Artists can poison their pics with deadly Nightshade to deter AI scrapers | Hacker News Reader

Artists can poison their pics with deadly Nightshade to deter AI scrapers

theregister.com

30 pointsmoijafcor25 comments

I’ve asked an artist if artists should be worried about the newest generative AI capabilities. This is his (translated) response:

> Artists? Pencil laborers, more like.. I am in favor of using AI in visuals. It will eliminate a lot of mere decorators, and won’t even slightly affect the artists. I hope AI as a technology has the same effect on the world of ART as the invention of photography had: it got rid of a lot of empty landscape copiers. Impressionism was born shortly after that. See, I believe many cursed photography, but Monet never saw it as a problem.

This. It’s no different than the impact word processors had on typists. The idea that low-level sign painting is “art” is silly.

And to the extent AI encroaches on “real”, it just enables artists to operate at a higher level, much like word processors allow authors to focus on craft rather than typing.

This is an extremely elitist take.

Obviously you're never worried about AI/Computers/Immigrants (underline your "reason" of choice) if you're in the top 1% of your field. It's like people who developed a bunch of their own libraries and then worked as a google Architect for 10 years giving everyone else advice on job hunting.

From a utilitarian point of view, I see his reason for saying it (not a great loss for society if the bottom 50% of some industry falls out), but it's still a very privileged take (some would even say tone-deaf)

Van Gogh, Manet, Gauguin would beg to differ! Being a true artist doesn’t imply being in the top 1%, having privilege or any economical reward at all.

…and that won’t change with AI as well, there will always be artists who will be recognized as great only after their death.

Adversarial optimization as a field is one that is very hyped by the media, but I am entirely unimpressed by it and most ML researchers I know are likewise bearish.

None of these techniques are general to all models (obviously) nor do they typically withstand adding a very small amount of noise to the image.

These don't actually work. Adversarial optimization is largely model-specific.

Curious if you have a resource to read that shows this?

This was the most surprising aspect to me, that such a solution would apparently work robustly? Does the paper cover this?

They claim it's not future proof, but does it actually work now?

Does it actually work now if you img2img your input files, or do any type of detection that a tool like Nightshade may be in use?

none of these solutions (especially for images) are robust at all. fundamentally this is not a technology that is going to work

What even is the point of using the word ‘deadly’ here? So obnoxious.

Same but for intelligence

I found this headline annoying at a visceral level. This wasn't click bait, it was click hate. I vote we just remove the title of this post, leave it blank.

Its a play on words, with the deadly nightshade plant. Poison is also the same play on words.

And folks, this is why the "softer arts" are just as much needed, cause even simple literary artistic devices are /whoosh.

Like apostrophes, right?

Pretty sure most people understood the reference to the plant - it's just silly phrasing given the context.

dupe/previously: https://news.ycombinator.com/item?id=39058428

There is already another submission on the front page, one that actually uses the original source and have hundreds of comments.

Why do we need another? Can we get these merged?

https://news.ycombinator.com/item?id=39058428

Any AI art/video/photography/music/text/etc generator company whose generates revenue needs to add watermarks to let the public know its AI generated(it's embedded with metadata specific to each company). This should be forced via legislation in all countries.

If they don't then whatever social network or other services where things can shared/viewed by large groups to millions & are posted publicly need to be labeled "We can not verify veracity of this content."

I want a real internet ..this AI stuff is just triple fold increasing fake crap on the Internet and in turn / time our trust in it!

The real internet died long ago. Even if the content you see was made by a real human, chances are it's been reposted a million times by bots, quite commonly strategically in order to evoke certain emotions.

As with anything, the services are geared towards the populous and there 85% just don't care.

We will have to continue to make our own casinos for real black jack and hookers.

They will care once scammer use deepfakes to scam them out of their monies and it hits as many of the populous as current scamming does. How many people still answer unknown numbers today (ten years from now) vs ten years ago?

If it collapses Twitter, Facebook and Tiktok due to a deluge of gibberish spam that decreases the value proposition for users, and users are forced back onto moderated forums or curated news sites, I see that as a huge win for humanity.

Yeah, no as another commenter said if AI spam is the downfall of twitter tik tok and facebook that will be a net benefit for all of society and the internet. The amount of misinformation and garbage ragebait I see my older relatives consuming on those sites is astounding.

Can someone explain how this differs from Glaze, which is apparently "defensive" while Nightshade is "offensive"? Their goals sound the same to me, maybe I missed something.

Based on their marketing, Glaze is “defensive” because it prevents training a model to have a coherent idea of what your style is, preventing people from generating art “in the style of andai”. Nightshade doesn’t protect you, but it behaves as “wrong” training data and makes it harder to train models in general. (I don’t claim that any of this works as advertised though.)

Reply on news.ycombinator.com