In this article, we will look at why this problem of hallucinations happens with LLMs and the techniques that can help make LLMs more dependable for answering.
The fabricated citation problem has a direct twin in financial planning. Clients now arrive having asked an LLM about NPS withdrawal tax treatment, 80C limits, or whether an endowment policy is a good investment — and get fluent, structurally perfect answers with invented section numbers and outdated thresholds. The model has learned what a tax rule looks like without knowing the current one.
LLM’s only know what they’ve been trained to know compounded by their context. You stumbled onto some text that aligns with a chapter in a book where those same words were followed by X Y Z.
It didn’t plagiarize the way you intended isn’t lying.
When they say they completed a task, they probably did. It just wasn’t what you consider complete or correct. You’re lying about what you want.
There’s always the phenomenon that people try to force s4177y models into roles they aren’t equipped for. Then equate that to “the state of AI”
We could bicker about all of that for days but the fact of the matter is that being wrong isn’t lying. Lying requires the intent to deceive. If you want to ascribe that to an LLM, we shall mock you.
Well, "Lying" makes for a catchy title, but I don't think it's what the author means. They say so directly: "We can't outright call this behavior lying, but it comes pretty close to a lie… lying usually implies an intention to deceive. A hallucination doesn't establish that intention."
Not hallucinations at all. Diagnosing LLMs with human pathology is a categorically incorrect assessment of machine limitations.
The article is spot on but stops short of correcting a nonsensical misnomer.
The fabricated citation problem has a direct twin in financial planning. Clients now arrive having asked an LLM about NPS withdrawal tax treatment, 80C limits, or whether an endowment policy is a good investment — and get fluent, structurally perfect answers with invented section numbers and outdated thresholds. The model has learned what a tax rule looks like without knowing the current one.
That sounds like a training corpus issue.
Not to mention many problems with output are actually issues with use prompting and their real knowledge about what they are asking.
if rlhf training makes models reluctant to admit uncertainty, what changes when being wrong costs them less?
LLM’s only know what they’ve been trained to know compounded by their context. You stumbled onto some text that aligns with a chapter in a book where those same words were followed by X Y Z.
It didn’t plagiarize the way you intended isn’t lying.
When they say they completed a task, they probably did. It just wasn’t what you consider complete or correct. You’re lying about what you want.
There’s always the phenomenon that people try to force s4177y models into roles they aren’t equipped for. Then equate that to “the state of AI”
We could bicker about all of that for days but the fact of the matter is that being wrong isn’t lying. Lying requires the intent to deceive. If you want to ascribe that to an LLM, we shall mock you.
Well, "Lying" makes for a catchy title, but I don't think it's what the author means. They say so directly: "We can't outright call this behavior lying, but it comes pretty close to a lie… lying usually implies an intention to deceive. A hallucination doesn't establish that intention."