Just ask them to hand write a sentence of your choosing on a piece of paper and put it in the photo. Even if AI manages to produce a "handwritten" sentence it'll likely look too perfect and stock.
this post was submitted on 20 Jul 2026
20 points (81.2% liked)
Fuck AI
7684 readers
1057 users here now
"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
founded 2 years ago
MODERATORS
A set of mechanical clocks with specified different times set on them.
Me, naked. It won't actually look like what i look like naked. It doesn't know how many nipples i have, et cetera
A ruler with two fingers of their hand above it seems like a good idea. Or an egg or another thing of s known size.
Or perhaps a photo of uncooked spaghetti on a plate if you throw them like the sticks of a mikado game. It would probably picture cooked spaghettis.
You might be hard-pressed to find something a human can make, but an AI cannot. (It also doesn't preclude an old-fashioned scam).
An AI would struggle with something like depicting the melancholia of having your partner die to a wasting disease, and when they do die, their estranged family comes in, taking almost everything you had in you shared apartment while you're helpless to stop them, because your relationship isn't legally recognised.
But so would a human. It's not impossible, mind, but 'twould certainly be difficult.
What about going the other way? Ask for something that an AI can do easily, but a human can't.
Consensual nudes.
A hand-drawn picture of Elon, his shrivelled, tiny excuse of a dick in hand. AI will refuse to render something like that because it involves
- nudity (all common models) or
- an accurate depiction of Elon (Grok)
The downside is: humans might refuse to "render" something like that as well...
Ask for multiple photos from different angles that show the same background scene behind the object too. It's difficult to get an image generator to generate the same thing several times from different angles consistently, especially the details in the background.
If there's a person in the photo it can be even more obvious when their face looks slightly different each time, like you're seeing a series of siblings substituted in for each other.
Qwen image edit plus, with the multi angle Lora
Since AI is so eager to please, I'd ask for an image of something that does not exist or is highly unlikely.
Like, a clover with five leaves (in winter) or a picture of the snowman in the garden )in summer ). Or a picture of two wine bottles stacked vertically on each other.
And do it in a way that real people have something to laugh.
Why not just ask for an image of the product with a date and timestamp in the photo? Keep it vague but expect to see it written on a sheet of paper. AI will likely confuse this with a digital date and timestamp, or it should mess up the text entirely.
I feel like the best bet would be to go with instructions that a person will understand, but an AI will likely take literally or get confused.
If someone asked me for a timestamp, I would not picture something handwritten.
Aim for something that will confuse computers, but not autistic people.
Though if we could correctly predict communication difficulties between autistic and allistic people, things would be better.
Proper keyboard layouts. Seems like they always screw that up somehow
Maybe bar codes (with numbers above/below), like on a regular product box. Of course i mean a working bar code that is 1:1 the same as the number written above or below it.
Complicated shadows in scenes. AI sucks at real world physics and light directions, and complicated shadows, like under a tree or a bunch of chairs and a table always look janky
The barcode standard is pretty convoluted to say the least. Any of the numbers can have up to 4 different sets of lines depending on a combination of the placement and the initial digit.
So it wouldn't be able to do it from picture sources, but I guess it could simply outsource the task to one of the many barcode generators - or steal their code to do it.
Just tried the keyboard, and unfortunately, that also worked. It was able to generate a quertz layout on a laptop no problem.
The barcode thing is an interesting idea, I will look into that.
Shadows might be difficult, but not really helping in my case. I can't ask a seller to please put the product outside and take a picture of it underneath a tree...
There are many different keyboard layouts for different languages
Most being qwerty
My keyboard is non US so enter button has a different size, left shift is smaller etc
May I see it?
(not parent commenter but) try searching japanese layout
Why were you buying a used full glass of wine online?
No, I just asked them to place a full glass of wine in front of the object I wanted to buy. As AI couldn't do that, I could at least be sure they'd actually have the product.
If you'd ask me that, I would tell you I wouldn't do that. That's an awful lot of trouble to go to, I don't just have open bottles of wine any time of the day. And once you pour it, you'd have to drink it, which I don't normally do during the day or even on weekdays. Especially filling it to the brim, so I can spill it or what? That's asking a lot of a random person on the internet to do just to sell something old.
Haha, noo, I ask for water, not wine. And yes, people actually do that and it did reveal a scammer once. And I usually only do it with more expensive items.
I just ask the person for a voice phone call to get a sense that they understand the item that they're selling. I haven't gotten caught in any scams, though maybe some items I've noped out on have been legit (no way to tell).
I don't think I've ever sold anything online, but if I tried and was asked for a voice call, I'd just say no.
edit: a single letter. I leave it to you to guess which.
I've only done it for larger items if that helps. Like $300+.
isn't that more likely to be ai if they do it? who would even agree to a request like that?
As AI couldn't do that, I could at least be sure they'd actually have the product.
... what?
That was pretty easy to follow. What didn't you grasp?
How does which kind of image an AI can or can not generate translate to an availability of a product?
EDIT: Oh, now I get it.
Will Smith eating spaghetti with his hands.
I wonder how much CO² was created by each shitty image to satisfy everyones curiosity. Obviously there is no thought given towards the consequences, no remorse for their ever increasing role in the avoidable deaths of millions of people that are too poor for anyone to give shit about. I wonder how many people in this awful thread also burn gasoline to get to work. How many of them will starve when most of the farmland is destroyed and still feel no remorse for the choices they made that lead to their own deaths and the destruction of everything they know? Just what the fuck is this thread?
Calm down. I don't use ai for anything I do, not even when the company I work for encourages me to do so. But I need to know how to defend myself against new scam techniques and thus I want to understand the weaknesses of the tools they use. I am totally on your side when it comes to pollutants. In fact, no, I don't burn gas, I don't even own a car or a drivers license for that matter. I take the bicycle to work every day, always did. I never fly in airplanes. I don't eat meat. I travel by train as much as I can, if I have to travel. But I need to know what I can do about all the fake images online. So I don't understand how it can make you so upset, its not like I'm promoting ai, im just trying to find a way to protect myself.
Maybe ASCII art?:)
Tiny contextual details often give it away.
Not sure if it still works but for the longest time, "bulldozer" really gave it trouble
"AI" can't generate any images because it doesn't exist.
As far as computer programs, they typically generate images by averaging over some "big" dataset. I imagine the best way to have the computer fail at its task would be to ask for images that aren't in the dataset or that can't be created by the generation algorithm.
Can it generate clocks showing any given time (don’t choose 10:10, for example)?
Woah, that's a good one, the first try failed! Will look into it more.
Edit: Oh well, it nailed it on the second attempt. Too bad...
I honestly just think of LLMs as dementia patients, so that’s why I went in that direction. Maybe try other dementia tests (though obviously if a person with dementia is trying to assert their personhood, it won’t be effective).
I think that’s probably all kinds of offensive, and the only real defense for it is, ironically, heuristics. I don’t mean to offend, and my impression is probably based on my first hand experience of ADHD/pothead/CPTSD cognitive/memory issues and secondhand experience living with my grandmother for nearly a decade after her Alzheimer’s diagnosis. I’ve seen both of us do things that seem like they’re based on dream logic or autopilot, and they seem to be similar to LLM hallucinations.
For me, it feels like my subconscious mind knows the sorts of things I do and carries them out without any conscious direction or awareness, meaning they don’t always make sense. Doing my hair before putting on a pullover top or washing and drying my hands before washing dishes are common examples from my daily life. My grandmother used to get a long stream of logical steps from an illogical starting point, like thinking her great uncle (who died in 1952) was trying to surprise her by taking her to the fair, so she needed to get dressed up and pack snacks and a camera, and anyone suggesting it wasn’t happening was in on the surprise.