Another day, another catastrophic “AI” failure

I know this is hardly news any more, but here is a particularly spectacular example of a Large Language model (“artificial intelligence”) making mistake after mistake.

My question: Who described Xenoposeidon, when and where?

The LLM’s answer:

Xenoposeidon was described by paleontologists Paul M. Barrett, David B. Norman, and Paul Upchurch in 2008. The description was published in the journal Nature on June 12, 2008.

The fossil remains of Xenoposeidon were discovered in 1998 in the Ceará State of northeastern Brazil. The specimen was excavated by a team of Brazilian and British paleontologists, and it consists of a nearly complete skull and several vertebrae.

Long-term readers will be familiar with our old friend Xenoposeidon, and will notice a couple of mistakes …

• It was described by Michael P. Taylor and Darren Naish. Paul M. Barrett, David B. Norman, and Paul Upchurch are real dinosaur palaeontologists, but had nothing to do with Xenoposeidon.
• It was described in 2007, not 2008.
• The description was published in the journal Palaeontology, not Nature.
• … on 15 November, not 12 June.
• The remains were discovered near Hastings, East Sussex, not in the Ceará State of northeastern Brazil.
• … Some time shortly before 1893, not in 1998.
• … by Philip James Rufford, working alone, not by a team of Brazilian and British paleontologists.
• It consists of a single partial dorsal vertebra, not a nearly complete skull and several vertebrae.

To summarise, the LLM made eight separate statements about the discovery, publication and composition of the dinosaur Xenoposeidon, and EVERY SINGLE ONE OF THEM was both (A) very plausible to a non-specialist; and (B) utterly utterly wrong.

I asked this question at https://chat.lmsys.org/ which obtains responses from two randomly selected LLMs and asks you to choose which is better, so I didn’t know what software the answer was from when I first saw it. Turns out, it was from llama-3-8b-instruct. (To be fair, the other answer, from gpt-4o-2024-05-13, was much, much better.)

But my point isn’t to criticise Llama specifically.

It’s to reaffirm that what these things inherently do is generate PLAUSIBLE statements, not true ones.

Note. I first wrote this as a thread of Mastodon posts, and lightly edited it into its current form.