I would like to introduce to you now Niharika Singhal, and she's a lawyer with a lot of experience in IT laws, in data protection, in intellectual property as well, also in financial technology, so I don't know anything about that, but I'm really, really delighted, and also in artificial intelligence, of course. So she's a perfect fit for this session. And her talk on free software and openness of artificial intelligence, she will shed a little bit more light on terms like ethical AI, on open source AI, and responsible AI. Thank you. Absolutely. Thank you so much. And good afternoon to everybody here. As I already mentioned, my name is Neharika, and I work at the Free Software Foundation Europe. And today I'm also going to be speaking as a consortium member of the Zoom project. This is an EU-funded project that basically helps to develop tools and provide guidance to startups, companies, developers who want to know and understand appropriate licensing strategies for their business models, which is modeled on the three O's. So open hardware, open data, and open software. And yeah, so before I jump right into what my presentation is all about, I would like to give you a brief glimpse into how our website looks like, what we've been doing so far, so that it actually gives you a flavor of what my talk is gonna be about. Okay, so. Can we turn the presentation? Yeah, okay, good. So this is the website of Zoom. And as you see, we're collaborating with 14 partners across Europe. So those are our partners. And we have some universities here, innovation hubs and of course Free Software Foundation Europe. So what we do is basically collaborate on these three O's. We have materials produced on open hardware, data and software in the form of reports. So we've produced some literature review regarding legal cases. We've provided a licensing framework. And these are huge documents, which is why I'm not opening them right now. But it's for you to just know that it's there for anybody who's interested. You can definitely have a look into these documents. So we've provided a literature review. We've provided some business cases, which would be really important, interesting for, say, startups and companies. Maybe some of you would relate to these business cases that we presented here, and this would prove to be insightful for you. And we've also produced certain policy briefs that we're presenting to the Commission regarding the crux of our contribution to this project and FCFE has produced this policy paper. I, together with my colleague Lucas Lozota, have produced this policy brief number three on free software and AI openness and that's exactly what I'm going to be talking about today. So let's dive into it then. Okay, so today's talk is all going to be about the concerning trend of AI proliferation or licenses, AI licenses with additional behavioral restrictions that are based on ethical concentrations. with additional behavioral restrictions that are based on ethical concentrations. And yeah, so talking about how these licenses are based on ethical concentrations, I'd like to first enunciate as to how openness as a concept is subsumed into the definitions of free software and open source. Now free software provides you with four freedoms. Freedom to study, use, share and improve. And open source is regulated, the definition open source is regulated by OSI, open source initiative, and which states that open source is not just limited to access of source code, but also you need to comply with these criteria here. So yeah, these four freedoms that you see out here can be given by way of a software license. And yeah, so moving on to the next slide. Reasons for engaging with free software. Now there are multiple reasons to engage with free software because while proprietary licenses are fundamentally incompatible with each other, free software licenses on the other hand are are well-documented, well, they're standardized, and have fit to the complex legal issues. And so they seek to curb this problem of proliferation of licenses by increasing legal interoperability. And they also simplify license adoption. So essentially, this free and open-source movement has had a dramatic effect on the software development in these past few decades. Earlier, the proprietary systems used to lock in users, they used to block access, and as a result of this, it stagnated innovation. But by way of free software licenses like Apache and MIT, they've opened the cord to make global collaboration easier, and that's actually enabled and fostered more innovation. since I mentioned how free software and open source licenses are adding to it. If you look at it, free software licenses are actually promoting digital commons. It's promoting reciprocity, altruism, you name it, social justice. And that's great news because free software powers much of the Internet's infrastructure, and also key machine learning frameworks and libraries such as Kubernetes, PyTorch, Linux. And yeah, so free software has just emerged as an indispensable part of AI research and development, and rightly so. I would say that free software contributes to AI accessibility, reusability, and fairness. But AI systems don't necessarily operate like traditional software. They have multiple interconnected components and require distinct development processes and rely on specialized resources, mostly available in the hands of few tech companies. So although the idea of free software movement is mapped into the concept of open AI, we must be really careful because both these are fundamentally different based on how they're built. And so there are a lot of components at play here. Sometimes the model weights would be open source but the weight might not be. Also while the number of publicly available models has been growing exponentially in the recent past, what is particularly concerning, as I mentioned, is this popularization of the term open in relation to AI systems. Many of these models proclaim to be open source, but actually in reality are not using any of these open source models, any of the open source licenses. More broadly, the term open has been used to define models that provide minimal transparency and reusability to the users. So if we must apply the definition or the traditional definition of free software onto these AI systems, then we must not forget the key pillars that actually make them free and open, such as transparency. In the context of AI systems, this would mean the ability to wet and access the source code, reusability, the ability to allow a third party to reuse the source code to build their own AI models, and enablement is to disclose sufficient information that enables a third party to rebuild a model provided that they have the same computational resources as identified by the AI community. And oversight is the ability to verify and audit the source code by anyone, which actually also curtails this problem of discriminatory effects of AI that we see most often these days, which is a hot topic in the EU especially. often these days, which is a hot topic in the EU especially. So now, as you know, the concept of open AI, as I've been mentioning, is a very incumbent concept. That's primarily due to the fact that we don't really have a robust definition for AI systems, but that's going to change with the implementation of the AI ad in Europe. And also, we don't have a robust definition for what openness means in the context of AI systems. But having said that, I would actually take a moment to applaud OSI, Open Source Initiative, which has actually spearheaded this initiative of creating a robust definition for open AI, which takes into account not only the four traditional freedoms as provided by free software, that is to use, study, share, and improve, but they also take into concentration a wide gamut of AI technologies. And it fills me with a lot of pride and delight to announce that the leading partners of Zoom Consortium are also aiding OSI in this initiative. So, yeah, it's something to look out for, definitely. So, until we have this definition, what we see is that AI labeled as open actually exists on a long gradient. Now at the end of this gradient we have a handful of maximally open AI efforts and on the other end of the spectrum we have something like LAMA 2 and now more recently Lama 3, released in April, that basically proclaim to be open source. However, they actually levy a lot of restrictions onto the use and distribution of the license. And they also fail to provide any meaningful transparency to the users. Therefore, what open AI looks like in practice could vary. It's a very wide concept and could actually vary from disclosing the training and evaluation data sets of the AI to making, say, documentation regarding the data cards or models publicly available to, yeah, just making the model weights publicly available and there's this excellent paper by David Vitter and also Suleiman that dwelled into this exact long gradient of open AI and for anybody who want to who wants to deep dive into this subject can have a look at it. So given the lack of any definition, a legitimate claim to open AI must be the one that actually rests on the curated definitions of free software and open source software, which take with them a rich legacy of over 40 years of democratizing control over software. So, in the last decade, there have been diverse groups and individuals that have departed from exclusively using free software licenses to creating licensing schemes or developing new licensing schemes that levy restrictions which are related to field of endeavor, behavior, community management, commercial practices, and ethical incompatibility. And for example, in 2021, the OES had released the hypocritic license, which basically prohibits the use of software in violation of the universal standards for human rights. And now, this practice has spilled over to creation of surmoto code of ethics for AI systems with restrictive characteristics. For example, Meta's Lama 2, which has an entire appendix dedicated to its prohibited users and yeah it also requires like a special license from Meta in case you involve a large number of users and similarly Lama 3 which got released on 18th April has another index another appendix which carries the same prohibited uses, and yeah, definitely not so open source. And we have a similar long list for BigSign's OpenRELM license. So essentially, sorry, yeah, so essentially what we're trying to say is that what are the implications of use of licenses with additional behavior restrictions? First is the barriers against use and reuse. Now, as I've already displayed, and I can go back in my slides, these terms and conditions that you see here are absolutely vague and ambiguous to begin with, and I've highlighted a few here. And what happens because of the use or implication of these vague terms and conditions? It actually creates a broad overarching prohibition of its use, and this has a hindering effect onto downstream integration and application of AI systems for any of the users. Yeah. So then hurdles to adaption and improvement. Now these hurdles can be by way of unauthorizing any derivative work. And I can also give you an example which I think would be better suited for this audience since we're talking about art and free software. So I'm pretty sure everybody knows about stable diffusion. There's this foundation model which basically converts text to images. And it has this amazing ability to create accurate medical images. But if they use, and when they use the Creative ML OpenRail M license, what happens is that this license actually prohibits the use of the software for generating any images for medical advice or for any medical result interpretation. So as you see right at the outset, we have this prohibition to use the AI for bettering the health industry, for the medical industry, and so on and so forth. So those are some prohibitive or prohibitory uses of these restrictive licenses that actually claim to be open. And yeah, weakening of oversight, hindrances to control over technology. So the consequence of this long gradient of AI openness is that it increases the loss of control that the users have over their technology because of this problem of vendor lock-in. And you have to constantly depend on the vendors or the providers by way of strict control of APIs. And yeah, weakening of external oversight and transparency. So although proprietary AI models could also be transparent. Free software models actually provide you with this ability to access the source code, to verify it, and as I said, it also has this effect of reducing these discriminatory effects. And with the global explosion of AI systems or LLMs that we have these days, we are also now having a lot of standardized risk assessments, model evaluation assessments. And having a free software license really helps because by way of that, you can actually, as users, anybody can verify and inspect the source code. So it brings in more transparency to the users. So, yeah, in conclusion to our contribution to the Zoom project, we've basically come up with a few recommendations. So the first one is to basically preserve the openness in AI. And second is to keep the licenses interoperable with free software licenses. Now, we know that with the progress of time, new and dedicated AI licenses are going to be a natural progression and a much desired phenomenon. The only thing that we plead is that they should be interoperable with the free software licenses. And that is the only way that we can actually maintain the reusability, sustainability, and transparency of the AI systems. And this is also another way of maintaining the legal interoperability between the existing free software licenses and the AI licenses of the future. Then talking about ethical compliance check and also the crux of today's talk. So ethics as a subject is deeply rooted in societies and societal values that differ from one jurisdiction to another and problems of transparency, accessibility, explainability, interpretability. All of these challenges or all of these problems pose an ethically challenging problem for the deployment of AI systems. Therefore, before embedding any sort of ethics into technologies, we should be absolutely careful. And there have been various approaches that have been developed by various AI practitioners and researchers, such as creating, for example, a framework for assessing the risks of open foundation models, to creating tools for auditing the AI systems that audit the fairness, explainability, and robustness of these AI systems, to also establishing ethics review board. Now, for example, Meta has its own oversight board. Microsoft has this Aether Committee, and Google DeepMind has this Responsibility and Sustainability Council. So essentially, there are multiple ways to ensure that these AI systems are ethical. However, any restrictive terms and conditions posed onto AI systems based on ethical considerations are better suited only for regulation and not by way of licensing schemes. And that's exactly what we're pleading before the commission. In essence, there is no need, or I would say there's this phenomena of openness or perniciousness of selling everything as open. To begin with, there is no need to have everything as open and free, and sometimes the law also prohibits it. But if you're calling something free and open, you're actually equating it to the standards that FreeSoft and Open Source have created and have lived for over 40 years. And essentially you're just watering down the standards by equating your licenses to these licenses, to the free and open source licenses. So perhaps Meta had good intentions in mind, but it could call it a responsible license. It could call it whatever you would like, but not free and open source license, because these definitions are fixed. These definitions carry a meaning behind it. So, in essence, ethics could be within the purview of regulations, but not licensing schemes, because licenses are not a substitute for governance and licenses can never be a substitute for legislation. So yeah with that I would like to wrap up my talk. I hope it was interesting and insightful and give you an insight into what we've been working on in the Zoom project. So yeah, thank you very much. I'm trying to move on to the last slide, but for some reason, yeah, it's just fixed. But okay, yeah. And I also open the floor for any of the questions you want to ask now or towards the end if we have a panel discussion. So if you have questions that you can only ask now and you don't want to wait, then you can, of course, do that. But otherwise, I would suggest that we collect them in the end of the full session. And I'll just be quiet for a second to see your answer. I'd also like to quickly add, so what I displayed in the beginning, I've also captured this in the form of a QR code. You can easily scan this and look into our website and also the policy brief. So feel free to have a look and share it with anybody you think would find it useful too. Thank you.