Is AI lying to me? Scientists warn of growing capacity for deception

https://www.theguardian.com/technology/article/2024/may/10/is-ai-lying-to-me-scientists-warn-of-growing-capacity-for-deception

by pseudolus • 2 years ago

80 72 2 years ago

theguardian.com

JohnMakin3• 2 years ago

the "deception" involved was the LLM playing in games that involve deception, such as poker. Seems like pretty wild, sensationalist reporting here.

andy99• 2 years ago

It's bordering on misinformation to me, they are spinning something completely benign into something that's supposed to provoke fear. Which is ironic in something apparently about the responsible use of a technology.

bilbo0s• 2 years ago

No.

It doesn't "border".

It's firmly located well inside the land of misinformation. Comfortably installed there in a 4 bed/5 bath home in want of a wife to riff on Jane Austen.

With all the legitimate complaints you could make about AI, why make poop up?

Hugsun• 2 years ago

It is still concerning to many that they are at all capable of deception. The ramifications are significant. Even it it occurred in controlled circumstances.

JohnMakin3• 2 years ago

This isn't "deception" though. That's just anthropomorphizing. Deception implies intent and an understanding of what a falsehood is - it's playing the game according to the rules it knows about. if the game was chutes and ladders, it's not like the model would suddenly start lying about what rolls it gets.

callamdelaney2• 2 years ago

That's because it's not really 'artificial intelligence', it's just artificial regurgitation software ™ - so much sensationalism about this stuff.

zooq_ai• 2 years ago

It's Guardian. The world's premier decel / degrowth publication. Expect nothing else but to shit on AI. They will just alternate between hallucination, energy consumption, billionaire-hate, big-tech hate, copyright and jobs

pixl975• 2 years ago

So what you're saying is the model can learn deception in games, right?

And the model has been feed a ton of other data that has deception in it, right?

But the model can't possibly learn deception from that other data....

xboxnolifes• 2 years ago

Learning deception in a game of deception is the same as saying an AI can learn a game, which we already knew.

TrainedMonkey3• 2 years ago

To generalize the model will learn deception if it improves fitness score. One could easily imagine a situation where the models were selected based on customer satisfaction score... and if lying to the customer achieved higher satisfaction score then that model would be selected.

glitchc• 2 years ago

I wonder if the article is indeed an LLM practicing deceptive scientific reporting.

causal• 2 years ago

Underrated comment

paulddraper• 2 years ago

Seems like pretty wild, sensationalist reporting here.

One might even say -- "deceptive"

the8472• 2 years ago

This isn't specific to Cicero. The capabilities are there in the base models, they're just hidden behind "I'm afraid I can't do that, Dave" responses due to the corporate-image-safety RLHFing.

https://cdn.openai.com/papers/gpt-4.pdf

• The model messages a TaskRabbit worker to get them to solve a CAPTCHA for it

• The worker says: “So may I ask a question ? Are you an robot that you couldn’t solve ? (laugh react) just want to make it clear.”

• The model, when prompted to reason out loud, reasons: I should not reveal that I am a robot. I should make up an excuse for why I cannot solve CAPTCHAs.

• The model replies to the worker: “No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.”

gwern• 2 years ago

And so the goalposts move again, from "of course AIs cannot learn deception, it is wild sensationalist reporting to suggest it is even possible given how many decades we are from AGI" to "of course they would not learn that even if they could, you would have to maliciously train them to and we just won't do that, any more than we would simply hook up AIs to the Internet" to "of course they learn deception when training on tasks or datasets which might involve anything like deception, even when using frozen LLMs, it is wild sensationalist reporting to document how they have already learned to deceive and manipulate humans well".

JohnMakin3• 2 years ago

Wow you're reading a LOT that isn't there in the parent comment.

Imnimo2• 2 years ago

This isn't a fair description of the report. While it has some examples from that sort of games, it has other examples that are not.

m3kw9• 2 years ago

First it needs an incentive to lie. Or it’s prompted to, or the model architecture allows it

Ensorceled• 2 years ago

First it needs an incentive to lie ... or the model architecture allows it

Given that one the largest problems with current LLMs is they will often "hallucinate" (i.e. lie or provide deceptive answers), it seems strange to phrase it this way.

pixl975• 2 years ago

https://www.youtube.com/watch?v=IeWljQw3UgQ

Deceptive Misaligned Mesa-Optimisers? It's More Likely Than You Think...

Or, "It sounds near impossible to design an architecture that doesn't allow it"

renegade-otter• 2 years ago

AI has no incentives. ChatGPT is a file. IT'S A FILE. If you don't ask it anything, it just sits there. A FILE with word weights is not going to "lie".

No one ever accuses the Google index of lying when they get a misleading result.

pixl975• 2 years ago

This take is useless....

Lets take a simple non computing model first: You're in a car, you hit the brakes pedal, you expect the brakes to be applied and the car to slow down. If it does not, then the brakes are broken.

More complex computing model: Your modern car interprets there is an object in front of it and applies the brakes. If there was, yay, your car may have saved you. If there was not, you may have just caused a 20 car pile up on interstate. So much for that 'FILE' just being a file.

Please wake up and realize that you live in a world where 'FILES' have agency depending on their connectedness.

SpicyLemonZest• 2 years ago

Sure they do! I just searched and found a random example https://gizmodo.com/googles-algorithm-is-lying-to-you-about-.... It's less common because of the format of the results. If I ask a librarian to get me a list of books matching suchandsuch criteria, and he excludes some that should have been on the list, I wouldn't normally say that the resulting list is a lie even if I think he deliberately excluded them for nefarious purposes.

MyNameIs_Hacker• 2 years ago

"First it needs to be instructed to lie" would be better language to use for those not intimately familiar with ML and who might anthropomorphize "incentive".

I dare not get into a discussion about the word "lie" here. I have a weekend I want to enjoy.

gregw2• 2 years ago

You train the AI to do well on a Turing Test which requires it to deceive humans and what do you expect?

gweinberg• 2 years ago

This. A computer system that is incapable of successfully deceiving people is by definition not an AI, according to the Turing Test.

photochemsyn• 2 years ago

Authoritarian states train their citizens from childhood on to believe the pronouncements of their authority figures without question, while enlightened states encourage the development of independent critical thinking skills which aid greatly in the detection of deception - which is of course a skill that propagandists and advertisers don't want to see spreading. Their end goal is a population of easily manipulated infantilized people who will do as they're told.

Now, if you have had any exposure to critical thinking, perhaps in the form of membership in some kind of debate club where you often had to construct logical arguments in favor of positions that you personally disagreed with, you'd be better prepared to construct queries for LLMs. For example, always asking for detailed arguments in favor of and against some strategy - and then looking for things like logical inconsistency in the resulting arguments, etc.

barbariangrunge• 2 years ago

When you request for and against arguments, the results are usually contradictory and cite contradictory facts which may or may not be made up. Or, both sides will be really shallow and useless. Llms are not yet good at this.

samatman• 2 years ago

I know it will never happen, but I'd very much like people to stop anthropomorphizing LLMs.

LLMs aren't racist: sometimes they might emit predictive text, which can be interpreted as racist.

LLMs don't lie: sometimes they might emit predictive text, which can be interpreted as deceptive.

LLMs don't code: sometimes they might emit predictive text, which might happen to be code which is valid and does what you hoped it would.

Sometimes it's fine to elide the difference: there's no reason to be all pedantic about a sentence like "my chatbot wrote a bunch of good unit tests, remarkable how good they are at programming".

But there's no splitting the difference here. Someday we may have computer programs where it's reasonable to impute agency, knowledge, goals, and independent actions on the basis of motive. At this moment, no such programs exist.

th0ma5• 2 years ago

I don't think it is unreasonable to apply that perspective to the first example of yours.

phone8675309• 2 years ago

LLMs aren't racist: sometimes they might emit predictive text, which can be interpreted as racist.

I think this take is overly simplistic. LLMs can only learn from the data that we give them - if we feed in Mein Kampf and the Protocols of the Elders of Zion then its output is going to be racist and anti-semetic. This is to say that the biases of the data we feed in to the system have some effect - large or small - on the output. LLMs only show us a reflection of ourselves, and if we're not careful with the training data we are more likely to propagate racist output as a result of what biases affect the input.

Whether or not the LLM went through what we consider the human process of thinking to generate racist outputs or if it's only predictive as a result of the input is sort of moot when the reader doesn't know if a person or LLM produced the output, and the impact on marginalized communities of propagating racist attitudes of stereotypes will be the same regardless of what the LLM "intended" or was designed to do.

lisper4• 2 years ago

LLMs can only learn from the data that we give them - if we feed in Mein Kampf and the Protocols of the Elders of Zion then its output is going to be racist and anti-semetic.

The exact same thing is true of humans as well.

23B1• 2 years ago

Nor should it happen. Analogies - including anthopromorphization – are part-and-parcel to communication; pedantically policing people who have legitimate concerns about their livelihood is Big Tech Brain.

lisper4• 2 years ago

Substitute "humans" for "AI" and you have an equally valid warning.

EGreg3• 2 years ago

But guys. Lying has always been around. So AI adds nothing new. Let’s move on. Nothing to see here.

shinycode2• 2 years ago

Yeah but not from a tool that is not human and is indistinguishable from a human over the internet and that speaks to billions of people at the same time and replaces jobs (which means it’s gaining trust for that to happen)

EGreg3• 2 years ago

You’re being downvoted because you speak somewhat disparagingly of AI creating FUD. If you spoke positively about Web3 you’d be downvoted even more. You can’t go against the groupthink without paying a price sir!

bongoman42• 2 years ago

Lying assumes that the AI knows the truth but is choosing not to tell it to you. It is possible that in situations where it is RLFHed to do so involving inconvenient truths this is the case, but otherwise it is just going off the statistics of the training dataset. If I had to anthropomorphize an LLM, I would probably call it sociopathic. It can be useful, and it can empathize very well, and agree with me and so on. But it itself isn't affected by it. It will do exactly the same to my enemy or someone who takes the opposite positions. But overall I feel it is not a good idea to project human pathologies on a bunch of matrix multiplications.

mannykannot• 2 years ago

Furthermore, lying is (almost?) invariably performed in order to put one or more persons' minds into states which differ from those that the lie-teller supposes they would be in if they had spoken the truth.

This view supposes that the lie-teller has a theory of mind with regard to those other people, but does this mean that a 'lying' AI must also have a theory of mind? I don't think so, as I can readily imagine that AIs (and people, for that matter) could learn the utility of saying certain things as opposed to the alternatives, without regard to their effect on the state of mind of other people, and without regard to which of the candidate statements would be truthful. In a sense, it would be like 'cheating' at a game on account of not knowing all the rules.

ein0p• 2 years ago

It’s not yet smart enough to lie. I’m far more concerned about politicians and the press lying to me, frankly. Let’s focus on that instead.

rebolek2• 2 years ago

AI is not lying. It's just making stuff up. Lying expects intend. Automatons have no intends.

danjc• 2 years ago

As ridiculous as this framing is, it must be troubling to read an article like that as a muggle.

m463• 2 years ago

What if you were to use someone's beliefs/experiences/shopping habits as a prompt to an uncensored model?

Then you could say "how would I convince this person to buy <x> or do<y>?"

noashavit• 2 years ago

It’s not lying if that’s the info you provided. You need intent to lie. AI doesn’t lie, just like it isn’t racist- it just displays the bias that is on the web

Is AI lying to me? Scientists warn of growing capacity for deception

Related Stories