Powerful release.
Hi!
Incredible work. I was wondering if there was any chance you folks might consider a move away from openrail license? I know it's in the name of harm reduction, which I am ABSOLUTELY in support of, just I'm of the opinion that the use cases it restricts (lots of scenarios don't use models that aren't open, PLENTY of alternatives that are truly permissive) make it so that the net good this model is capable of having on the world is limited, while bad actors don't really care very much about licenses anyway, and so the openrail isn't actually much incentive to stop. In fact, most won't read or understand it.
What it DOES do, is give you the ability to take legal action. So... if you folks are actually trying to lawyer up and slam people for misusing this, which would require some pretty unbelievable auditing and time to even identify clearly if at all possible, then maybe the openrail is good for you. But EXTREMELY similar (albeit maybe a little worse!) solutions exist for the ones who WILL read the license, and they'll just go do bad things elsewhere. You wave PERSONAL liability, but when you actually permissively open something like by using apache-2.0, you've also completely waved personal liability, so that point is moot.
Alternatively, permissively opening this (while it will likely increase negative uses a little, theoretically) allows it to be used in so many more positive situations it mightn't have been used for that I believe the offset makes licensing permissively a clear decision. An open model allows this to be used in technologies like augmented and assisted speech, integrating this into disabled people's lives for example, so that they may be able to speak using devices they already have using edge compute for realistic and clear speech. It's empowering and beautiful, and it's just the first example that comes to mind. I'd never augment someone's voice, a part of themselves that they rely on to communicate with the world, with something they didn't completely own.
Regardless if there's a dataset thing that stops you or if your team just says "no, we disagree", great work with this tech.
Cheers.
Realistically, what does this license object us to do? It seems like it's pretty much permissive.
AI Summary:
If you are a developer or company: You can commercially use and build upon this model with confidence, as long as your application does not fall into the explicitly prohibited categories. You must ensure your users also comply with these rules.
If you are a researcher: You are free to study, experiment, and create derivative models, but you cannot use the model for any of the restricted purposes, even in a research setting.
Key Takeaway: This license is permissive but not laissez-faire. It encourages open innovation while legally attempting to prevent a specific set of harmful and unethical applications. It is a direct response to the unique risks posed by powerful AI models.
Thanks a lot for sharing your thoughts — we really appreciate the care and depth behind your perspective.
We agree with much of what you said, especially regarding the importance of accessibility, positive use cases, and the reality that malicious actors rarely follow licenses anyway.
For this release, though, we chose to stay with the OpenRAIL license not as a way to “police” misuse, but to clearly communicate responsible-use expectations for a technology that can meaningfully impact people’s identity and safety. In practice, OpenRAIL serves less as a restriction and more as a framework that:
- Clarifies rights and responsibilities for both developers and users.
- Helps organizations, accessibility groups, and public institutions evaluate the model with explicit ethical and usage boundaries.
- Provides a widely recognized standard that gives many teams the confidence to adopt the model in compliant and sensitive environments.
We absolutely agree that positive applications—like assistive speech or communication support—are essential, and we want our work to empower those scenarios.
Our use of OpenRAIL isn’t meant to stand in the way of that.
Rather, it allows many organizations that care about ethical requirements to more comfortably and safely integrate the model.
If you have a specific project in mind that you feel cannot move forward due to the OpenRAIL license, please feel free to email us at [email protected] with a brief description.
Depending on the project’s purpose and nature, we may be able to offer a project-specific license that better fits your needs.
We’re continually reevaluating licensing approaches as the ecosystem evolves, and feedback like yours directly informs those discussions.
Thank you again for taking the time to share this — we truly appreciate it.
Hi! Thanks for both of your responses @WansooKimHF & @ZuzEL . I really was waiting for a moment in which I had time to properly explain all of this but I'm so busy right now I'm unfortunately not able to give the responses I feel your comments deserve. Regardless, I'll offer the short version:
ZuzEL, you're EXACTLY right. It's "pretty much permissive". That is right on the money. And effectively, it is permissive, for almost anyone doing anything not evil. The problem is that there are a lot of people who are of the mind, me included, that open source releases are like human rights. If any of them are infringed upon at all, the whole thing comes tumbling down in ways that are tremendously hard to track and reproduce.
There are smarter people than I who share this opinion, like maybe the open source institute, who have great materials that are probably better than I will do in a rush here. But the idea is this:
There are generally four main protections to open source software that need to be maintained in order to truly be open.
- freedom to use
- freedom to share
- freedom to study
- freedom to modify
I'm sorry if I'm getting those wrong off the top of my head but with a little digging, again, you'll find resources that make the point better than I could!
I will say, Effectively on the ground if we are talking about trying to actually use this for alternative and augmented speech for disabled folks, I would either:
1: be making a harness for an individual person, where likely the resources are low and going through communications to actually get the "okay" is so often a lost cause in terms of time sink and effective restrictions at the end of it. Realistically I'd be coming with no proof and no outcome and asking you to trust me with a free license you're not giving to anyone else. OR
2: I'd release an open source harness that would have to have the technology included AS open source, truly. If that was the case, people would just be able to take your tech from my release and then release it open, overstepping what you're doing here. I'm a free software advocate, but I'm not advocating for maliciously overturning scientist's intentions.
There are so many options that are just truly permissive that in a lot of workflows and situations, the boss just says "only pick truly permissive stuff or I wont even look at the spec before refusing.". If I'm on the clock to make a solution, taking a shot in the dark to see if a team is going to permit some nobody to use their tech freely at high risk is not something they're going to okay. Not just that, but I build solutions for people on the ground, custom. I'm nothing, really, not even a small fish in the pond. More like a flake of fish food! I'm not even a blip on GitHub to point to. I'm a member of my community. And so at the very high level, sure, restrictions, but at the lowest level on the ground actually doing work for real humans it's also a complete non-starter for me in several different ways, and that is just my actual honest perspective and reality. I can't even imagine how many gotchas there is out there for other people in situations I don't have the perspective to see.
TLDR; it's not about the actual restrictions of the license. It's about what a license with ANY restrictions, even small, actually does. It has MASSIVE implications, way, way past what is written.
While I sincerely appreciate your respect for the serious ethical implications of machine learning development, I fundamentally disagree that openrail serves your goal. But I'm just one person!
You say openrail is doing something in practice.... you sure? How have you verified that? Have you any idea what it actually is effectively doing, or how difficult that kind of qualitative research would be to do holistically or even close to accurately? Licensing creates ripples that are so far reaching by the time the butterfly has reached the end user it's become some awful eldritch creature, just in my experience of an actual person actually doing things with people who NEED this technology, not writing papers. That isn't to glorify that, I love papers!!
ALL OF THAT SAID:
I am not making any assumptions on what this team actually thinks or has done, and I definitely don't assume it's at all malicious or because of lack of intelligence or understanding. Quite the opposite. This is genuinely very impressive work and your response was one of the better I've received, and trust me, I've done this a good number of times. You can actually audit that on Huggingface, some of them (most of them) devolve into useless fights. I've made enough substantial changes doing this though that I continue.
Cheers and hope you both have a good day