SafetyInsights.org

This paper challenges the label of AI hallucinations – arguing instead that these falsehoods better represent bullshit.

That is, bullshit, in the Frankfurtian sense (‘On Bullshit’ published in 2005), the models are “in an important way indifferent to the truth of their outputs”.

This isn’t BS in the sense of junk data or analysis, but in their indifference to accuracy in the pursuit of mimicking human speech.

LLMs aren’t lying or hallucinating

First, when LLMs make factual errors, they aren’t lying because lying requires an intention to deceive. That is, “to make a believed-false statement to another person with the intention that the other person believe that statement to be true”.

In contrast, LLMs “are simply not designed to accurately represent the way the world is”.

Calling the LLM falsehoods ‘hallucinations’ is an inapt metaphor: it “anthropomorphises the LLMs”, giving the impression that they are misrepresenting the world and describing what they see. This is misleading because “LLMs do not perceive, so they surely do not ‘mis-perceive’”.

Also, they argue that the process leading to falsehoods isn’t unusual, because the “very same process occurs when its outputs happen to be true”. Therefore, attributing the falsehoods to hallucinations allows the AI “creators to ‘blame the AI model for faulty outputs instead of taking responsibility for the outputs themselves’”.

Calling the AI bullshit confabulation is also problematic according to the authors, as this, too, “risks anthropomorphising the LLMs”, suggesting that there “is something exceptional occurring when the LLM makes a false utterance”, which isn’t the case for LLMs.

LLMs’ outputs are best understood as “bullshit” in the Frankfurtian sense.

Bullshit, in the Frankfurt sense, is defined as a “lack of concern with truth, or an indifference to how things really are”.

In contrast, LLMs are “designed to produce text that looks truth-apt without any actual concern for truth”.

Instead, LLMs use “reams of available text and probability calculations in order to create seemingly-human-produced writing”. If LLMs have a primary goal, then it’s to produce human-like text by estimating the likelihood that a particular word will appear next.

Therefore, they are not “designed to transmit information” or “represent the world at all”, but designed to “provide a normal-seeming response to a prompt, not to convey information that is helpful to their interlocutor”.

Soft vs hard bullshit

The authors draw a distinction between soft and hard bullshit.

Soft bullshit: This is “bullshit produced without the intention to mislead the hearer regarding the utterer’s agenda”.

At the least, ChatGPT is a soft bullshitter, since it doesn’t “intend to convey truths” and “is indifferent to the truth value of its utterances”. Even if it does have intentions, it’s intentions are “aimed at being convincing rather than accurate”.

Hard bullshit: This is bullshit with “intention to mislead the audience about the utterer’s agenda”.

ChatGPT may also be a hard bullshitter, depending on whether we can ascribe it to have ‘intentions’.

If its primary function is to imitate human speech, and if this is intentional, then “ChatGPT is attempting to deceive the audience about its agenda,” specifically “it’s trying to seem like something that has an agenda, when in many cases it does not”.

Conclusion

They argue that the choice of terminology matters a lot:

·         Calling LLM falsehoods ‘hallucinations’ can mislead and misinform the public, policymakers and other parties

·         It leads to the “wrong attitude towards the machine when it gets things right” and suggests “solutions to the inaccuracy problems which might not work”

·         Also, trying to “hook the chatbot up to some sort of database, search engine, or computational program” to improve accuracy, “doesn’t work very well either”, because LLMs aren’t designed to transmit information, but only to produce human-like convincing text

·         In contrast, describing AI misrepresentations as bullshit is “both a more useful and more accurate way of predicting and discussing the behaviour of these systems”

·         This is because calling their outputs bullshit “avoids the implications that perceiving or remembering is going on in the workings of the LLM”

·         And “indifference to the truth is extremely dangerous”, because “by the mindlessly frivolous attitude that accepts the proliferation of bullshit as innocuous, an indispensable human treasure is squandered”

Hence, “calling these inaccuracies ‘bullshit’ rather than ‘hallucinations’ isn’t just more accurate…it’s good science and technology communication”.

Ref: Hicks, M. T., Humphries, J., & Slater, J. (2024). ChatGPT is bullshit. Ethics and Information Technology, 26(2), 1-10.

Shout me a coffee (one-off or monthly recurring)

Study link: https://link.springer.com/content/pdf/10.1007/s10676-024-09775-5.pdf

LinkedIn post: https://www.linkedin.com/pulse/chatgpt-bullshit-ben-hutchinson-4cg5c