Recently, there has been considerable interest in large language models: machine learning systems which produce human-like text and dialogue. Applications of these systems have been plagued by persistent inaccuracies in their output; these are often called “AI hallucinations”. We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs. We distinguish two ways in which the models can be said to be bullshitters, and argue that they clearly meet at least one of these definitions. We further argue that describing AI misrepresentations as bullshit is both a more useful and more accurate way of predicting and discussing the behaviour of these systems.
Carl T. Bergstrom, 13 February 2023:
Meta. OpenAI. Google.
Your AI chatbot is not hallucinating.
It’s bullshitting.
It’s bullshitting, because that’s what you designed it to do. You designed it to generate seemingly authoritative text “with a blatant disregard for truth and logical coherence,” i.e., to bullshit.
I confess myself a bit baffled by people who act like “how to interact with ChatGPT” is a useful classroom skill. It’s not a word processor or a spreadsheet; it doesn’t have documented, well-defined, reproducible behaviors. No, it’s not remotely analogous to a calculator. Calculators are built to be right, not to sound convincing. It’s a bullshit fountain. Stop acting like you’re a waterbender making emotive shapes by expressing your will in the medium of liquid bullshit. The lesson one needs about a bullshit fountain is not to swim in it.
Imagine still not realizing what a useful skill bullshitting is. Literally, hundreds of millions of people are professional bullshitters. So many people go do bullshit every day, all day. Having a machine that can produce the same or better bullshit than them frees them from suffering through doing all that bullshit. I can’t think of something that is more bullshit than pretending like there is no benefit from automating the bullshit out of our lives.
Except it’s not really being automated out of our lives, is it? I find it hard to imagine how increasing the rate at which bullshit can be produced leads to a world with less bullshit in it.
It saves us from doing the bullshit that we are currently suffering through right now. It is rapidly getting better at doing it as well. 2 years ago, the best llms were preschool level, now they are high-school level, or arguably better.
Sure, if we were already living in a world where nobody had any reason to produce BS, then it would be weird if we needed machines that could do it. The fact of the matter is though that we all use BS daily because it makes our lives better. The code that runs most apps you use could be way better, but it’s not, it’s BS. It gets the job done. The customer support people are making BS that at least gets you what you need. The teachers wade through hours of BS to find the same spelling mistakes, grammar mistakes, logical errors. You think they like doing that BS? Nope.
BS machines have relieved so many people of so much BS and its only just the very beginning. This is the worst the BS machines will ever be, and it is improving at a blinding fast speed. The sooner people realise this, the sooner they can start trying to imagine the implications. Nearly everyone complaining about how useless they are, always point to the worse instances of outdated one-shot responses. They never talk about how awful the Claude Opus agent workflows are. That’s because the people who know what that stuff is realize what we are on the cusp of. An intelligence revolution is happening, some people have seen it already, many people will see it soon. Denying it is like scoffing at the idea that people will ever want their own computer.
Thank you, that is such a beautiful and liberating vision for the future!
imagine a world with 10^^^ times more bullshit, but all the human bullshiters are unemployed! Able to do what they want (except pay rent).
Problem is that it doesn’t automate away the bullshit in our lives. We’re creating even more bullshit that we’re forced to deal with online and at our jobs. Sure we can use the bullshit generator to respond to bullshit, but how do you know what’s bullshit in the first place, are you going to ask your bullshit generator to sort that out for you as well?
Control the language and you control the thought. I pitched a fit when “hallucinate” was put forward by the tech giants to describe their LLMs’ falsehoods, and it mostly fell on deaf ears in my circles. Hallucinating isn’t what these things do. They bullshit.
The use of anthropomorphic language to describe LLMs is infuriating. I don’t even think bullshit is a good term, because among other things it implies intent or agency. Maybe the LLM produces something that you could call bullshit, but to bulshit is a human thing and I’d argue that only reason that what the LLM is producing can be called bullshit is because there’s a person involved in the process.
Probably better to think about it in terms of lossy compression. Even if that’s not quite right, it’s less inaccurate and it doesn’t obfuscate difference between what the person brings to the table and what the LLM is actually doing.
“confabulate” is, imo, the closest we have (i don’t remember who originally used this analogy, unfortunately)
Tbh half the reason teachers are struggling to tell is that bullshitting your way through things is a time-honored element of human thought and word. I myself bullshitted my way through most of my schooling until my professional degree. Really the AI is just imitating us better than we realized.
We further argue that describing AI misrepresentations as bullshit is both a more useful and more accurate way of predicting and discussing the behaviour of these systems.
Bullshit is a far better description for sure.
