No, OpenAI's image-making DALL·E 2 doesn't understand some secret language

AI text-to-image generation models are all the rage right now. You give them a simple description of a scene, such as "a vulture typing on a laptop," and they come up with an illustration that resembles that description. That's the theory, anyway. But developers who have special access to OpenAI's text-to-image engine DALL·E 2 …

[Filippo]

> Hilton believes that it shows the model doesn't have a secret understanding of some unknown language, but instead demonstrates the random nature of AI.

But it's not random, though. It's true that the words don't appear to have a specific meaning, but at the same time Daras' weird prompt really does produce images of birds eating bugs - consistently so, not randomly. Hilton's prompt doesn't produce bugs specifically, but it does produce animals and not, I dunno, cars.

To me, the results show the black box nature of AI, more than its randomness. We just don't know why it has that behavior, but it's not random.

The reason this could be an important distinction is that this behavior could be the basis for adversarial attacks. What if, by wearing a t-shirt with an apparently nonsense word on it, I could consistently fool face-recognition?

[ThatOne]

It's just a kind of adversarial attack: Clearly those gibberish words trigger something in the AI's internal workings. We've seen it before, although in a different form.

We foolishly expect that the AI only heeds proper instructions in English from the guy who is supposed to issue instructions, but clearly the system is open to unwanted suggestions.

[Claptrap314]

"Computer!"

[Michael Wojcik]

It might be more precise to say we know why in general, but not specifically.

Deep-learning architectures are a stack of neural nets, and typically most of the discrimination is done by convolutional layers. Convolutions are good for detecting signals probabilistically: they try to maximize the area of the intersection between the incoming signal and a reference signal.

If you stack a lot of convolutions and you don't provide the equivalent of a null hypothesis in your training – a bucket that the model can dump low-probability inputs into – then even a weak signal will be "recognized" as something from the training.

Basically, the input runs through the stack and has to end up in some category. DALL-E I think (can't be bothered to look it up) is a transformer architecture, so it's all based on attention, which gives it a certain amount of "memory" or context. So the context in which the nonsense words appear will affect what categories get chosen.

In short, what's happening here isn't surprising, and in fact is precisely what you'd expect from a large transformer DL model. And it is a fine example of an inexplicable model, as you suggested. This is a big problem for the deployment of DL architectures, as many researchers have discussed at length.

Note that there are other approaches to ML and "AI" (insofar as "AI" means anything) that are explicable or at least more transparent. Deep-stack ANN architectures like DL are particularly resistant to explication and other types of analysis.

[b0llchit]

Undecided between gibberish and divinity

Why exactly DALL·E 2 associates those images with the gibberish inputs remains unclear.

The nature of statistical inference machines is statistics. The term "gibberish" is no statistical term but a manifestation of the nature of statistics.

I'd suggest to deal with it without assigning divinity. If you really need to assign divinity, then, by all means, worship me (your monetary donations are greatly appreciated).

[OhForF']

AI libel

"What happens when you're defamed by an internet giant's code? Can you sue? Whom do you sue?"

You can always fling a sue-ball at whoever you want to.

"Is there a resonable chance to win a case for slander" is a different question.

I don't see why you should not be able to sue an internet giant but if they can show the statement was generated by an algorithm (that was not created specifically to create content dissing you) it will be hard to infer malicious intent.

Things might change if the internet giant continues to post the message after it was made aware it is untrue.

There seem to be some issues if you want to sue the AI.

Lately articles if AI's can be recognized as inventors or be sued keep springing up.

When did it become the latest fashion to treat Artificial Idiocy as sentient?

[ThatOne]

Re: AI libel

> I don't see why you should not be able to sue an internet giant

Don't you? First, suing costs money, a lot of money which the victims most likely can't spare. Second, you and your ambulance chaser don't have the proverbial snowball's chance against the army of the internet giant's top lawyers. And even if you manage to secure a legal eagle, your money will eventually run out and you'll be forced to fold. It will cost the internet giant back-of-the-sofa money but it will totally bankrupt you, for nothing.

_{(Didn't downvote you though.)}

[Yet Another Anonymous coward]

Re: AI libel

And the small print that you didn't sign says the case will be held in California, and if when you say "suing American corporation" as the reason for visiting the border guard doesn't let you in .

[Michael Wojcik]

Re: AI libel

Also, this is the US, where the bar for libel is considerably higher than in many other countries. And not just because of the First Amendment.

(Personally, I think this is a Good Thing, but this probably isn't the place to debate that.)

[emfiliane]

Re: AI libel

The law around this might become very interesting in a hurry, if an argument holds that the AI is akin to an employee acting in an official capacity as a spokesperson of the company, or that it's akin to a manufacturing defect which injured someone.

The mere fact that they threw it out there without sufficient quality control doesn't sway juries.

[ThatOne]

Re: AI libel

> if an argument holds that the AI is akin to an employee acting in an official capacity

Then you create a subsidiary to put all those "employees" in. Let them sue that, it's just a postal box...

[Androgynous Cow Herd]

Re: AI libel

Who do you sue when defamed by AI?

Simples: Stephen Thaler

[elsergiovolador]

A dog

AI is like a dog on a leash. If you let it bite someone, you are responsible.

Not sure why there is even a debate about this, perhaps because as usual big corporations don't want to take responsibility.

[Anonymous Coward]

big corporations don't want to take responsibility.

I bet they want the earnings though ...

[vogon00]

Re: A dog

"AI is like a dog on a leash. If you let it bite someone, you are responsible.

I wish it would be like that. AFAIK, and under English law, it's easy to identify the person who libels you as they are the ones that publish the libel. If it's a corporate body like the tabloid press publishing stuff, then it's up to them to defend their statements - assuming one afford (Or find 'pro-bono') someone willing to represent you.

Only rich people can sue for liable. (Not mentioning the recently concluded 'JD vs AH Round 1' b.s. to avoid legitimising the storm-in-a-teacup created by the media).

In these days of syndication etc., who is liable? The entity in control of the AI who created the erroneous shite that they published in the first place, the entity that re-tweeted it, the website that published the re-tweet or the programmer who created the AI? Good luck with that, as the website publisher/re-tweeter etc. will pass the buck upstream (Assuming they disclose their source!).

We're all familiar with 'Sorry, the computer says no'....but such decisions are (currently) usually algorithmic and can be evaluated/reverse engineered if necessary to correct error. If you can find someone able to 'back trace' the operations/factors leading to an AI decision...I'll show you a bullshitter!

My personal opinion is that AI is good for rules-bound repetitive simple tasks (Like image processing leading to a go/nogo decision or something basically statistical etc.) but as soon as you let it do something that involves actual perception or judgement you're gonna be in a world of hurt, either because the original programmer was crap, the dataset it was trained on was crap, the interpretation of it's output was crap, and the shit output will be accepted as gospel by those desperate to join the AI buzz.

Sorry for any typo...I can confirm that the sharper the kitchen knife, the less it hurts....especially after two or three glasses of vino rosso.

[Michael Wojcik]

Re: A dog

If you let it bite someone, you are responsible.

As a question of morality, or of law?

Not sure why there is even a debate about this

Perhaps because pat slogans don't generally resolve complex questions to everyone's satisfaction.

[Anonymous Coward]

I think it's patently obvious

That "machine" learning is most certainly not equal to "human" learning.