杏吧原创

IT ALL started when I typed a perfectly reasonable prompt into Midjourney, one of several apps on the market that can create an image based on text. 鈥淪kull space laser dinosaur starship explosion,鈥� I wrote. Midjourney processed for a few seconds, and returned four images, one of which was strangely accurate: a dinosaur-looking skull screamed out of the void of space, trailing fire (pictured, above). It resembled an illustration from Heavy Metal magazine, and perhaps art from the magazine influenced its creation.

You see, Midjourney and similar products like DALL-E, Stable Diffusion and Make-A-Video work their magic 鈥� translating words into pictures or video 鈥� by using vast 鈥渢raining data sets鈥� of images scraped from the internet. It is very possible that Midjourney partly honed its craft by consuming Heavy Metal comics, pretty much the way I did as a kid. But unlike 12-year-old me, Midjourney didn鈥檛 know what it was seeing. It relied on millions of unsuspecting humans for that.

Text-to-image AIs identify images by looking at the text that people have used to describe those pictures online. When Midjourney got my prompt, it contemplated images that random people had described as 鈥渄inosaur鈥� or 鈥渓aser鈥� and so on, then used what is called a to add a bunch of random chaos to those pictures. Once they were suitably scrambled, Midjourney 鈥渦pscaled鈥� them, removing noise and sharpening focus. Its work is so good that using it recently won first place for digital images at the Colorado State Fair.

But there are major ethical issues raised by the success of such AIs. The biggest has to do with those training data sets. Reporters at Motherboard that the data set used by Stable Diffusion contained images of terrorist violence and revenge pornography. When they prompted Stable Diffusion with the words 鈥淚SIS execution鈥�, the algorithm returned pictures that resembled real-life beheadings. Using an online tool called , anyone can search the data set that Stable Diffusion uses, LAION-5B. That is how the reporters discovered Stable Diffusion was training on a data set that contains actual photos of extreme violence.

Google also uses a LAION data set for an AI called Imagen, which hasn鈥檛 yet been released to the public. Both Google and Stability AI, which makes Stable Diffusion, told Motherboard they are working on ways to prevent the public from seeing images based on offensive and illegal pictures in the data set. A representative from LAION also noted that the images in its data set are 鈥渁lready available in the public internet on publicly available websites鈥�.

But even if this problem is fixed 鈥� or companies choose better data sets 鈥� there is still the question of all the other pictures online that are being transformed into AI-generated masterpieces. As many artists have pointed out, their work is being used without compensation. Adding insult to injury, , creating illustrations and even movies by using data sets stocked with art ripped off from artists who post their work online.

Some AI researchers argue that their algorithms aren鈥檛 stealing from artists so much as learning from them, just as human artists learn from each other. But a more ethical approach would be for companies to acknowledge their debt to artists and create a model of , much like radio stations first did in radio鈥檚 early days. Back then, musicians created groups like BMI to collectively license their music to radio stations 鈥� then BMI would pay artists based on how often their songs were played. Perhaps artists and art institutions today could form a 鈥渃ollecting society鈥� that would allow companies to license their artwork en masse for data sets. Artists could be paid based on how often an algorithm trained on the data set produced an image for commercial use.

Assuming we can iron out the many issues with training data, there is also the problem of just how good these algorithms are at creating realistic images. What if an army of robots starts churning out propaganda that can鈥檛 be distinguished from reality? But honestly, of all the potential harms of text-to-image AI systems, I think this is the least worrisome. People don鈥檛 need propaganda to be 鈥渞ealistic鈥� to fall for it 鈥� we fall for it all the time. The solution to propaganda is human moderation of our online communities, alongside better media education. In a sense, this is the solution to our data set problems, too.

To create ethical AI systems, we need to acknowledge the people whose work makes those systems so magical. We can鈥檛 simply snarf up every image online 鈥� we need humans to curate those data sets and we need to pay them to do it.