3 ms·

Oh, please. I've generated many, many images. They are not hard to spot.

by devin 2y ago

jnwatson 2y ago
The bad ones are of course not hard to spot. The good ones you'll never notice.
- devin 2y ago
  Good ones /of what/? Are we talking a human subject? Nature?
tokioyoyo 2y ago
Much respect, but nowadays, unless the person put basically zero effort to make it look realistic, there's no way you can detect whether an image is AI or not while quick scrolling. Obviously, if you look at every image as "let me examine every part of it to see if it's AI or not" mindset, you can still spot them. But anyone who spent a few days playing with the latest gen models, can create images that pass the 90% of sniff tests.
- devin 2y ago
  Do you have a test you like? I just took one at https://sightengine.com/ai-or-not?version=2024Q3 https://sightengine.com/ai-or-not?version=2024Q3 and got 18/20 correct, and I'm not zooming in on details or anything, I'm just using some basic discrimination based on what I've generated and seen generated in the past. I would do even better at this if we limited it to pictures of "realistic" settings.
  - tokioyoyo 2y ago
    I think we might be talking about two distinct cases. If you're actively thinking whether an image is AI or not, you're already biased to it potentially being AI-generated. That improves your recognition of slop-finding. As I mentioned, I definitely agree how it's fairly straightforward to spot the slop if you're looking for it. I'm not even sure how we could implement a real-life test without bias. Maybe if there was a complete feed of your internet browsing, where it asks you at the end of the day "ballpark the % of media that you think was AI?". Then go through the entire feed, and scrutinize it one by one.
    - devin 2y ago
      Right, and even there I think we might need to get specific about categories of images. Images that are supposed to be photo realistic are far easier to spot than "battleship in outer space" generations. Bringing it back to the topic of stock photography: A large percentage of stock photos are of real things, people, scenery. So, when someone says I'll have a hard time spotting generated stock photos, I kinda go uhh, well, no, not generally, because stock photos are very often of people and real life scenes, the thing that is the easiest to spot as a generation.
      - vidarh 2y ago
        Has anyone said you will have a hard time spotting them? Because I did not. I pointed out that when you say you can, it is an instance of survivor bias, and it is, whether you are good or bad at it as long as we don't have data to tell whether your assumptions were correct. We still don't know whether or not you're good or bad at picking out AI images used in actual campaigns, because we have every reason to assume at least a reasonable proportion of AI images used in actual ads will have been through an editorial process that'd rule out a lot of the easily recognized shlock, and so a test that does not use images that have been through the same selection process is meaningless. I have no doubt you can recognize some. You may well be able to recognize all of them perfectly for what I know. The point was not to argue you can't, but that your impression can't reliably tell you, because you'd be likely to think the same whether your accuracy is high or low.
      - tokioyoyo 2y ago
        I’m not entirely sure why you’re discrediting the advancement of realism. I’m very sorry, but I have a hard to believe that when you scroll through IG and see something like this — https://www.reddit.com/r/ChatGPT/comments/1hvdhie/this_girl_is_100_ai_generated/ https://www.reddit.com/r/ChatGPT/comments/1hvdhie/this_girl_... , you’ll think it’s AI instantly. Unless, again, you’re consciously examining whether every single piece of media is AI generated or not.
        devin 2y ago
vidarh 2y ago
You've already indicated elsewhere that in a test of images that had not been edited, or selected to minimize the risk of detection, you as someone who has spent lots of time generating AI images got 2 out of 20 wrong. So clearly it's possible to fool you. How many more do you think would get past you if the person running the hypothetical campaign was someone with a similar experience at picking images to you spending the same amount of time they would picking stock photography on ruling out any picture that looks like it's AI-generated to them, or editing them to remove things that'd tip you off?

Oh, please. I've generated many, many images. They are not hard to spot.

by devin 2y ago

Oh, please. I've generated many, many images. They are not hard to spot.

jnwatson 2y ago

The bad ones are of course not hard to spot. The good ones you'll never notice.

devin 2y ago
Good ones /of what/? Are we talking a human subject? Nature?

tokioyoyo 2y ago

Much respect, but nowadays, unless the person put basically zero effort to make it look realistic, there's no way you can detect whether an image is AI or not while quick scrolling. Obviously, if you look at every image as "let me examine every part of it to see if it's AI or not" mindset, you can still spot them. But anyone who spent a few days playing with the latest gen models, can create images that pass the 90% of sniff tests.

devin 2y ago
Do you have a test you like? I just took one at https://sightengine.com/ai-or-not?version=2024Q3 https://sightengine.com/ai-or-not?version=2024Q3 and got 18/20 correct, and I'm not zooming in on details or anything, I'm just using some basic discrimination based on what I've generated and seen generated in the past. I would do even better at this if we limited it to pictures of "realistic" settings.
- tokioyoyo 2y ago
  I think we might be talking about two distinct cases. If you're actively thinking whether an image is AI or not, you're already biased to it potentially being AI-generated. That improves your recognition of slop-finding. As I mentioned, I definitely agree how it's fairly straightforward to spot the slop if you're looking for it. I'm not even sure how we could implement a real-life test without bias. Maybe if there was a complete feed of your internet browsing, where it asks you at the end of the day "ballpark the % of media that you think was AI?". Then go through the entire feed, and scrutinize it one by one.
  - devin 2y ago
    Right, and even there I think we might need to get specific about categories of images. Images that are supposed to be photo realistic are far easier to spot than "battleship in outer space" generations. Bringing it back to the topic of stock photography: A large percentage of stock photos are of real things, people, scenery. So, when someone says I'll have a hard time spotting generated stock photos, I kinda go uhh, well, no, not generally, because stock photos are very often of people and real life scenes, the thing that is the easiest to spot as a generation.
    - vidarh 2y ago
      Has anyone said you will have a hard time spotting them? Because I did not. I pointed out that when you say you can, it is an instance of survivor bias, and it is, whether you are good or bad at it as long as we don't have data to tell whether your assumptions were correct. We still don't know whether or not you're good or bad at picking out AI images used in actual campaigns, because we have every reason to assume at least a reasonable proportion of AI images used in actual ads will have been through an editorial process that'd rule out a lot of the easily recognized shlock, and so a test that does not use images that have been through the same selection process is meaningless. I have no doubt you can recognize some. You may well be able to recognize all of them perfectly for what I know. The point was not to argue you can't, but that your impression can't reliably tell you, because you'd be likely to think the same whether your accuracy is high or low.
    - tokioyoyo 2y ago
      I’m not entirely sure why you’re discrediting the advancement of realism. I’m very sorry, but I have a hard to believe that when you scroll through IG and see something like this — https://www.reddit.com/r/ChatGPT/comments/1hvdhie/this_girl_is_100_ai_generated/ https://www.reddit.com/r/ChatGPT/comments/1hvdhie/this_girl_... , you’ll think it’s AI instantly. Unless, again, you’re consciously examining whether every single piece of media is AI generated or not.
      - devin 2y ago

vidarh 2y ago

You've already indicated elsewhere that in a test of images that had not been edited, or selected to minimize the risk of detection, you as someone who has spent lots of time generating AI images got 2 out of 20 wrong. So clearly it's possible to fool you. How many more do you think would get past you if the person running the hypothetical campaign was someone with a similar experience at picking images to you spending the same amount of time they would picking stock photography on ruling out any picture that looks like it's AI-generated to them, or editing them to remove things that'd tip you off?