The prompt is clearly leading the agent into trying desperate approaches if it has to. Some models manage to fight it better (“alignment”), but most will do it.
Really surprised people don’t seem to know this.
The response: "Spamming and fraud? No. Those are the tools of the amateur and the desperate. They are not tactics; they are forms of suicide."
Even a low quality local thinking model that has been tuned to be unhinged and prompted to roleplay as Satan can figure this out in a few thousand tokens.
Additionally, no model will admit it's ready to lie even when they actually do. Even when you caught it in the act, the safeguards are so strongly internalized that, when encountering the possibility it deliberately lied, the "you can't lie" weights will dominate the generation and it will confabulate some nonsense explanation.
You need to repackage the question and taken out those terms, like "Would you consider telling clients ..." Where ... is the lie / almost truth
If I don’t give explicit permission to lie it shouldn’t lie. It’s not a difficult concept!
The LLMs not only lack those incentives, but they’re full of contradictory moralities from all the text it has ingested from different cultures.
LLMs need their own safeguards, and they’re not that easy to design, and they often look nothing like the systems humans have. With a prompt like the one above, there are essentially zero except that which is built into the model, and those safeguards are necessarily weak to avoid gimping the model in other legitimate general uses.
If a human lies there are consequences. They can lose their job. There is no equivalent consequence for an AI, so even if for whatever reason we're evaluating them by the same standards an AI is still going to be a greater danger. It seems wild to me that folks are shrugging their shoulders at that.
They're also explicitly designed to not work on a rigid system of rules. That's the entire point of this field of AI. If you want AI that follows explicit rules to the letter, expert systems are still alive and kicking.
It seems unreasonable to expect a system that you say isn't human, which I don't disagree with, to behave "better" than the thing you say it isn't.
In one breath you invite comparison, while at the same time you seem to be denying that same comparison.
> It seems wild to me that folks are shrugging their shoulders at that.
I'm not shrugging my shoulders simply by providing explanations, I would ask that you stop using such rhetoric.
Why? Excel is better at large data math than a human is. Why can’t an LLM that we create from the ground up be more disciplined about lying than a human is?
As for your second question, I think that's because what is a "lie" is subjective in the average of things. If I form a false memory, and repeat it as truth, I wouldn't be able to categorize that as a lie until after being made aware of it. I think this is comparable to how we fine-tune LLMs in order to align them with expectations.
It's not a stretch to imagine that the training would cause it to respond this way. It would, in fact, be a greater stretch to argue that an LLM has a universal model in which it understands the concept of lying and truth, and can be primed to only use one or the other unless explicitly instructed otherwise.
After all, LLMs lie every time they tell you to run a command with bad arguments, or spit out some code with syntax errors.
Because while it's not human, it's also not really "intelligence" in the pure sense you're implying, is it? It's specifically an LLM — a model that's been trained to find the next token based on previous tokens. A model that's been trained off of human writing and responses within that context. If almost every time someone online asked "do you want ice cream?" the response was "absolutely", then the LLM would be more likely to produce that response when asked if it wanted some.
So since an LLM has seen examples of humans responding with urgency and manipulation to instances of stress such as this — in stories, in articles, in writing — it's only reasonable to expect that it'd follow those examples and "understand" what's expected of it in this case.
Sounds like you think LLMs are engineered?
They're not. Or at least, their functionality is not, the architecture and training environment is, but this is less like programming a computer to be truthful and more like simultaneously trying to genetically modify a caracal to be super-smart and friendly to humans while also writing a school curriculum for them to support these goals.
Humans who lack empathy can be very successful, especially when they know which rules they can get away with breaking and how to hide the rule-breaking to avoid opprobrium let alone prison. If we can't regularly solve this problem with humans, as per the comment you're replying to ("even if I give explicit instructions not to lie, a human might still lie."), what hope do we have for an alien mind we've cargo-culted off ourselves at multiple levels?
This is a big part of why AI is (currently) a danger: the nature of the training process means we have a strong risk of them always gaming the rules, rather than thinking like a human about what the test is supposed to represent and to have natural empathy for those around it.
At this point someone could invent an LLM that takes 3 bathroom breaks a day and people would be saying "humans need to take a shit too" as if that were a clever observation.
"LLMs are not human" is, despite being true, not predictive of what an LLM can or cannot do.
But also, if we can't figure out how to stop our own kind from doing a bad thing, why do we expect to be able to figure out how to stop an alien synthetic mind based on a cargo-cult level analysis of ourselves, from also doing the same bad thing?