Consent Is A Winning Strategy
I was working on a provisional patent draft for a new component of the Sublimate platform last week. I started writing about additional benefits of our method, above and beyond the direct, technical one. I wrote that the data collection our method relies on requires the participation and consent of the human subject being generated, and how that participation and consent improves the trust the subject has in the model. I explained how participation and consent also improve the provenance of the data the model is trained on — as compared with foundation models trained on data of questionable provenance.
(Clockwise from top left) Generations from Advisor Ian Colon, Advisor Christina Greer, Co-Founder & CEO Emi Kolawole and Founder & CTO Sarah Laszlo’s Sublimate models.
What I realized after writing all this, though, was that I had the relationship between our technology and the participation and consent of our generation subjects backwards. It's not that our users’ trust has an incidental side benefit for our technology. It's that the participation and consent of our users allows us to build models that could not be built without their participation and consent.
Once I realized that, a whole list of other model capabilities that we could build with similar data collection techniques started emitting from my fingers — capabilities that we can have because our users participate in the creation of the data that their models are built on.
I got excited about this. Part of the reason that I started Sublimate was that I spent a lot of time thinking quite hard about the abuses to human dignity that can arise from the models that do not forefront the participation and consent of the individuals in the training dataset. We all know: AI that can generate indistinguishable likeness depictions of individuals are coming. And with them, the potential for deepfakes, nasty stereotypical depictions of every kind, regrettable damage to brand integrity, and socially and culturally unacceptable depictions of people are all real risks that come along for the ride.
The solution to this, though, is not to give up on generative AI — a conclusion that would have made my younger self, a futurist and an AI optimist, very sad. The solution is also not to put a new band-aid on a foundation model every time it generates something objectionable — because, the number of objectionable possibilities and the creativity of bad actors are both infinite; band-aids on foundation models are an unwinnable arms race. Rather, I have come to believe, the solution is to build models of individuals that are controlled by individuals.
As we have done this at Sublimate, we have found that there is a joy that comes from being able to depict oneself however one wants, without the fear of what a bad actor might do with one's likeness.
We are constantly amazed by the creativity and individuality that we see unleashed when we deliver a new model to a new user. This is the real promise of image generation models. This is the antidote to the abuses of foundation models. This is what we are building at Sublimate. Join us?
