Anthropic apologizes after one of its expert witnesses cited a fake article hallucinated by Claude in the company's legal battle with music publishers
Anthropic’s attorney from Latham: fictitious source was “an honest mistake” due to Latham’s using Claude to format the citations.
Can we normalize not calling them hallucinations? They’re not hallucinations. They are fabrications; lies. We should not be romanticizing a robot lying to us.
Pretty engrained vocabulary at this point. Lies implies intent. I would have preferred "errors"
Also, for the record, this is the most dystopian headline I've come across to date.
If a human does not know an answer to a question, yet they make some shit up instead of saying “I don’t know”, what would you call that?
If you train a parrot to say “I can do calculus!” and then you ask it if it can do calculus, it’ll say “I can do calculus!”. It can’t actually do calculus, so would you say the parrot is lying?
that's a lie. They knowingly made something up. The AI doesn't know what it's saying so it's not lying. "Hallucinating" isn't a perfect word but it's much more accurate than "lying."
"Fabrications" seems ok to me
Bullshit.
This is what I've been calling it. Not as a pejorative term, just descriptive. It has no concept of truth or not-truth, it just tells good-sounding stories. It's just bullshitting. It's a bullshit engine.
I like fabrication going forward. Clearly made up, doesn't imply intent
The word hallucination has zero implication of intent whatsoever. Last time I checked hallucination is an entirely involuntary experience, regardless of the context the word is used in.
They are called hallucination in computer science not “to romanticize” it. It is called that because the output is totally random from the perspective of the input. If there is no logical path from input to the output, it is similar to a human hallucinating. Human sees no actual weird visual stimuli that results in them hallucinating a dragon, therefore the input info from their eyes has no bearing on what they imagine is actually there.
This is different from “fabrication” in that the AI intentionally creating fake info based on your input request would not be a hallucination, because there would be a relationship between input and output.
While you say you prefer “fabrication”, the word fabrication actually implies some intent that is absent from what we are referring to as AI hallucinations
I meant that fabrication doesn't imply intent as "lies" would.
It seems like you use the hallucinations term correctly, when output has no relation to input.
In this case, as in many numerous others, the Ai took input of "cite a source" and did as output cite a source as requested, but invented the content of the source. It fabricated, which means to make up, create.
Fabricate does not imply intent to deceive, where lie does.
I will agree that if the output is purely unrelated to the input, hallucination is still fine, but is absolutely a romanticized term when we're referring to this computer generated code... It's literally personification.
Everything an LLM outputs is hallucinated. That's how it works. Sometimes the hallucination matches reality, sometimes it doesn't.
Emm no... Why?
