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.

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.

Every generation prompt has this suffixed:

    The scene should feel incidentally noticed during daily life, not arranged for an advertisement. Natural perspective and subtle everyday wear. No cinematic grading, stock-photo staging, deliberate blur, artificial grain, fake low resolution, extreme background blur, exaggerated HDR or oversharpening. No collage, border, caption or added watermark.

At some point during the process of building up this collection of turd-polishing techniques, the thought should’ve occurred that this is actually a lot of effort to get generations confused for reality. But I suspect this process was being carried out by something that couldn’t have an original thought.

> 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?

Highest I got was 26 in a row. AI seems to love generating "close up" images of a subject. Its like every image has to have an obvious subject (that was in the prompt), and it struggles to make scenes without a clear subject and more chaotic vibe. I notice this a lot with these "AI movies", every single camera shot has an obvious subject that makes them unbearable to watch, they can't just chill out and have a messy scene.

Highest score is 16,800. Highest streak is 60 in a row.

I just came to HN to see if I won!

Also, I do research on how people perceive images and video. I’ve published a couple papers and did my dissertation on how folks understand moving images.

If you’re the developers of this game I’d love to help you parse this data. B L T P H D at V T dot E D U.

Something we’ve been discussing in our lab is if people can detect the “taste” of AI. Seems if you play this game enough times, you can at least recognize the differences pretty consistently. I’d love to know how rounds of play improves people’s accuracy.

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.

Neat app but it makes me sad. Also I'm very curious as to where the real photos come from and how we can verify that they are indeed real. The timer thing doesn't really add any value, just stress.

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-...

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

Also the time limit is stupid and pointless.

Good reality check for people who proudly boasted about "always" being able to spot AI images.

I guess the UI is ai generated, and it impossibly to see the images well enough to judge on a phone.

Hmm… training data siphon?

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.

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.

My biggest clue is the odd repeating details. Especially on grass, tree leaves, pores, etc

Not all image models (or sampler settings) do this, but with gpt it's very obvious.

With this in mind, I misjudged 3 images on the default game.

I failed horribly and might as well used a random coin toss

I don't get the scoring system: I got 0 points, but I got a streak of 2 in a row. Bizarre.

not doable in short time with high certainty which means that from now on an image is only worth a thousand words

Please give us a fraction or percent of how many we got correct.

Is this a crowd sourced training effort to make their image generator better?

Got 4k by just spamming 2.

I got 14/16 correct, and the two I got wrong had no people and I was unsure about both. Should have went with my gut.

undefined

Wow - I did much worse than I'd have thought.

The general lack of humans helps conceal the difference though. Bypasses the uncanny valley problem. I'd contend that most of us would do a bit better on truly random photos

Surprised how good the nature/scenery shots are though. When I last played with a (local) image gen I tried making forest scenes and it was terrible at it

TIME'S UP

2,050 points Best streak: 9 in a row

Hard af.

undefined

this is fair only if i have enough time to decipher.

12/17. Due to the speed, I felt like I was just randomly guessing, but 70% correct seems better than random guessing. I did suspect some things.

How mundane the AI images were felt like a means to trick people. It doesn’t seem like most people making AI images and posting them are generating still life images of everyday life. What would be the point?

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

best strategy: just click as fast as possible.

My method is simple:

Did I create the image myself?

Yes: Not AI

No: AI

The trouble with these tests are that they are themselves training AI. I don't want to let designers know the "tells".

Can I tell what’s AI generated?

Well if I can I’m not going to help train more slop image generation.

Bouncing on AI generated button as fast as possible without looking at the images got me 6,000 points.

No, but we still can tell that this web site is AI-generated slop.

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.

uh yeah...we need hardware attestation and a protocol to make it work cross platform like, yesterday...

holy shit