> There are real, legitimate reasons why SynthId is actively harmful that do not involve deceiving or manipulating people at all

I am legitimately curious: can you name some?

> Actually no it just makes me use a different model

Yes, this is a very good thing when "a different model" means "a worse model."

> People who want to deceive or manipulate are not using Google models anyways. They are going to use a model without safety rails

That's totally invalid logic. There are plenty of deception and manipulation use cases that don't run afoul of model safety rails at all. Trivially: Creating fake dating profiles to scam people. Fake product images. Fake insurance claims. Fake blackmail (e.g. of a person and another man/woman at a bar).

It doesn't mean a worse model. It may mean that at certain points in time now which tend to be very short lived, but model advancement will hit diminishing returns and at that point models will become commoditized. But even now which model is best is not always the SynthId models from Google.

In fact, the only thing allowing differentiation now is how compute heavy current architectures are. It's very possible this will turn out to not be necessary.

Also my logic was not "Nefarious uses require no safety rails". That was your logic you injected into the conversation. I was merely saying that nefarious users were more likely to use models with safety rails off.