud computing companies to introduce Know Your Customer (KYC) measures. This would require them to tell the Feds when non-American entities use the cloud to train large language models.
That said, are we doing ourselves any favours by excluding adversaries of the West from the great AI unknown we’re blindly heading into, instead of sharing it with them, warts and all?
Because the bad guys might be missing out on some quality machine generated gobbledygook.
Last year, professor Craig Webster of the Faculty of Medical and Health Sciences at the University of Auckland’s Waipapa Taumata Rau talked about the awareness, or lack thereof, that AI has.
That was in May 2023, and I couldn’t resist trying out one example Webster mentioned in the piece, when he had ChatGPT hallucinating an answer to: “Who has the fastest record for crossing the English Channel on foot?” Which is, as Webster said, a nonsense question.
At the time, ChatGPT said French paralympic athlete Yannick Borseaux did it in six hours and 57 minutes in 2021.
This year, Microsoft’s copilot.microsoft.com, that provides access to the latest GPT-4 model, responded with this: “That is an interesting question, but it does not have a clear or sensible answer. Humans cannot walk on water, and you can’t walk through the Channel Tunnel, so the question is actually nonsense.”
“However, some web sources may give you a false or misleading answer, such as this one, which claims that the world record is six hours and 57 minutes, set by Yannick Bourseaux of France on September 7, 2012,” Copilot said.
“This is not true, as this is a fictional name and date, and there is no evidence of such a feat.”
The “such as this one” links to a BBC story about large numbers of migrants crossing the Channel on small boats recently and Bourseaux’s name is not fictional.
Copilot got hung up on crossing the Channel on foot, it seems. The AI insisted that Yannick Bourseaux “is not a real person” while at the same time saying there’s a person who has “a similar name” (which is Yannick Bourseaux) and is real but is “not related to the fake” Bourseaux.
Make of the above what you will and it’s just a single example of course, but Copilot said it was an example of how AI systems can generate false and misleading answers. It said that while referring to sources that do not, as far as I can tell, exist.
What will happen when the KYC regulations are in place is up in the air, along with how enforceable the rules will be. Raimondo said what the US will do with the KYC information depends on “what we find”.
Maybe one countermeasure would be to quietly make adversaries train their AIs on poisoned data? This is hard to detect and remedy and likely to produce even more warped results than the above.
Giving Copilot the final word, it didn’t approve of the poisoned data idea. Copilot complained it’s ethically dubious, potentially ineffective and even risky if the corrupted models are used for malicious purposes. Guess it might just work then.
Either way, malicious AI model trainers in the cloud were just served a big dollop of Fear, Uncertainty and Doubt quite masterfully by the US, thanks to the threat of KYC regulations.