3 ms·

> We havent had an ability to tell whats real and whats not for a long time. Photos have been able to be doctored since well before photoshop, writing style im

> We havent had an ability to tell whats real and whats not for a long time. Photos have been able to be doctored since well before photoshop, writing style imitated, signatures forged etc. This is a change of degree, thats all. This kind of argument rubs me the wrong way. It's like saying a printing press is just a faster scriptorium, or a nuclear bomb is just a change in degree from a hand grenade. It's misleading because it ignores the monumental difference in society these later inventions brought about, which were not caused by the previous inventions, and never would have been. It ends up being an argument to not engage with the reality of the situation.

the_snooze 3y ago
I agree. There's this weird contradiction in tech circles to hand-wave away legitimate concerns about scale (e.g., ubiquitous surveillance is just a policeman at scale), while simultaneously putting a lot of emphasis and value on whether a system is scalable. The more complicated reality is that things qualitatively change at scale. We need to have honest sober discussions about the trade-offs and unknowns involved, instead of just YOLOing blindly into moving fast and breaking things.
OliveMate 3y ago
I feel the issue is with human filters. Creating a seamless and believable image edit requires time and effort, regardless of how experienced the editor is. You were limited by what images you could source, how they'd fit together (does the lighting and perspective match?), and there was often a contextual paper trail people could follow ('I recognise that picture of X'). With these image generation models, the only things filtered by human involvement are the creation of the prompt, and whether or not they believe any generated image is believable. The required time and effort is minimal, and on top of that you can generate hundreds by the time a single person can make a convincing Photoshop edit. --- In terms of being able to sift through generated images across the wider internet, I've no idea what can be done. But I feel a potential defence against them would be for news sources to publish the exact images they were sent online, keeping the original resolution, removing cropping, and retaining Exif data.

the_snooze 3y ago

I agree. There's this weird contradiction in tech circles to hand-wave away legitimate concerns about scale (e.g., ubiquitous surveillance is just a policeman at scale), while simultaneously putting a lot of emphasis and value on whether a system is scalable. The more complicated reality is that things qualitatively change at scale. We need to have honest sober discussions about the trade-offs and unknowns involved, instead of just YOLOing blindly into moving fast and breaking things.

OliveMate 3y ago

I feel the issue is with human filters. Creating a seamless and believable image edit requires time and effort, regardless of how experienced the editor is. You were limited by what images you could source, how they'd fit together (does the lighting and perspective match?), and there was often a contextual paper trail people could follow ('I recognise that picture of X'). With these image generation models, the only things filtered by human involvement are the creation of the prompt, and whether or not they believe any generated image is believable. The required time and effort is minimal, and on top of that you can generate hundreds by the time a single person can make a convincing Photoshop edit. --- In terms of being able to sift through generated images across the wider internet, I've no idea what can be done. But I feel a potential defence against them would be for news sources to publish the exact images they were sent online, keeping the original resolution, removing cropping, and retaining Exif data.