But even without filing lawsuits, artists have a chance to fight back against AI using tech. MIT Technology Review got an exclusive look at a new open source tool still in development called Nightshade, which can be added by artists to their imagery before they upload it to the web, altering pixels in a way invisible to the human eye, but that “poisons” the art for any AI models seeking to train on it.
Excellent. This is exactly the kind of clever thinking we need to stop major corporations from stealing everyone’s creative works for their own further gain. I hope we can develop these poisons further, to the point of making these “AI” tools entirely useless.
Get permission, or get poisoned.
Thom Holwerda,
It’s a clever idea, however I predict this will not work out long term. As it becomes more widespread the technique will no longer be able to fly under the radar and countermeasures will be developed. A “Hacker” need only masquerade as an artist to generate samples and train an AI to detect said “poison” and subsequently remove it.
The whole reason nightshade works at all is because algorithms see the changes that are invisible to humans, but this can be fixed. The fact that these changes are so easily visible to AI will be Nightshade’s achilles heel.
Alfman,
This is a “cat and mouse” problem, but the cat is much more clever in this case. And we already have excellent models:
https://www.reddit.com/r/StableDiffusion/
How are they going to put the genie back?
I don’t think this is solving the right problem, and will probably be left as a footnote in the generative AI history.
Basically there are two concerns of artists wrt. to the new generative methods:
1) They can “learn” from these artists
2) They can “replace” these artists
I would say (1) is unavoidable, and (2) is overblown. Let me explain.
The method seems to be aimed at preventing using artists’ work in training data. Normally this data would be used for learning (a) new concepts, or (b) new styles. So for example, after ingesting 100s of “dogs”, and a few Picasso paintings, you can ask prompts like “a dog on a chandelier, in Picasso style”, and get a good enough imitation of that artist.
First, modern large scale ML organizations will not just randomly image search for dogs, and use them for training (some do, many did in the past, but this is no longer the common practice). And even if they did, it is unlikely they will run into 100s of samples from this method, it is more likely they will run into regular pictures. So this will not cover cased (a).
And if a model is “fine tuned” for a specific style, it would be obvious that the training data is “poisoned”. What the model owner would do, would be a simple photoshop filter to clean any noise, including hidden ones. Since this is an active iterative process, the “poison” would not work for case (b) either.
That leaves the (2), which is machine learning models replacing artists. This is a very long discussion, but it seems like AI will “augment” proper artisans, giving them tools to become much productive, and only replace those who do repetitive work, like legal assistants searching though 100s of caseworks, “fivers” designing a simple logo, and ironically mechanical turk users tagging machine learning training data. If you know your craft, there is nothing to be worried about, at least for now.
And if you are really interested in the state of the art in image generation, especially free ones, there is a nice tutorial here:
https://octoml.ai/blog/the-beginners-guide-to-fine-tuning-stable-diffusion/
Also check out:
https://github.com/AUTOMATIC1111/stable-diffusion-webui
And don’t be afraid to use free colab resources:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Online-Services
sukru,
Agreed. Like it or not, AI is likely here to stay (and keep improving).
Workers are going to have to adapt. Naturally some will refuse over ideological reasons, but it will come at their own expense as they loose out on efficiency gains. I expect there will be niches where the sentimental value of a human touch is more important than raw efficiency, but on the whole I believe this will be the exception rather than the norm. Consumers and businesses are more price sensitive than they care about worker well being. So it’s in our best interest to embrace more efficient tools, because if we don’t someone else surely will.
Other than the “for now” part, I’m less sure about the long term stability and middle class economic health. Technological revolutions have rendered large numbers of employees redundant throughout history. Their saving grace tends to be new job openings that they must retraining for. This creates anxiety for many, but isn’t necessarily bad as long as the opportunities are good. What does concern me however is that historically efficiency gains tend to benefit owners far more than the workforce who get compensated disproportionately less, even as their efficiency goes up.
Just for simplification, let’s say that some new technology allows one worker to do the work of two former workers. There might even be a pay increase for remaining workers, but if the employer is able to layoff half the workforce (or alternately double production without doubling salaries), then macro-economically this is a bad deal for workers. When efficiency gains are used to give owners a larger share of GDP, then it ends up accelerating social wealth gaps. The wealth gap is already extremely skewed today. If nothing is done to ensure that the efficiencies of the next technological revolution help the middle class, then I think there’s a very real possibility that it could lead to an increase in poverty overall.
Alfman,
Tat is a fair point. But as you mentioned adaptability is the keyword.
We had the same transition when ATMs were first introduced. Tellers no longer needing to count cash, they were able to do higher value jobs. Which, against the expectations, led to more bankers being hired, since they could not open more profitable branches. Downside? If your only contribution was counting cash, you were out of a job (like manual elevator operators, horse buggy drivers, and those horses as well).
I am on the “hopeful” camp wrt. AI. But unfortunately, if not prepared well, yes, some people will be unable to adapt.
sukru,
I assume you meant “could now open”, haha.
Did the switch to ATMs end up end up increasing the number of banking jobs relative to the size of the banking industry? I don’t know the answer, but I am curious if you have a source for it?
This is relevant to many types of jobs today, like grocery stores replacing cashiers with self checkout. As a consumer I am not a fan of this BTW because prices go up (I am genuinely shocked at how much food prices are increasing) while quality of service becomes worse. Same deal with the local hardware store. Once autonomous vehicles become more mature, long haul truck drivers are at serious risk of loosing their jobs too. A local Wendys restaurant is trialing a configuration whereby customer line up in front of screens to take orders with no humans in the customer area to take orders or even help at all without getting out of line and going to the kitchen. And just for the record, prices did not go down. Soon enough the same fate may come to those working the kitchen. My personal feelings about this aside, I think it’s very fair and necessary to ask about what actually happens when disruptive technologies displace several millions of jobs?
I’m not trying to persuade or dissuade anyone on the merits of automation. I am not anti-tech or anti-progress, However I think we need to put the breaks on the cavalier attitude that tech fixes everything and society doesn’t needs to worry about displacing jobs because new jobs are around the corner. While it is true that historically we’ve been able to offset technological redundancy with more new jobs, there isn’t a fundamental economic law that makes this so. Our markets are maturing and are run by corporations seeking profits above all else. If they can get the job done with fewer workers, they will and that’s exactly what technological automation promises to do for them. We’ve been very fortunate to have access to new markets as job creators since this has allowed us to completely ignore the role that technology has in displacing workers. But…there is no guarantee a constant stream of new markets can go on indefinitely. Once new markets substantially dry up, the effects of technologically induced job losses will be economically far more dire than anyone is expecting because the masses are ill-prepared to handle such conditions and laissez-faire policy will fail us.
End of easy buck for content creators? How sad…