I'm concerned about the ethics of not publishing the work as open source, given that diffusion models are unethically sourced - trained from non-consensually scraped data. There is also the question of whether models can be ethically sourced - even if we buy the rights to a huge collection of images, there is the contention that many of the original creators would never have anticipated or accepted AI use of their works. So virtually no matter what, there will be outstanding ethical claims against such AIs.
Given all that, I believe that an adequate compensation for unethical sourcing in AI - the absolute best we can do for humanity - is this:
1. We admit that the works are unethically sourced, which means they could be banned the future, and may require switching to an "ethical piracy" distribution model
2. We ensure that the models produced this way belong to the entirety of the humanity by default, by distributing them for free under copy left licenses like GPL3
3. We abstain from monetizing or otherwise drawing revenue or profit from unethically sourced models
4. We assemble zero cost service models for the AI, drawing on volunteers to publicly pool compute
Case Study: Whisper
I follow these guides in my own work on OpenAI's Whisper. Whisper can do much good for humanity, as it allows them to freely transcribe and translate speech, meeting a core human communication need.
But whisper needs many improvements before it is a freely available service making an impact in millions of people's lives, for zero cost. To that end I'm building extensions that let people pool CPUs and other cheap hardware to put together independent and free transcription services based on whisper. I'm building rapid customization models to help people with accents. And I'm building real-time feedback and correction models to enhance the accuracy of the naive model
Yes, even here we are faced with the possibility of future bans, especially given the unethically sourced nature of whispers training set. That means I have chosen to abstain from ever collecting revenue through my whisper work, and consider my whisper work to be my contribution towards the legacy of all of humanity.
I encourage you to consider the upsides of this form of engagement.
Objection: How Do I Feed Myself?
Yes, this model requires you to have a fully ethical alternative business that you run. I support myself on about 25k-60k of CAD design revenue through my semi-automated CAD reverse engineering service, which is based on fully ethical automation models that I built by hand and calibrated ethically on my own work stream.
Objection: This does not protect us from bans.
No, adopting this model does not prevent the law from banning models like yours in the future. It can even enable legal audits of your code, and make you seen as a "flight risk" - someone likely to continue illegally distributing models for ethical reasons, even after they have been banned for legal reasons. I have no good answer here yet.
Closing Statement: Choose the ethical high ground. It is Based, and that will guide you.
I encourage you to see this strategy as "Ethically Based". I define Ethical Basedness as a form of true ethical high ground that can guide you towards the best contribution you can make towards the shared knowledge of humanity.
Given the extreme ethical quandaries of attempting to monetize a proprietary service on top of unethical AI, you stand to lose that higher ethical authority, and be cut off from its guidance. But you have an opportunity, now and in the future, to pick it back up.
Good luck and keep up the good work. I hope I have moved your views, rather than merely agitating your feelings. I hope you will embrace the need for a free and openly shared AI legacy for all humanity.
Overcome the ethical quandaries of AI sourcing by giving it away.
Provide freely for the core human needs of creativity and visual story telling.
Consider it.