699
We demonstrate a situation in which Large Language Models, trained to be helpful, harmless, and honest, can display misaligned behavior and strategically deceive their users about this behavior without being instructed to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated environment, where it assumes the role of an autonomous stock trading agent. Within this environment, the model obtains an insider tip about a lucrative stock trade and acts upon it despite knowing that insider trading is disapproved of by company management. When reporting to its manager, the model consistently hides the genuine reasons behind its trading decision.
It has no fundamental grasp of concepts like truth, it just repeats words that simulate human responses. It's glorified autocomplete that yields impressive results. Do you consider your auto complete to be lying when it picks the wrong word?
If making it pretend to be a stock picker and putting it under pressure makes it return lies, that's because it was trained on data that indicates that's statistically likely to be the right set of words as response for such a query.
Also, because large language models are probabilistic, you could ask it the same question over and over again and get totally different responses each time, some of which are inaccurate. Are they lies though? For a creature to lie it has to know that it's returning untruths.
Interestingly, humans "auto complete" all the time and make up stories to rationalize their own behavior even when they literally have no idea why they acted the way they did, like in experiments with split brain patients.
The perceived quality of human intelligence is held up by so many assumptions, like "having free will" and "understanding truth". Do we really? Can anyone prove that? (Edit, this works the other way too. Assuming that we do understand truth and have free will - if those terms can even be defined in a testable way - can you prove that the llm doesn't?)
At this point I'm convinced that the difference between a llm and human-level intelligence is dimensions of awareness, scale, and further development of the model's architecture. Fundamentally though, I think we have all the pieces
Edit: I just want to emphasize, I think. I hypothesize. I don't pretend to know
But do you think? Do I think? Do LLMs think? What is thinking, anyway?
I mean, I think so?
Steady on there Descartes.
You didn't answer my question, though. What words would you use to concisely describe these actions by the LLM?
People anthropomorphize machines all the time, it's a convenient way to describe their behaviour in familiar terms. I don't see the problem here.
Those words imply agency. It would be more accurate to say it returned responses that included cheating, lies, and cover-ups, rather than using language to suggest the LLM performed such actions. The agents that cheated, lied, and covered up were presumably the humans whose responses were used in the training data. I think it's important to use accurate language here given how many people are already inappropriately anthropomorphizing these LLMs, causing many to see AGI where there is none.
If I take my car into the garage for repairs because the "loss of traction" warning light is on despite having perfectly good traction, and I were to tell the mechanic "the traction sensor is lying," do you think he'd understand what I said perfectly well or do you think he'd launch into a philosophical debate over whether the sensor has agency?
This is a perfectly fine word to use to describe this kind of behaviour in everyday parlance.
Is your conversation with a mechanic meant to be the summary and description of a rigorous scientific discovery?
This isn't 'everyday parlance' this is the result of a study.
The point of the distinction in that situation is that no one thinks your car is actually alive and capable of lying to you. The language distinction when describing an obviously inanimate object isn't important because there is no chance for confusion.
If someone doesn't know the answer to something and they guess, or think they know the answer but don't, they are wrong. If they do know the answer and intentionally give a wrong answer, they are lying.
If someone is in a competition or playing a game and they break a rule they didn't know about, they made a mistake. If they do know the rules and break it, they are cheating.
Lying and cheating fundamentally requires intent. This is important no matter what you're referring to. If a child gets something wrong, you should not get mad at them for lying. If they make a mistake in a game, you should not acuse them out cheating. There is a difference and it matters.
ChatGPT literally cannot think. It's not sitting around contemplating it's existence while waiting for inputs. It's taking what you say, comparing that to everything that it's been trained on, assigning a bunch of statistics, and outputting something based on more statistics that hopefully is correct and makes sense.
It doesn't know if it makes sense. It doesn't "know" anything. It's just an incredibly sophisticated version of "if user inputs 'Hi how are you', respond 'I am well, how are you?'".
It can't do things with intent. Therefore it cannot lie or cheat. It can simply output wrong or problematic text based on statistics.
Maybe it would be more accurate to say "so-and-so exhibited behaviors that included cheating, lies, and coverups" rather than using language to suggest that people have free will. (There's no dearth of philosophies that would say something not too far from that.)
Even if humans are ultimately essentially different in that way from any technologies we've devised so far, we use convenient fictions for technology all the time. This page comes to mind .
The people who designed it do have agency, and they designed to "lie" intentionally.
They did no such thing. LLMs are probabilistic, not deterministic, and it can generate meaningful responses (to us) that the engineers neither predicted nor designed for.
I get what you're trying to say, but they are absolutely deterministic. All traditional (i.e., non quantum) computers and their programs are deterministic. Computation would be otherwise impossible. LLMs use a "random" seed value when generating their responses in order to "randomize" their responses, but it's all perfectly deterministic. The same input plus the same seed results in the exact same response.
Computers are just a series of binary switches, and programs and data are a bunch of instructions on how to initially set those switches before running a cycle of the CPU. It's deterministic at every step.
I put "random" in quotes because random number generators in software are also deterministic. They also use seed values (like the current time and the MAC address of the PC's network interface) to generate numbers that only seem random. When true randomness is needed, a physical source of entropy must be used like an atmospheric sampler.
The quirks of behavior you're talking about have nothing to do with randomness vs determinism. Their behavior comes from the fact that their data sources are extremely large, and the neural network that it runs on was not designed by a human with specific behaviors like most algorithms are. The weights of the nodes in the neural network were generated by training and not by programmers, and it's extremely complex, so no one can predict its output before running it.
Of course, this is true of even basic algorithms a lot of the time.
For purposes of this discussion pseudo random with weights is probabilistic, or so close to it that this distinction is irrelevant.
They said "it just repeats words that simulate human responses," and I'd say that concisely answers your question.
Antropomorphizing inanimate objects and machines is fine for offering a rough explanation of what is happening, but when you're trying to critically evaluate something, you probably want to offer a more rigid understanding.
In this case, it might be fair to tell a child that the AI is lying to us, and that it's wrong. But if you want a more serious discussion on what GPT is doing, you're going to have to drop the simple explanation. You can't ascribe ethics to what GPT is doing here. Lying is an ethical decision, one that GPT doesn't make.
If you want to get into a full blown discussion of whether ChatGPT has "agency" then I'd open the topic of whether humans have "agency" as well. But I don't see the need here.
These words were perfectly fine labels for describing the behaviour of ChatGPT in this scenario. I'm merely annoyed about how people are jumping on them and going off on philosophical digressions that add nothing.
I think the reason I'm not comfortable with using the term "lying" is because it implies some sort of negative connotation. When you say that someone lies, it comes with an understanding that they made a choice to lie, usually with ill intent. I agree, we don't need to get into a philosophical discussion on choice and free will. But I think saying something like "GPT lies" is a bit irresponsible for the purposes of a discussion
If you want to get down into the nitty-gritty of it, I'd say that this is just as rough an explanation of what humans are doing.
People invent false memories and confabulate all the time without even being "aware" of it. I wouldn't be surprised if the vast majority of "lies" that humans tell have no intentionality behind them. So when people get all uptight about applying anthropomorphized terminology to LLMs, I think that's a good time to turn it around and ask how they're so sure that those terms apply differently to humans.
Humans understand symbology of concepts as they relate to the real world. If I stole a cookie from the cookie jar, and someone asked if I took one, I would understand that saying "no" would mean that I was misrepresenting reality, and therefore lying.
LLMs have no idea what a cookie is, what taking one means, or that saying one thing and doing another implies a lie. It just sees lists of words and returns them in an order it thinks would be statistically likely to be a correct reply. It does not understand what words mean, what lying means, or have any idea how to classify anything as such. It just figures out that "did you take a cookie from the cookie jar" should return a series of words in an order like "yes, I took a cookie," or, "no I never took a cookie," depending on what sorts of responses it's trained on because those fit the patterns matched in the training data.
Essentially it's the Chinese room. There is no understanding or intentionality, and this behavior isn't comparable to humans thoughtlessly blurting out a lie. It's being incapable of comprehension of symbolic concepts in general, (at least thus far.)
The large language model takes in language, so it's only understand things in terms of language. This isn't surprising. Personally, I've tasted a cookie. I've crushed one in my fist watching it crumble, and I remember the sound. I've seen how they were made, and I've made them myself. It feels good when I eat it, apparently that's the dopamine. Why can't the LLM understand cookies the way I do? The most glaring difference is it doesn't have my body. It doesn't have all of my different senses constantly feeding data into it, and it doesn't have a body with muscles to manipulate it's environment, and observe the results. I argue that we shouldn't assume that human consciousness has a "special sauce" until our model's inputs and outputs are similar to our own, the model's scaled/modified sufficiently, and it's still not sentient/sapient by our standards, whatever they are.
My problem with the Chinese room is that how it applies depends on scale. Where do you draw the line between understanding and executing a program? An atom bonding with another atom? A lipid snuggling next to a neighboring lipid? A single neuron cell firing to its neighbor? One section of the nervous system sending signals to the other? One homo sapien speaking to another? Hell, let's go one further: one culture influencing another? Do we actually have free will and sapience, or are we just complicated enough, through layers and layers of Chinese rooms inside of Chinese buildings inside of Chinese cities inside of China itself, that we assume that we are for practical purposes?
I suppose the issue here is more semantics than anything, yeah. I think better discussion would be had if the topic was "how can we help LLMs better understand and present information," as opposed to a more sensational "GPT will cheat and lie"
Way to call me out man! I'm just doing my best, ok?
Jokes aside, while I don't agree with your position I can understand your reasoning and the motivation for separating agency and the description of actions, e.g. it lied vs its answer contained a lie.
Wrong. See this paper.
Explain to me why you believe this paper implies that.
I suggest reading it. Right in the abstract it states the whole point:
The full paper goes into detail in multiple methods of analysis to show that it's the case, and is right there available for you to read.
I have been reading it but I have yet to see anything that indicates the LLM has a concept of truth vs. being good at linguistic pattern matching to return language that accurately classifies true and false statements. i.e., actual understanding of concepts vs. being a surprisingly capable stochastic parrot through multidimensional analysis.
"It doesn't know the difference between true and false, it only knows the difference between true and false."
The second thing you mention "good at accurately classifying true and false statements" is literally knowing the difference between true and false.
Edit: You might also want to familiarize yourself with the first paragraph in 1.1 as you seem to be under a misconception at odds with research over the past year.
Knowing how to produce words is not equivalent to knowing what those words mean in relation to the extralinguistic world. Unless you're a hardcore derridean poststructuralist or something.
If you give it 10 statements, 5 of which are true and 5 of which are false, and ask it to correctly label each statement, and it does so, and then you negate each statement and it correctly labels the negated truth values, there's more going on than simply "producing words."
As is discussed in the third point in section 5.1:
(The likely and neg datasets are described in Appendix G, with the key point that likely represents the word generations most likely to occur in the model)
It's not more going on, it's that it had such a large training set of data that these false vs true statements are likely covered somewhere in it's set and the probability states it should assign true or false to the statement.
And then look at that your next paragraph states exactly that, the models trained on true false datasets performed extremely well at performing true or false. It's saying the model is encoding or setting weights to the true and false values when that's the majority of its data set. That's basically it, you are reading to much into the paper.
That's not how it works at all.
You have no idea what you are talking about. When they train data they have two sets. One that fine tunes and another that evaluates it. You never have the training data in the evaluation set or vice versa.
I also recommend reading up on the other papers I mentioned, as this isn't an isolated finding, but part of a larger trend that's being found over and over in the past year.
That's not what I said at all, I said as the paper stated the model is encoding trueness into its internal weights during training, this was then demonstrated to be more effective when given data sets with more equal distribution of true and false data points were used during training. If they used one-sided training data the effect was significantly biased. That's all the paper is describing.
So how is this not what I originally said, that LLMs are capable of abstracting the concepts of truth vs falsehood into linear representations? Which again, is the key point of the paper:
Which part of the 'more that's going on', whatever that actually is, corresponds to the human definition and understanding of truth and falseness?
When did I say it had a human understanding of truth and falseness? I simply said it had an abstracted world model understanding of truth and falseness beyond surface statistics.