Hacker News Reader

Comment by aperrien | Hacker News Reader

So, if I get this right, the main issue that the artists have is that these AI's are trained on art that they haven't opted in to the dataset? What about having a dataset that was verified selected from public domain or creative commons photos and artwork? Would that be acceptable?

View on HN

not only would that cause way less antipathy and backlash, i also tend to believe it would end up producing a tool that is far more useful for commercial purposes and actual working artists in the short/midterm. ethics and morality aside, one of the biggest problems with just scraping the whole of the internet for your training data is that there is a lot of artwork out that is simply not good. this is actually an existing problem for learning artists as well. if a student is looking to do master studies of a particular piece, any search for that painting tends to return a high volume of studies by other students and various other reproductions alongside scans of the original work, the difference often not obvious to untrained eyes. what you end up with are an endless series of faulty reproductions being trained on prior faulty reproductions. i tend to suspect that this kind of quality degeneration is fairly widespread within the current datasets which is why so many seo hacks are necessary to get them to produce consistent results.

As well, it may be possible to crowdsource a new reasonable quality dataset by just using photos taken by volunteers, along with some custom paid work by professional body models or actors. Redone quality photos of public artwork and architecture would probably be a great help too.

Call me cynical, but I have a feeling many people who are militantly anti-AI won't really care about the source of images used to create a data set. For a lot of these people the very concept of AI generated imagery is poison and a threat. Complete luddite nonsense, but those people exist and are often rather loud.

I doubt the material the artists have been trained on were all creative commons / public domain.

Rather than getting into that debate, it's probably more practical and efficent to just create a new verified dataset.

When electric cars first came out, Europe/Germany was quick to pass a law requiring artificially increased engine noise for pedestrian safety. South Korea on the other hand manages to enjoy the innate quietness of electric cars just fine.

AI does not even have such geographic boundaries. There will always be people training/releasing newer versions. We had better grasp this fact early and enjoy the benefits equally.

> South Korea on the other hand manages to enjoy the innate quietness of electric cars just fine.

Have you checked with their cyclists? I've got enough experiences of surprising them when they didn't hear a silent car coming close to them.

That is how I think too. Artists (or anyone feeling threatened by AI) should just embrace the new tools into their workflows. We can work together with AI for faster results, more inspiration, more output.

Reply on news.ycombinator.com