Alright y'all.
We are going to start here in just a moment.
I'd like to introduce you to Nahara.
Nahara, I tried, I really did.
She is coming from Estonia.
She loves some free software, data protection, geopolitics, and wondering about different
cultures and countries.
She also spends some time doing photography and some perfumery.
Now that's interesting.
She's a certified scuba diver.
We've got multiple athletes today.
This is awesome.
This is truly awesome.
Take it away.
Thank you.
Thanks everybody.
And welcome to Vostim.
Good evening.
My name is Nahara Ika.
And I work as project manager at Free Software Foundation Europe.
And today I'm going to also be speaking as a consortium member of the Zoom project, which
is funded by the European Commission.
And it basically aims to integrate the three O's.
Open hardware, open data, and open software into an innovation-driven policy.
So first off, I'm going to set the agenda for today.
Since my talk is based on the AI license proliferation, based on ethical considerations, I'm going
to start with how openness is also an ethical consideration, followed by the reasons for
engaging with free software, and what open AI should mean in practice versus its current
state.
And finally, I'll end with what are the imposition of additional behavioral restrictions by licensing
and the implications.
So first of all, the concept of openness is subsumed in the definitions provided by Free
Software and Open Source by FSF and OSI respectively.
And Free Software provides you with four freedoms, that is, the freedom to study, use, share,
and improve.
So essentially, anybody can use and distribute Free Software by way of a license, a software
license, in a non-exclusive manner.
Then there are multiple reasons for engaging with Free Software.
Primarily, proprietary licenses, as you know, are fundamentally incompatible with each other.
But Free Software licenses are well standardized, well documented, and have widths to the complex
legal issues.
And so the C2Ca tail the problem of license proliferation by making it more legal interoperable
and also by making license adoption easier.
Yeah, and if you're talking about ethics, then Free Software also helps in providing,
or rather, promoting digital garments.
It helps to promote altruism, democratization of knowledge, and reciprocity, most importantly.
And now, if you just put this concept into AI systems, then Free Software also helps in
promoting accessibility, transparency, fairness, explainability.
So yeah.
Okay, so then AI systems don't really operate as traditional software.
There are multiple interconnected components, as you see, training data, model architecture,
and they require distinct development process and rely mostly on specialized resources in
the hands of a few big tech companies.
So the ideology of Free Software is essentially mapped into the concept of open AI, but we
must be wary about the fact that AI is built differently than a traditional piece of it.
And so there are a lot of components at play here.
So sometimes the code around the model could be open source or Free Software oriented,
but say the model isn't open source.
So it's fundamentally not the same as a traditional software.
And so what is particularly concerning is the popularization of the term open when it
comes to AI systems.
If you must apply the spirit of the traditional definition of Free Software and open source,
then we must not forget the key pillars that actually make it open, which are transparency,
reusability, oversight, and enablement.
So now transparency in the context of AI could mean the ability to access the source
code or read the source code.
Reusability could mean to enable any third parties to reuse the code, the data or documentation.
Reusability is basically enabling the ability to inspect and verifying the source code or
the documentation or rather even data about configuration of the AI system.
And enablement is basically by disclosing sufficient details of how the AI is built in order to
enable a third party to rebuild the same AI system provided that given the fact that
necessary computational resources are provided and these should be identified by the community
building of the AI.
Now as you know, the concept of open AI is pre-encombered.
That is primarily due to the fact that the definition of AI systems itself is not clearly
defined, which would change of course with the AI Act.
But the concept of openness in AI is also not clearly defined.
Now OSI has taken a great initiative in this regard by defining open source AI and while
doing so, then not only endorsing the four traditional freedoms as provided by FreeSoft
Fed, as you can see from the definition itself, but they're also trying to widen the spectrum
of the definition by including diverse types of AI technologies.
And as you see, it also says that they support the efforts on these issues, including appropriate
government regulation when it comes to having ethical component built into the definition
itself.
So until a really important addition to this is the fact, and I take immense pride in saying
this, that leading members of the Zoom consortium are also cooperating and collaborating towards
this effort.
So I'm hopeful that we would have a comprehensive definition of openness in AI, which would
really enable all AI users to use licensing schemes appropriately.
So until we have a definition that's been finalized as regards openness in AI as concerned,
what we see is that AI labeled as open actually exists on a long gradient.
Now on one end of this gradient, as you see, we have a handful of maximally open AI systems
such as Luther AI.
It's a nonprofit and it has licensed its model under Apache 2.0.
There's also GPT-New York that is built and developed by Luther AI.
And it's also made all its model weights and parameters.
Also the documentation and the data around its training and configuration absolutely
publicly accessible.
Whereas on the other hand, you have AI systems like Lama 2, which basically claim themselves
by Meta, which claim themselves to be open source.
But Ashley Forbid uses from its use to build other language models.
They also provide meaningless or rather not a very meaningful description of the language
definition or some kind of transparency regarding the data that's being used to build the AI system.
So yeah, well, now given the fact that there's any lack of definition, what we see is that
openness actually exists on a long gradient.
And if we need to use the term open or free, we need to actually conform to the principles
provided by free software and open source software, which take with them a rich history
of 40 years of success in having control over software.
So, in the last decade, what we observe is that there's been few diverse groups and individuals
who've departed from using free software licenses exclusively to creating certain licenses
that actually prioritize restrictions on the use and distribution of software.
And this primarily relate to field of endeavor, behavior, community management, commercial
practice and ethical compatibility.
So for instance, in 2021, there was Hippocratic License 3.0, which was developed and released
by OES, and this specifically prohibits its use for the use of free software in violation
of the universal standards of human rights.
And this practice has now also spilled over to creation of sewer motor ethic codes for
AI systems, and that has led to the creation of AI licenses with restrictive additional
behaviors.
For example, we have Lama 2 by Meta, and as you see, they have an entire appendix dedicated
to a lot of prohibitory uses.
There's also a similar list by Big Science Opal Rail M license.
And so essentially, what are the implications of use of these licenses with additional behavioral
restrictions?
Now, as I see, it also basically creates barriers against use and reuse.
As I've just displayed, there are certain terms and conditions which are absolutely
ambiguous, and what happens is that the use of these wake terminologies create a very
overarching prohibitory use for downstream integration and application of AI systems.
And hurdles to adaption and improvement, this is basically by unauthorizing derivative
work or by prohibiting copy left licenses.
Hindrance to control over technology.
So the consequence of this long gradient of openness of AI has led to the users not having
appropriate control over the technology because of, you know, by blocking interoperability.
And yeah, a weakening of oversight and transparency.
So proprietary AI systems could also be transparent, but free software basically provides the ability
to rep the source code and also improve it.
And this also helps to minimize the discriminatory effects of AI systems.
So in conclusion to our contribution to the Zoom project, we majorly recommend for recommendations
to everyone first is preserve openness in AI.
Now there's been a dissonance amongst the marketing pitch of these AI efforts versus
restriction to software freedom, and this disables control, transparency, and oversight
over technology.
So there is an imperative need to preserve the openness.
Then we talk about keeping licenses interoperable with free software licenses.
Now the emergence of dedicated AI licenses is perceived as a national progression and
a well-desired phenomenon.
We only plead in the bargain is that these licenses should actually be interoperable
with the free software licenses so that we actually make the AI systems reusable, accessible,
and sustainable.
And yeah, talking about ethical compliance.
Ethics is actually deeply rooted in societal values, which actually differ from jurisdiction
to jurisdiction.
So in order to actually apply any ethical restrictions, we need to be very, very careful before embedding
these into technologies.
Any kind of restrictive practices based on ethical considerations should be under the
purview of law and regulation and not licensing.
Schemes because licenses aren't a substitute for regulation and they cannot be a substitute
for good governance and for legislation.
So yeah, with this I'd like to wrap up my talk.
I hope it was insightful and I hope the message is loud and clear.
If you call yourself as an open, licensed, following the principles of open-source software
or free software, we should and we must respect the freedoms that it comes or it rather provides.
So yeah, thank you.
Unfortunately, we don't have time for time for any questions.
Yes.
Yes.
All right, we will be starting with Steph here in just a couple minutes.
So don't go in too far.
Yes, sir.