2023/10/12 07:50:19

Verify that the system you’ve received gives the recommendations, predictions or decisions it’s supposed to give

While I think this is remarkably important, I believe this is covered by the first point and second point. Maybe it can be added to the second, or perhaps third point (... and verify) This would make the freedoms back to four, but that's just a lateral remark, not a requirement.

pchestek

2023/10/16 01:15:02

"supposed to give" is problematic for me - according to whom? If I intentionally bias the model, it will be giving the recommendations I think it is "supposed to give." It also has an air of value judgment

Guest Keller2023/10/19 13:03:52

My concern here (which we discussed at length in the ATO workshop) is that as this is technologically infeasible for non-deterministic or probabilistic systems, which is a large amount of algorithmic applications in the ML/AI space. As stated, this assumes that all outcomes are able to be found repeatedly and according to a testing criteria. That's just not the case for many kinds of ML/AI, or even for a majority of them.

Guest Singleton2023/10/24 00:35:06

I agree with Carlo that this should be covered by the first point of "Study" to understand how it comes to its recommendations, predictions, or decisions and with Pam that its unclear what an AI system is "supposed" to give.

Carlo Piana

2023/10/12 07:54:36

Being Open Source doesn’t mean an AI system will also be safe, secure, trustworthy, explainable and fair but it won’t be an impediment either.

True statement, but is it called for? Maybe we should turn it to positive: being Open Source would increase the chances to make it so. Also I would change "mean" with "guarantee", or "cause", which is more precise and referring to an effect of a status, or part thereof.

Anne-Marie Scott

2023/10/12 13:30:11

If this is being kept in some form, watch language - it's not consistent with "fair, reliable, transparent, trustworthy, secure and safe" at the start. And I prefer "explainable" over "transparent" since it has some defined meaning in the field.

smaffulli

2023/10/12 21:21:07

both terms, transparent and explainable appear in different policy documents. We may need to define a collective term for all of these adjectives to simplify the language.

pchestek

2023/10/16 01:18:36

I'm very confused by it. At the beginning we say these are the assurances that open source licenses provide but they aren't going to necessarily true for Open Source AI-licensed systems?

Carlo Piana

2023/10/12 07:58:50

learning

I think we should at least stress the basic training material and process. In my experience, one of the benefit of Open Source is reproducibility (getting the sources, you can end up with a fully working application without relying on the pre-built one), which is necessary for making sure software works as intended and as described in the published source code, even if full reproducibility is a chimera. In AI it's all the more a chimera, but it is nonetheless a desirable goal, let's state it.

Carlo Piana

2023/10/12 08:03:50

AI adopted

But we are not necessarily bound to it. A branch of AI for instance is object boundary recognition in a picture. Does this fall within "predictions" (no), recommendation (probably not), decisions (probably not). But this does not matter, if it's not a hard definition.

Seo-Young Isabelle Hwang

2023/10/12 12:42:34

Adopting a global organization's definition makes sense for universal consensus. But we also need to check Carlo's comment in above. Because AI has evolved over time and will continue to do so. So.. below is a rough idea. We might start with an easy, general definition and then adding more definitions. For example, 'an AI system is a machine-based system that demonstrates human intelligence or behavior. (refer to Emerging Tech Series AI v1.1 (un.org)) and add UNESCO's one for more precise.

rvs

2023/10/19 11:05:44

Great questions, but I feel for me the meta question here is: why are we even including this section here? My answer is to scope our efforts to a particular kind of technology. As any scoping exercises go it always is (a) impossible to define with a 100% crisp boundary (b) extremely important to put some kind of a stake in the ground In that sense I feel that while Carlo's comment is valid, it still falls within a blurry boundary of the UNESCO definition.

Thierry Carrez

2023/10/12 12:03:49

decades of Open Source

"Open source" as a term was coined in 1998, so this phrasing open up to the usual rebuttals. I'd say "Open source has demonstrated..." instead.

amanda casari

2023/10/19 13:21:47

+1 to adjusting to "Open source has demonstrated that..."

Thierry Carrez

2023/10/12 12:09:25

How we can get the benefits of Open Source AI

This paragraph feels out of place in the preamble, as it's part of evaluating Open Source AI (once defined). I'd move it later in the doc, for example to the "checklist" part

Thierry Carrez

2023/10/12 12:14:27

explain how it comes to its recommendations, predictions or decisions

I think this is good, but also very likely to trigger a lot of discussion. Given enough complexity, AI systems exhibit emergent properties that are hard to explain (even LLMs, which are relatively simple "next word" generators, start to exhibit emergent properties)

zack

2023/10/14 13:46:48

Counterpoint: this requirement states that the license must guarantee the freedom to "*study* how", not to achieve the full understanding of. (You might think of this as an obligation of enabling the study, not as an obligation of achieving explainability.)

Guest Mendez2023/10/12 13:13:13

Open Source Definition

Link to the OSD? This is going to go far and wide to unknown audiences we hope, who may not be familiar.

Anne-Marie Scott

2023/10/12 13:21:26

(this is A-M before logging in)

Anne-Marie Scott

2023/10/12 13:27:03

fair, reliable, transparent, trustworthy, secure and safe

"fair" and "safe" could be read as loaded terms and might seem to contradict what we say is out of scope?

amanda casari

2023/10/19 13:29:56

(I was Guest Keller before logging in) From this comment thread, it sounds like we all find these terms to be potentially admirable but also areas which would be the desired effect (not the same as the definitions). I'd like to propose changing this section to be: "Open Source AI needs to ensure essential freedoms which enable users to build and deploy AI systems that are atomic, independent, verifiable, and modifiable." (Edited)

amanda casari

2023/10/19 13:32:54

Reasoning here (and open to new words): Atomic - "Open Source AI" should be defined by components which are parts which can be integrated into many different technology stacks. Independent - "Open Source AI" components should be able to integrated or be integrated into many different kinds of tech stacks. It's not Open Source AI if it only runs on a specific proprietary stack.

amanda casari

2023/10/19 13:35:01

2/x Verifiable - We talked @ ATO that if inspectability and transparency is desired, there are many ways to ensure this across technology systems to ensure that what you wanted is what you got. (we centered the conversation on "provenance") Modifiable - This is key and should be part of the definition to allow for a spectrum of interpretations and licenses (similar to the current OSD).

smaffulli

2023/10/12 21:04:24

Fair point: the text is here is to highlight that Open Source AI can be fair, secure, etc. It's not to say that OSAI *is* fair or safe. Is there an agreement that it should be possible for a fair/responsible AI to also be Open Source?

zack

2023/10/14 13:38:51

(minor) "follow" a definition sounds unclear to my (non-native) ears. Maybe "adhere to" would be more precise?

zack

2023/10/14 13:41:20

make predictions, recommendations, or decisions

I also have doubts about adopting this definition. In particular, it seems to me that the case of generative models is not covered by the "make predictions, recommendations, or decisions" part of the UNESCO definition. Whereas our definition is not only about generative AI systems, it is certainly *also* about them.

MZhou

2023/10/18 18:15:06

I agree. better to replace all of them as "outputs", maybe. I commented on the other sentences with the there words

rvs

2023/10/19 11:09:07

I guess I agree that GenAI doesn't quite fall into the same "blurry boundary" of this definition. That maybe a problem for us. Not sure whether we're willing to craft our own definition (based on UNESCO) OR there's a way to shop for a more precise definition from global organizations.

amanda casari

2023/10/19 13:21:08

If the goal here is to target specific classes of algorithms when used in technology systems, these do define certain types of those. As phrased, this also can define software patterns and algorithms as simple as adding in a linear regression with a feedback mechanism into your application. In short: if someone has inputs into a model which is using many different kinds of statistical methods, and incorporates those next time, it would also fit under this definition.

zack

2023/10/14 13:43:54

although it doesn’t prevent it

shouldn't we be bolder than this? In particular, we should say that Open Source AI is a necessary (but not sufficient) condition for ethical/trustworthy/responsible AI

pchestek

2023/10/16 00:52:15

"; AND a checklist ..."

pchestek

2023/10/16 00:56:07

autonomy, transparency, and collaborative

I would add "predictability" as a benefit

pchestek

2023/10/16 01:11:06

without any limitations

This is not quite accurate. The OSD allows for some limitations, e.g., copyright notice. In particular copyleft is a significant limitation

rvs

2023/10/19 11:19:48

At the risk of splitting hairs here: I feel that "use" is a loaded term. I think we need use vs. modify dichotomy (but perhaps with a better term than use). The idea being that we're striving for maximum end user freedoms (where the system is staying as-is -- no modifications whatsover) while we may agree that "derivative works" type of limitations may still apply.

Luis Villa

2023/12/01 02:08:15

I would have said “without limitations that further the other purposes of the definition”. I realize that’s a big loophole, but smaller than just removing “without restrictions” altogether.

pchestek

2023/10/16 01:11:57

without any limitations

Same comment as about about limitations

MZhou

2023/10/18 18:10:33

ditto. even the most permissive licenses requires people to keep the original license and copyright notices.

MZhou

2023/10/18 18:01:13

make modifications

"make any modifications" should be better than "make modifications". "Any" requires full access to all the details of the model, data, code.

rvs

2023/10/19 11:12:04

+1 to adding "any" here

MZhou

2023/10/18 18:02:24

trustworthy

Isn't being "open source" a prerequisite to be "trustworthy"?

MZhou

2023/10/18 18:05:35

recommendations, predictions or decisions

machine learning models are basically either discriminative or generative. recommendations, predictions, and decisions are pretty much the same thing. And they don't cover the outputs of generative models. I suggest simply replacing the sentence as -- explain how it comes to its outputs. In this way we cover any type of models.

MZhou

2023/10/18 18:06:59

the system

Does "the system" contain the original training data?

rvs

2023/10/19 11:16:33

Why is this relevant? Personally I'm reading system to be a shortened version of "AI system" -- just a noun for what we're describing here.

MZhou

2023/10/18 18:09:15

recommendations, predictions or decisions

this only covers some usage of AI systems. Should use "modify the system to change any of its internal components as well as its inputs and outputs to adapt to your needs". The current clause only covers outputs, but do not covers other usage

MZhou

2023/10/18 18:13:39

[Provide an example, based on machine learning?]

Maybe provide examples in the appendix? It can be difficult to pick the best examples in the main text.

rvs

2023/10/19 11:12:25

[Provide an example, based on machine learning?]

Is there an action item for somebody to work on this?

rvs

2023/10/19 11:13:59

ethical, trustworthy or responsible

This needs to be synchronized with "Why we need Open Source AI" paragraph

rvs

2023/10/19 11:25:03

TODO

Idea: while building this paragraph, should we start from the list of existing licenses we'd use as guinea pigs for the evaluation criteria we would be using? Obvious things like RAIL, etc. The expectation would be to have a sort of a "unit test" bundle of licenses where we would run any updates to this Checklist through that list of licenses and see whether it would make sense or not.

The Open Source AI Definition

version 0.0.2

Image Not Showing Possible Reasons

The image file may be corrupted
The server hosting the image is unavailable
The image path is incorrect
The image format is not supported

Learn More →

Note: This document is made of three parts: A preamble, stating the intentions of this document; the Definition of Open Source AI itself; a checklist to evaluate licenses.
We follow the definition of AI adopted by UNESCO https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449

An AI system is a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. AI systems are designed to operate with varying levels of autonomy.

Preamble

Why we need Open Source Artificial Intelligence (AI)

The decades of Open Source experience have demonstrated that massive benefits accrue to everyone when you remove the barriers to learning, using, sharing and improving software systems. These benefits are the result of using licenses that follow the Open Source Definition. The benefits can be distilled to autonomy, transparency, and collaborative improvement.

Everyone needs these benefits in AI. We need essential freedoms to enable users to build and deploy AI systems that are fair, reliable, transparent, trustworthy, secure and safe.

How we can get the benefits of Open Source AI

A precondition for a system to be Open Source software is that developers must have unrestricted access to the “preferred form to make modifications to the work”.
For AI systems, the preferred form to make modifications to the work depends on the specific kind of AI.
[Provide an example, based on machine learning?]

Out of scope issues

The Open Source AI Definition doesn’t say how to develop and deploy an AI system that is ethical, trustworthy or responsible, although it doesn’t prevent it. What makes an AI system ethical, responsible or trustworthy is a separate discussion.

What is Open Source AI

To be Open Source, an AI system needs to make its components available under licenses that individually grant the freedoms to:

Study how the AI system has been built, inspect its components, understand and explain how it comes to its recommendations, predictions or decisions
Use the system for any purpose without any limitations and without having to ask for permission
Modify the system to change its recommendations, predictions or decisions to adapt to your needs
Share the system unmodified or as modified by you, for any purpose without any limitations
Verify that the system you’ve received gives the recommendations, predictions or decisions it’s supposed to give

Being Open Source doesn't mean an AI system will also be safe, secure, trustworthy, explainable and fair but it won’t be an impediment either.
[Provide an example, based on machine learning?]

Checklist to evaluate licenses

TODO

The Open Source AI Definition

version 0.0.2

Preamble

Why we need Open Source Artificial Intelligence (AI)

How we can get the benefits of Open Source AI

Out of scope issues

What is Open Source AI

Checklist to evaluate licenses

Read more

Answers to frequently asked questions

Checklists-MOF

The Open Source AI Definition v1.0-RC2

The Open Source AI Definition v.0.0.8