2 ms·

this is the kind of thing that algorithms are really good at (determining provenance/AI traces, at scale). search has been trying to fingerprint users forever,

this is the kind of thing that algorithms are really good at (determining provenance/AI traces, at scale). search has been trying to fingerprint users forever, its time to point their agorithmic guns at something far more useful for the actual end user. it makes sense for search algorithms to spot these and then flag them accordingly if they have not been disclosed.

criley2 3y ago
Algorithms are NOT good at detecting AI, and every attempt so far has been laughably terrible. For example the internet is already filling up with students complaining about teachers failing their school work for false positives on AI text generation. And there's the reality of the Survivorship Bias when it comes to examples: the only times you see AI art are the times it didn't fool you, not the full set of all AI images you saw. Similarly, if the algorithm can find the easy ones but miss the same better ones, is it really that useful? Or is it just training more realistic AI to evade both you and the algorithms detection? The idea that people will fundamentally be able to differentiate AI and human works is nonsensical from a future perspective and is an aberration of AI quality for the next few years at most. If you're not preparing for fully indistinguishable AI text/image/video/audio, then you're not preparing for the future.
- macNchz 3y ago
  Realistically I think the only way to create verifiably-human-made digital content is basically a end to end chain of trusted computing DRM from keyboard or camera lens to final product, which seems dystopian in its own way, impractical, and a high value target for people to try to bypass.
  - pixl97 3y ago
    Authoritarians: "Go on, I would like to know more"
orbital-decay 3y ago
With an automatic system, what will actually happen is even more algorithmic bureaucracy, and real people being banned for false positives or for using any ML-assisted techniques at all. These algorithms are snake oil, there's simply no real way to detect it reliably. Youtube is enough of a dumpster fire already and its automated systems are being exploited left and right. Don't make any mistakes, this requirement is meant to cover Youtube's ass, not to do anything useful to you. How it will be enforced is entirely up to them. I suspect it will be much closer to "occasionally banning someone to appease the outraged crowd" rather than to "algorithmic banhammer".

criley2 3y ago

Algorithms are NOT good at detecting AI, and every attempt so far has been laughably terrible. For example the internet is already filling up with students complaining about teachers failing their school work for false positives on AI text generation. And there's the reality of the Survivorship Bias when it comes to examples: the only times you see AI art are the times it didn't fool you, not the full set of all AI images you saw. Similarly, if the algorithm can find the easy ones but miss the same better ones, is it really that useful? Or is it just training more realistic AI to evade both you and the algorithms detection? The idea that people will fundamentally be able to differentiate AI and human works is nonsensical from a future perspective and is an aberration of AI quality for the next few years at most. If you're not preparing for fully indistinguishable AI text/image/video/audio, then you're not preparing for the future.

macNchz 3y ago
Realistically I think the only way to create verifiably-human-made digital content is basically a end to end chain of trusted computing DRM from keyboard or camera lens to final product, which seems dystopian in its own way, impractical, and a high value target for people to try to bypass.
- pixl97 3y ago
  Authoritarians: "Go on, I would like to know more"

orbital-decay 3y ago

With an automatic system, what will actually happen is even more algorithmic bureaucracy, and real people being banned for false positives or for using any ML-assisted techniques at all. These algorithms are snake oil, there's simply no real way to detect it reliably. Youtube is enough of a dumpster fire already and its automated systems are being exploited left and right. Don't make any mistakes, this requirement is meant to cover Youtube's ass, not to do anything useful to you. How it will be enforced is entirely up to them. I suspect it will be much closer to "occasionally banning someone to appease the outraged crowd" rather than to "algorithmic banhammer".