YHK - yet another hacker news clone

My method is simple:

Did I create the image myself?

Yes: Not AI

No: AI

> 60 seconds to find out.

> 90s round, 10s per image

> Play 60 seconds

This is confusing. What does "90s round" mean if it's a 60 second game? Is it about the 1990s?

My most effective strategy is to ask: is there any reason for a human to take this picture?

One AI-generated image was a bedside night stand, with no clutter other than a set of keys on top of a receipt. It looked perfectly real, no glitches or weirdness... other than the idea of someone taking a picture of it.

Not a perfect strategy, though; I thought this one was AI for just that reason: https://commons.wikimedia.org/wiki/File:Potted_plant_1_2018-...

Not bad. I’d appreciate a much longer time limit. Got really rushed at the end. But this was crazy hard on my phone, I was straight guessing most of the time. I think I could be a little better with more time on a desktop screen.

Another idea would be games that were all pictures of a single category- like 10 AI and 10 real pictures of basketballs, or food on plates, etc.

I realized this is only somewhat difficult because of the 10s timer. Most people probably don’t spend more than 10s looking at a picture on social media, excellent project.

We get it. AI slop is getting good enough to replace photography. This isn't good for humans, and makes the world a worse place.

The fact that the suck machine works isn't something to be proud of.

Hmm… training data siphon?

Impossible to do on a phone. Zooming in is misinterpreted as a choice. Fix the game.

Also the time limit is stupid and pointless.

The time limit did make it quite challenging. I got 1,000 points with a max streak of 8 in a row. On second playthrough I got 3,550 / 12. It seems the AI images all come from a particular area of latent space that's quite easy to recognize once you get dialled in.

So it's more of a "can you detect the GPT Image 2.5 house style", which is an easier problem.

I'd appreciate an untimed mode.

Now add Geoguesser-like battle royale mode with accuracy points for guessing which model generated the image

I hadn't been exposed to the new GPT-Image-2.5 model before this, so didn't have any "tells" to go off of, and found it almost impossible to begin with, but it seems the brain is pretty good at adapting for visual stuff! After 2 rounds, I was able to build up a 16-image combo, answering most within a second or two.

I think the "tell" I ended up with is that almost all of the AI images tend to center around a very obvious main "subject" (a flower, a bike, a wrench, etc) – presumably an inherent artefact of generating from a prompt – with everything else around the subject having a very strange depth to it. The depth, focus, and bokeh around the main subject just never look quite right. If you look at an image and find it has a very strong central subject with a little too much depth separation than expected, there's a good chance it's AI.

I recognise that this method of finding a consistent "tell" will probably not hold for future models, unfortunately. Everything else looks almost perfect at this point.

It would be cool if they published stats on how well people do on average, and how that changes over time and with different models. I feel like in a few years we'll either all be much better at this or completely desensitized, and sadly I'm guessing we're trending towards the latter.