AI hallucinations are getting worse - and they're here to stay (New Scientist)

[highplainsdem]

https://www.newscientist.com/article/2479545-ai-hallucinations-are-getting-worse-and-theyre-here-to-stay/

Earlier thread about the worsening accuracy: https://www.democraticunderground.com/100220267171

From New Scientist yesterday:

-snip-

These models work by repeatedly answering the question of “what is a likely next word” to formulate answers to prompts, and so they aren’t processing information in the usual sense of trying to understand what information is available in a body of text, says Bender. But many tech companies still frequently use the term “hallucinations” when describing output errors.

“‘Hallucination’ as a term is doubly problematic,” says Bender. “On the one hand, it suggests that incorrect outputs are an aberration, perhaps one that can be mitigated, whereas the rest of the time the systems are grounded, reliable and trustworthy. On the other hand, it functions to anthropomorphise the machines – hallucination refers to perceiving something that is not there [and] large language models do not perceive anything.”

Arvind Narayanan at Princeton University says that the issue goes beyond hallucination. Models also sometimes make other mistakes, such as drawing upon unreliable sources or using outdated information. And simply throwing more training data and computing power at AI hasn’t necessarily helped.

The upshot is, we may have to live with error-prone AI. Narayanan said in a social media post that it may be best in some cases to only use such models for tasks when fact-checking the AI answer would still be faster than doing the research yourself. But the best move may be to completely avoid relying on AI chatbots to provide factual information, says Bender.

The question of course, is why we should find error-prone AI in any way acceptable. Especially considering everything harmful about generative AI besides the fact that it can't be trusted to provide the right answer.