I appreciate that you got someone to speak who understands both the technical and art sides of the issue fairly well. That said, there are a few parts that bear clarification/rebuttal: Competition in relation to fair use: The speaker uses style as an example here, but copyright does not protect styles, only actual works. Anti-AI image modification processes: The claims made by the U of C and others are considered dubious at best by many in the AI community, in part because they require the altered images to make up a substantial portion of the training data, can (at least in some cases) be undone, and do perceptibly alter some images. Copyright registration: Technically speaking, all creative works are granted copyright in the US once published. Registration is required if someone wants to pursue legal action regarding a work, however. The speaker made it sound like unregistered works have no status at all, which is untrue. Permission/compensation: I’m skeptical of such a system working out, largely due to the prospects of civil disobedience (general consumers ignoring any restrictions), models developed in other jurisdictions, and the ability to generate models locally (large firms may have an advantage in regards to the latest and greatest models, but the increasing power of consumer-grade hardware will continue to expand what kinds of models can be built by individuals or small groups). The speaker seems to assume that generative models will be confined to large corporations, which is already untrue in some domains such as voice synthesis.
I appreciate that you got someone to speak who understands both the technical and art sides of the issue fairly well. That said, there are a few parts that bear clarification/rebuttal:
Competition in relation to fair use: The speaker uses style as an example here, but copyright does not protect styles, only actual works.
Anti-AI image modification processes: The claims made by the U of C and others are considered dubious at best by many in the AI community, in part because they require the altered images to make up a substantial portion of the training data, can (at least in some cases) be undone, and do perceptibly alter some images.
Copyright registration: Technically speaking, all creative works are granted copyright in the US once published. Registration is required if someone wants to pursue legal action regarding a work, however. The speaker made it sound like unregistered works have no status at all, which is untrue.
Permission/compensation: I’m skeptical of such a system working out, largely due to the prospects of civil disobedience (general consumers ignoring any restrictions), models developed in other jurisdictions, and the ability to generate models locally (large firms may have an advantage in regards to the latest and greatest models, but the increasing power of consumer-grade hardware will continue to expand what kinds of models can be built by individuals or small groups). The speaker seems to assume that generative models will be confined to large corporations, which is already untrue in some domains such as voice synthesis.