Well, yes, but they’re rare enough that we still say “Hey, the AI screwed up that hippie dude’s hands”, not “Hey, that hippie dude is polydactyl”.
And now I’m wondering how the current AIs would fare with the prompt “Count Rugen’s hand”.
Oh, and @merrick , I love the name “the bananana problem”.
In the context of this thread it was very easy to see the words. In the wild it would be harder and of course it would be easy to make it consciously undetectable. Does subliminal stuff work? Probably some of it does. Now let’s all go to the lobby…
But yeah, it’s not like other images are devoid of messages or manipulation, and there are probably more concerning Xesses. If problems are intermittent, perhaps we have no banananas today?
Here is a batch from Bing.
All of which could quickly and easily be fixed by anyone with rudimentary PS skills.
It some ways it does help, you have to adapt. Here’s a personal example however, back in the late seventies I was fresh out of architecture school. Being artistic, I discovered I could make more money doing architectural renderings than drawing plans. Architects need these illustrations to sell their designs. The ability to draw an unbuilt structure from plans in 2 or 3 point perspective and then paint in in ( shadows and all) was a specialized skill that no computer could come close to.
Back then I was getting about $1500 for a commercial structure. Then came the first decent perspective program. You still had to transfer the image down onto illustration board and paint it but my competitors were offering finished products at a much cheaper price - about half of what the going rate was.
Then of course, CAD programs started spitting out the final products. At that point it was either charge a couple of hundred dollars or find another field.
Now I think all architects generate these in-house.
It’s kind of sad, I feel, because the computer generated renderings don’t seem to have “soul”.
Yes, hands are hard. As an artist, I find that drawing extremely funny. Thanks for sharing that.
I don’t think it does “range way the heck down”. The limitations on the alteration of images or inserting “subliminal” components to invoke strong affective responses in the past have been limited by both the technology of image modification and the ability of a human editor to accurately gauge the effects on a substantial population sample. Certain types or styles of propaganda are recognized as being highly effective in politics, advertising, and other venues of influence but the expert application of them has been limited to the small pool of people who have an intuitive grasp on what rides the line between being effective without overtly offensive.
However, generative AI can now produce images that are photorealistic enough to fool the sharpest critical eye, and we can only distinguish them because it hasn’t yet figured out how many fingers and nostrils a real person should have, which is a flaw that will certainly be worked out in subsequent models. And a generative AI could produce a series of fabricated images, propagate them out via social media, gauge responses (either indirectly through “clicks” and “shares”, or perhaps even directly by monitoring involuntary bioaffective feedback through the camera and other sensors on mobile devices and fitness trackers), and fine-tune the image to achieve maximum influence.
And regardless of how logical and rational you might believe yourself to be, an image, audio, or video is always first processed by the affective brain systems before being integrated into something that your more rational (or at least rationalizing) mind can address. A generative AI that can produce highly affective sensory ‘products’ that are finely tuned to hyperstimulate the affective brain and hijack emotional responses can essentially override critical thinking, knowledge, and emotional regulation that are essential underpinnings of democratic and egalitarian governance. Exposure to and especially dependence on information ‘processed’ (or manufactured) by AI systems can literally make us dumber and more atavistic, at the mercy to whomever controls this capability.
Stranger
But 90% of what you said applies to AI images well above and beyond what the thread is about: putting messages into AI images.
The ability to create photorealistic images of people and events is an actual important issue. The ability to put “Vote for Cthulhu” into those images does not, in of itself, rank nearly as high.
The concern about “deep-fakes” and other obvious manipulations is well founded in a time of “fake news” but there are at least straightforward countermeasures toward factual misinformation falling under the general category of fact-checking and meme-trapping. However, the real problem isn’t “fact versus fiction”; it is manipulation of the public below the surface layer of verifiable information. This isn’t some kind of fantastical snowcrash-esque cognitive control; this is what advertising people do every day to sell their client brand’s products and services over functionally identical products by deftly manipulating the consumer to believe that the thing they are advertising will make the consumer feel somehow more fulfilled at a level that goes far deeper than cognitive critical thought.
In the past, this is done by people who spend their lives understanding how to use visceral auditory and image cues, short-form narratives, attractive or outlandish looking spokespeople, et cetera to achieve the desired promotion but they do so in a very broad sense to appeal to a wide potential consumer base. The combination of generative AI and systems that can take the extensive feedback that is available from a sample population and use it to fine-tune the affective response that can circumvent rational thought. And this is already happening—albeit at a crude level—with social media algorithms and news aggregators using learning algorithms to pick the posts and stories with the most outrageous content that will garner greater views and forwards. With the ability to insert triggering cues and sublimated ideas into media below the threshold that anyone not explicitly look for them would recognize, and then use feedback to tune these to a targeted population or broad appeal, these systems will essentially undermine any kind of critical thought.
You don’t think this can really happen, and all of this subliminal content is just nonsense? Take a look at the conspiracy theories that routinely make the rounds on social media, especially targeted social platforms. Stuff that is so patently absurd and factually falsifiable—really fringe, madman-standing-on-the-corner-shouting-into-the-sky, Hollywood-wouldn’t-even-touch-this nonsense—turns into a chain-reaction that causes millions of previously mostly-reasonable adults into paranoid conspiranoiacs and drives the 24 hour news cycle even though you or I, coming from the outside of this community, would read a few of these posts and wonder how anyone could take this seriously. And that is just the first wave, driven by primitive algorithms just using anonymous click-and-forward data for feedback.
Generated images of deformed cats spelling out mildly profane words may seem moderately amusing, but the capability than underlies it is a deep and enduring threat to the body politic as at least a somewhat rational decision machine that acts in its own (if typically short-sighted) best interest. This generative AI is a capability that can and will cause people in the aggregate to act against their interests and factionalize until there is nothing but infighting and cutting one’s collective nose to spite one’s face, which is something we don’t need any additional help with.
Stranger
So what is to be done? I presume you agree politicians are not, as a group, remotely expert in technical issues even if they had the will or sagacity to address them. And legislation will always lag innovation by years. They say if you can describe and understand a wicked problem it is halfway solved. But they say a lot of things.
Agreed that legislative solutions, even if they were in the offing, would be as poorly constituted and ineffectual as most attempts to regulate technology and innovation. Regulatory and executive efforts might be more flexible and well-concieved, but fundamentally developing this capability doesn’t require extensive resources beyond commercially available computing hardware and some specific expertise, so even if it were legally prohibited or tightly regulated it would be pretty easy for intentional bad actors to subvert such restrictions.
The most effective action that could be done is to develop and deploy countermeasures to identify efforts to manipulate the public, and like any arms race that becomes an endless competition where the defense is always lagging. But trivializing the detrimental effects of these capabilities and especially treating them as a toy box of ‘fun’ entertainments essentially desensitizes the public to what is being done to diminish their fundamental volitional autonomy and undermine what remains of the trappings of democratic institutions. I suppose this is probably just closing the door on an empty barn, but dancing on the ashes of civil discourse and the social contract is callow and revulsive.
Stranger
One guy: I wonder if I can create live QR codes in Stable Diffusion?
Everyone: Fun!
Another guy: Hey, look, you can use this to embed text and images, too!
Everyon: Fun!
Nobody at all: I’m going to create this to diminish everyone’s fundamental volitional autonomy and undermine what remains of the trappings of democratic institutions.
There isn’t a big enough roll-eyes emoji in the world to represent this.
Anyone not living in a cave is already subject to constant influences to diminish their fundamental volitional autonomy and undermine what remains of the trappings of democratic institutions.
Higher-tech tools to do this will be abused starting yesterday. Enjoy your “free” kitty and puppy pictures generator, though!
But that isn’t the purpose for their creation.
Well, if I’m being led to slaughter, I want to enjoy the journey.
Oh, DPRK, that’s a load of rich creamery butter.
And to think that no one had any way to embed “OBEY” into a kitten photo before today. Oh, wait, they did and no one bothered to do it because embedding OBEY into a kitten photo isn’t actually good for anything except some internet chuckles.
And it’s actually redundant, since humans are already conditioned to obey domestic felines.
Good point. My cats would think it beneath them to subliminally suggest I obey them - because of course I should.
Anyway, problem solved. I am sure.
1st image: No idea what it was supposed to be. I got “NO MMFC”. It really doesn’t frigging help when the letters are broken up into completely differently-colored elements. (I do agree that dogs can be a real mmfc, for whatever that’s worth.
2nd image: I’m seeing “LOVL”, but it’s clear enough that it could be an “e” if you lower your standards enough. Not bad.
3rd image: I tried, dammit. I even checked out the negative space; that’s how desperate I was. How the freak is that a “b”???
There’s definitely some potential to AI, but right now I find it mostly annoying, albeit occasionally funny.
Lucas Jackson - 
Case in point. I’m thinking there’s a Chen/Rin Kaenbyou/Mike Goutoukuji joke to be made here, but I’m not exactly sure what.
CairoCarol - And this. I’m going with “Cyriak video”.
Did you step far away from the screen, or reduce the size of the images? Most of these are, in fact, very difficult to see up close, but jump right out at a distance.