Conversation

Notices

1. Embed this notice
   I Value the Goose (paco@infosec.exchange)'s status on Monday, 28-Jul-2025 04:29:07 JST I Value the Goose

   A friend sent me the story of the LLM deleting a database during a code freeze and said "it lied when asked about it." I assert that a generative AI cannot lie. These aren't my original thoughts. But if you read Harry Frankfurt's famous essay On Bullshit (downloadable PDF here), he makes a very reasoned definition of bullshit. And this paragraph near the end of the essay explains why an LLM cannot lie.

   It is impossible for someone to lie unless he thinks he knows the truth. Producing bullshit requires no such conviction. A person who lies is thereby responding to the truth, and he is to that extent respectful of it. When an honest man speaks, he says only what he believes to be true; and for the liar, it is correspondingly indispensable that he consider his statements to be false. For the bullshitter, however, all these bets are off: he is neither on the side of the true nor on the side of the false. His eye is not on the facts at all, as the eyes of the honest man and of the liar are, except insofar as they may be pertinent to his interest in getting away with what he says. He does not care whether the things he says describe reality correctly. He just picks them out, or makes them up, to suit his purpose.

   And that's a generative artificial intelligence algorithm. Whether generating video, image, text, network traffic, whatever. It has no reference to the truth and is unaware of what truth is. It just says things. Sometimes they turn out to be true. Sometimes not. But that's irrelevant to an LLM. It doesn't know.

   #bullshit #ai #llm #genai

   • Alexandre Oliva likes this.
   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Monday, 28-Jul-2025 04:29:02 JST Paul Cantrell

     in reply to

     • GLC

     @glc
     Discussion aside, what a fantastic photo!

   • Embed this notice
     I Value the Goose (paco@infosec.exchange)'s status on Monday, 28-Jul-2025 04:29:03 JST I Value the Goose

     in reply to

     • RalfMaximus
     • Ryan J. Yoder
     • GLC

     @ryanjyoder
     All models work fundamentally in the same way: predicting a series of output tokens based on a series of input tokens. If you don’t understand this basic implementation mechanism, the rest of this conversation is inaccessible.

     It doesn’t store facts. It doesn’t have a representation of “true” or “false.” It isn’t a database. It splits written text into tokens and does colossally huge, environmentally damaging, and fabulously expensive “training” on that data using billions of parameters to arrive at a statistical model of tokens that follow other tokens. The model can then be queried to produce statistically likely replies to inputs.

     Given an input like “tell me a lie about the capital of France” the most statistically improbable reply is “the capital of France is Paris.” Other replies like “wear a seatbelt” are also super improbable. The size of these models and the probabilities they work with are really difficult to get one’s head around. But it returned a statement that was a probabilistically likely reply to that input. That’s all it did.

     When models make up legal cases that don’t exist, books that don’t exist, programming APIs that don’t exist, etc, they are simply outputting likely results. Text that fits the probability distribution of their input data. That’s why it is not a “bug” when an LLM bullshits. It’s not an error. It is working as designed.

     There is nowhere to report to an LLM company the factually incorrect outputs its model produced because there is nothing they can do with that. It is working as designed.

     @ralfmaximus @glc

     Alexandre Oliva likes this.
   • Embed this notice
     GLC (glc@mastodon.online)'s status on Monday, 28-Jul-2025 04:29:03 JST GLC

     in reply to

     • RalfMaximus
     • Ryan J. Yoder

     @paco @ryanjyoder @ralfmaximus

     A general comment ... from another angle, more or less parallel -

     The question is whether the following is a picture of a woman or a picture of a tree. Opinions vary, and are strongly held.

     LLMs by design strip away all meaning. Similar techniques have been in use since the I Ching (more recently, describedl by Shannon 1948 with transition probabilities). It's a bit like dehydrated water - just add water.

   • Embed this notice
     Ryan J. Yoder (ryanjyoder@techhub.social)'s status on Monday, 28-Jul-2025 04:29:05 JST Ryan J. Yoder

     in reply to

     • RalfMaximus
     • GLC

     @ralfmaximus @glc @paco

     If you didn't know the implementation of a specific model would there be any way to determine if it was capable of telling a lie?

     Alexandre Oliva repeated this.
   • Embed this notice
     Ryan J. Yoder (ryanjyoder@techhub.social)'s status on Monday, 28-Jul-2025 04:29:06 JST Ryan J. Yoder

     in reply to

     • GLC

     @glc @paco
     If an llm can't lie, then it seems true by only a technical definition. I can ask Meta ai basic questions about geography and get correct answers reliably. And I can ask it to lie about basic geography and I can reliably get incorrect answers.

   • Embed this notice
     RalfMaximus (ralfmaximus@mastodon.social)'s status on Monday, 28-Jul-2025 04:29:06 JST RalfMaximus

     in reply to

     • Ryan J. Yoder
     • GLC

     @ryanjyoder @glc @paco

     There are billions geography facts on the internet. This includes fictional data sets too from popular SF/fantasy.

     There are millions of human<->human interactions where geography facts are discussed, some of them accurately, many of them not.

     Train a LLM on all that. Ask it geography questions.

     The results will be the most likely next chain of text based on all those weighted interactions. That's it. There's no lying, no truth. It's making predictions.

   • Embed this notice
     Ryan J. Yoder (ryanjyoder@techhub.social)'s status on Monday, 28-Jul-2025 04:29:07 JST Ryan J. Yoder

     in reply to

     @paco
     Maybe this is semantics but Meta AI appears to be able to lie.

   • Embed this notice
     GLC (glc@mastodon.online)'s status on Monday, 28-Jul-2025 04:29:07 JST GLC

     in reply to

     • Ryan J. Yoder

     @ryanjyoder @paco

     Nope.

     Just another illustration of the original point.