undefined | Better HN

0 pointsMetaWhirledPeas2y ago0 comments

> Think about security or trust and safety or anti-scam or anti-fraud.

It doesn't matter what context I think about it in. It isn't going to work! And it will make things worse for everyone involved.

Hypothetically let's say we get to a point where everyone believes the detection is 100% accurate. Well that's all that means: everyone believes it. Meanwhile AI has just gotten better, and we're all more fooled than we were before. All we are really accomplishing is enhancing the training necessary for AI to elude detection.

And there will be an inherent bias toward false positives, because high detection rate will be the selling point. The truth is secondary, and there's no way to verify the results.

0 comments

rngname222y ago

It does work. If you absolutely need to know if an image is AI generated, you can just have a central authority in the system watch the person draw the picture on a piece of paper. Or drive to your house and hand you the paper and pencil and watch you draw it in person.

There are workflows or system designs that absolutely can and will solve for human-verified creation, they just might be incredibly costly or unscalable compared to existing solutions. It's all just tradeoffs. Might make existing business models no longer work. Might open new ones.

MetaWhirledPeasOP2y ago

> If you absolutely need to know if an image is AI generated, you can just have a central authority in the system watch the person draw the picture on a piece of paper. Or drive to your house and hand you the paper and pencil and watch you draw it in person. There are workflows or system designs that absolutely can and will solve for human-verified creation, they just might be incredibly costly or unscalable compared to existing solutions. It's all just tradeoffs. Might make existing business models no longer work. Might open new ones.

This is pretty much my point. Like you said, incredibly costly and unscalable. A non-solution! We're better off not pretending we can compute what is and isn't real.

rngname222y ago

I don't agree, I think it's just a matter of the right mix of distributed trust and creating incentives for honesty and penalties for dishonesty. As well as those costlier mechanisms for verification to be available for a subset of cases.

j / k navigate · click thread line to collapse

0 comments

rngname222y ago

MetaWhirledPeasOP2y ago

This is pretty much my point. Like you said, incredibly costly and unscalable. A non-solution! We're better off not pretending we can compute what is and isn't real.

rngname222y ago

j / k navigate · click thread line to collapse