The intellectual meat of the article seemed to be that the author found a connection between Harry Frankfurt's definition of [bullshit](https://www2.csudh.edu/ccauthen/576f12/frankfurt__harry_-_on...) and what we've been calling "hallucinations". Seemed reasonable but it wasn't exactly evident to me why I should care. We're replacing one word with another word. so what.
the best the authors could do was argue that there were some mysterious set of negative social consequences that we'd get from using "hallucinations," since it implies that the models are mistaken or misguided in an attempt at representing the truth, rather than attend to the fact that all the models are doing is generating text that _seems_ truthy, and has no awareness or attention to what the truth _actually_ is. One could probably just read the last paragraph and get all they needed to from the paper:
>We object to the term hallucination because it carries certain misleading implications ...
>
>Calling chatbot inaccuracies ‘hallucinations’ feeds in to overblown hype about their abilities among technology cheerleaders, and could lead to unnecessary consternation among the general public. It also suggests solutions to the inaccuracy problems which might not work, and could lead to misguided efforts at AI alignment amongst specialists. It can also lead to the wrong attitude towards the machine when it gets things right: the inaccuracies show that it is bullshitting, even when it’s right. Calling these inaccuracies ‘bullshit’ rather than ‘hallucinations’ isn’t just more accurate (as we’ve argued); it’s good science and technology communication in an area that sorely needs it.
So their argument is effectively:
1. it's wrong, or at least, Frankfurt's definition of "bullshit" fits better.
2. it could mislead the public or alignment researchers
On 1), I'm willing to concede that hallucinations might be the wrong term. But words have a life of their own, and it's too late to go back now. At least to late for one paper to change anything.
On 2), it seems plausible, but, regrettably, the paper spends less than a paragraph talking about it! None of the claims they're making are that complicated, and yet for some reason they fail to provide even a few falsifiable hypotheses about the main implication of their argument. Ok, sure. I can see that the term "hallucination" might be misleading. But are you really going to publish a paper just so you can "Uh, actually ..." everybody and argue that we're using the wrong word? How much is it going to mislead the public? The public is always mislead - why does this particular instance matter? How can we tell? Are alignment researches really going to be misled by choice of terminology? If they are, could you suggest a mechanism? How much would they be misled? If they are misled, why does it matter? What does "misled" even mean? How do we measure it? I could go on.
I'd like to imagine that a paper warning about the implications of using incorrect terminology would go into some detail about their claim and explore how those implications might play out - this paper's publication might have more to do with its title and topic than it does any important claims or results.
the best the authors could do was argue that there were some mysterious set of negative social consequences that we'd get from using "hallucinations," since it implies that the models are mistaken or misguided in an attempt at representing the truth, rather than attend to the fact that all the models are doing is generating text that _seems_ truthy, and has no awareness or attention to what the truth _actually_ is. One could probably just read the last paragraph and get all they needed to from the paper:
>We object to the term hallucination because it carries certain misleading implications ... > >Calling chatbot inaccuracies ‘hallucinations’ feeds in to overblown hype about their abilities among technology cheerleaders, and could lead to unnecessary consternation among the general public. It also suggests solutions to the inaccuracy problems which might not work, and could lead to misguided efforts at AI alignment amongst specialists. It can also lead to the wrong attitude towards the machine when it gets things right: the inaccuracies show that it is bullshitting, even when it’s right. Calling these inaccuracies ‘bullshit’ rather than ‘hallucinations’ isn’t just more accurate (as we’ve argued); it’s good science and technology communication in an area that sorely needs it.
So their argument is effectively: 1. it's wrong, or at least, Frankfurt's definition of "bullshit" fits better. 2. it could mislead the public or alignment researchers
On 1), I'm willing to concede that hallucinations might be the wrong term. But words have a life of their own, and it's too late to go back now. At least to late for one paper to change anything.
On 2), it seems plausible, but, regrettably, the paper spends less than a paragraph talking about it! None of the claims they're making are that complicated, and yet for some reason they fail to provide even a few falsifiable hypotheses about the main implication of their argument. Ok, sure. I can see that the term "hallucination" might be misleading. But are you really going to publish a paper just so you can "Uh, actually ..." everybody and argue that we're using the wrong word? How much is it going to mislead the public? The public is always mislead - why does this particular instance matter? How can we tell? Are alignment researches really going to be misled by choice of terminology? If they are, could you suggest a mechanism? How much would they be misled? If they are misled, why does it matter? What does "misled" even mean? How do we measure it? I could go on.
I'd like to imagine that a paper warning about the implications of using incorrect terminology would go into some detail about their claim and explore how those implications might play out - this paper's publication might have more to do with its title and topic than it does any important claims or results.