cross-posted from: https://programming.dev/post/3974080
Hey everyone. I made a casual survey to see if people can tell the difference between human-made and AI generated art. Any responses would be appreciated, I’m curious to see how accurately people can tell the difference (especially those familiar with AI image generation)
I only guessed a single one as generated by AI and I was wrong on that (the mouse in the boat drawing felt like unusual shaping and shading for a human). I really could not identify any telltale signs of AI in any of them, so answered entirely honestly that none of the others looked like AI generated.
To be honest, I expected that. The telltale signs people often talk about are only problems for bad AI art. Well done AI art really is indistinguishable. Stuff like weird fingers, faces, and teeth are only problems if the prompter is lazy and just picks the first thing generated (and doesn’t selectively regenerate). If you’re selective, you can get AI art without the things some people claim make AI art easy to recognize.
It’s like photoshop or movie CGI. Anyone can detect a bad photoshop and we’re used to seeing those. But well done photoshops by experts can be near impossible to detect (short of careful pixel level inspection, which doesn’t really apply to AI art). Yet, a lot of people are over confident in how well they can spot photoshops.
I wonder if this will change anyone’s mind? I’ve always wanted to do this for a few topics, including AI. I’ve also wanted to do this for trans vs cis people (so many transphobes claim they can “always tell”), movie CGI vs practical effects (see also: Captain Disillusion videos), and for various kinds of food and drink (so many people are elitist that something tastes better – simple example that I’ve actually seen disproven is the various kinds of eggs, including store bought vs locally sourced).
The scenery ones are near impossible to tell because you have no context. The pencil art one is impressive though.
There are things you can look for. When it isn’t generated, you can spot parts where the artist got lazy. Sometimes, if the art style allows for it, you can spot simple shapes that are left over, and the lighting.
Also things like the butterfly in one of them look off in a way that a human who’s seen one won’t draw it as if they’re actually capable of the rest of the image
I did bad.
But I expected to do bad. AI generation has become too good.
You tell yourself you can identify them, because sometimes you notice weird artifacts and spot the AI quickly. But we’re really only noticing the bad ones. We’ll never even know the good ones were AI most of the time, so we can’t balance how good we think we are at spotting them against how often we were actually wrong.
Hands are often a giveaway. The first image for example shows perfect hand proportion - even through a glass. AI isn’t there yet.
Hands are only a give away for bad AI art. There’s no shortage of examples with great hands (especially when using features like Stable Diffusion’s ControlNet, which allows you to give hints to the AI for the shape that something should match). Just so many people posting AI art generate once or twice and post that. If you’re more selective and selectively regenerate, you’ll be able to get much more believable results.
This is also a rapidly changing area, with the most cutting edge AI being way better than something from even a year ago. Used to be that no AI could do even remotely believable text, but in recent weeks, I’ve been seeing many examples of AI art that got small amounts of text perfect.
I was missing a “don’t know”/“can’t determine” option.
For photographs specifically and some types of paintings/artificial stuff, there are things you can look for. But for other things, I feel like, or at least to my knowledge, you can’t.
Like the pencil drawing. There’s not enough things it could be doing wrong. It’s a sketch. With simplistic but “error-excusing”/diffuse/transformable content.
The goal isn’t really to be a quiz, but rather just to see how susceptible people are to AI generated art. Many of the images I chose are intentionally vague, 80% of people so far got the line art sketch wrong, and that’s with knowing that many of these are AI generated. The results are definitely interesting to see.
A “don’t know” option would ruin the point since most people would just choose that. I want to see where people lean towards.
13/20, I work in AI. The paintings were the hardest for me, because the art style obfuscates some of the AI artefacts that can be tells.
