2023/10/27 17:45:41

make predictions, recommendations, or decisions influencing real or virtual environments

Reporting here a comment from previous version: this list of tasks for an AI system is not enough to cover current practices — in particular, generative AI ones.

smaffulli

2023/10/27 23:59:07

this is what the OECD and other regulations around the world adopted as a definition of AI system. Generative AI (like copilot's suggestions or DALL-E images) are suggestions, in this context. This is the document that introduced that definition: https://read.oecd-ilibrary.org/science-and-technology/scoping-the-oecd-ai-principles_d62f618a-en#page6

smaffulli

2023/12/01 10:59:09

The OECD updated its definition in November, I think it addresses this comment.

zack

2023/10/27 17:47:03

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”.

Having this outside the definition is a problem. Not how, for example and reference, in the Free Software definition "access to source code is a precondition for" is part of the definition, not outside of it.

Aspie96

2023/11/20 05:46:28

Agreed. A viable approach could be defining what an OS license is (which, of course, will not mention this point) and then what an OS AI system is: one under an OS license which is provided in the preferred form for modification (which might just mean an open and DRM-free format, preferably not quantized, for many kinds of model). (Edited)

zack

2023/10/27 17:50:57

For AI systems, the preferred form to make any modifications to the work depends on the specific kind of AI.

I find this problematic, because it is going to be a very important detail. If this is left out of the actual definition, then OSI will have from day 0 questions about what is part of the AI system and will end up having to maintain an "authoritative FAQ" that explains how to interpret the definition. While that might end up happening anyway, we should try to put "core notions" in the definition upfront, and it looks like we are doing the contrary here.

smaffulli

2023/12/01 10:57:51

Agreed. Let's get to the next phase of drafting, now that we have at least an early agreement about the initial freedoms we need to exercise.

zack

2023/10/27 17:51:31

[Provide an example, based on machine learning?]

The problem is that one example is not gonna cut it. We are going to need many of them, and anyway "definition by example" is not a great approach. It would be better to have a main definition of what is part of an AI system *in* the definition itself, so that when we talk about the right of modifying an AI system, people will understand what it means in practice.

smaffulli

2023/12/01 10:56:45

Agreed, we'll need more than one example

smaffulli

2023/10/28 00:00:50

decisions

A later (2022) OECD document removes the word "or decisions" saying in the notes ```Experts Working Group decided should be excluded here to clarify that an AI system does not make an actual decision, which is the remit of human creators and outside the scope of the AI system``` https://www.oecd-ilibrary.org/docserver/cb6d9eca-en.pdf?expires=1698451383&id=id&accname=guest&checksum=0673801DD780EFD29DCB04855221F918 (Edited)

Arun Gupta

2023/10/30 02:34:57

components

Are the components called out somewhere? Is it model, weights, data, or something else? (Edited)

smaffulli

2023/12/01 10:27:49

this is going to be the next step, identify components and the legal terms they're made available with

kwinkunks

2023/11/08 11:51:58

to change its recommendations, predictions or decisions to adapt

"...to adapt its recommendations, predictions or decisions" is less wordy.

kwinkunks

2023/11/08 11:53:47

Needs a question mark.

Shuji Sado (佐渡秀治)

2023/11/11 05:47:43

Use the system

Does "use" mean running the system? Or does it include copying or modifying?

smaffulli

2023/12/01 10:29:30

this should mean use the AI system to get its output. Modify and copy are different verbs

Shuji Sado (佐渡秀治)

2023/11/11 06:53:19

your

It is difficult to know who "your" are referring to. It seems to me that it could mean that it is limited to the person reading this document.

Simon Phipps

2023/11/14 08:28:10

available

should this specify "in the preferred form making modifications" at this point (see @Zack above) (Edited)

Simon Phipps

2023/11/14 08:31:00

Presumably Share also means "deploy for use by others" (since "Use" is a different right). Maybe say "Share and/or deploy"?

Wichmann

2023/11/17 22:40:23

s o

Simple thing, but for clarity, can we consistently use the serial comma? https://www.adamsdrafting.com/more-syntactic-ambiguity-the-serial-comma/

smaffulli

2023/12/01 10:39:11

we'll do a grammar editing once we have the final document.

Aspie96

2023/11/20 05:28:20

Use the system

The freedom to use the system should be the first freedom, not the second. This is because it's the most important freedom and in order to align with the definitions by the FSF, the EFSF (which has a different definition for "free software", but the same fist freedom) and the definition of free cultural works. SUGGESTION: Change the order of the freedoms so that the freedom to use the system is the first.

Aspie96

2023/11/20 05:31:14

to change its recommendations, predictions or decisions to adapt to your needs

This is way too narrow. One could desire to change the system in any other way too. It's possible that a modification does not lead to a difference in recommendations, predictions or decisions. One could change quantization or even apply lossless changes for whatever reason that produce different architectures but the same outputs. SUGGESTION: Reword this so that any modification of any soft whatsoever, including mixing with other models, is allowed, be it to change outputs or not.

Aspie96

2023/11/20 05:32:54

As an additional note, whether the modification has the purpose of adapting the model to one's need should not matter. One could wish to apply a modification they don't "need" and should be allowed to. SUGGESTION: replace "to adapt to your needs" with "as you wish".

Aspie96

2023/11/20 05:36:46

For AI systems, the preferred form to make any modifications to the work depends on the specific kind of AI.

The source code is the preferred form to make modifications to the work. If the work never existed in a format which is easy for modification, it still (always) has a source code: whatever form is the best (aka "least bad"). At minimum, however, it must be in a format compatible with open source software: a free and DRM-free format. SUGGESTION: Address free/open and proprietary formats somewhere.

smaffulli

2023/12/01 10:37:47

That's the intention. We'll clarify this topic as we add more details to the definition in the next drafting phase

Aspie96

2023/11/20 05:43:10

Share the system with or without modifications, for any purpose

This should mention sharing as open source only. This is to allow copyleft licenses. Personally, I don't think using copyleft licenses for AI models would be a good idea at all. However, copyleft licenses are definitely open source and, desirable or not, should be allowed by the definition. SUGGESTION: add "as open source".

smaffulli

2023/12/01 10:36:57

this feels redundant.

Aspie96

2023/11/20 05:49:01

Out of scope issues

I couldn't possibly agree more that this is out of scope. Should this even be mentioned, however? How to deploy a system which is accurate and works well is arguably even more crucial. But it's also out of scope, so it's not part of the document.

smaffulli

2023/12/01 10:23:43

Given how some of the stakeholders are focused on field-of-use restrictions, It's important to state here that this is not the place for those conversations. I agree that this sentence is not enough to explain the depth of the issue though. We may want to link to a deeper document explaining why these field-of-use and other topics are not fit to be included in the OSAID. Do you feel like taking a stab at drafting such document?

Luis Villa

2023/11/27 05:00:26

licenses

“licenses” is too narrow, given that many AI systems will be available under non-license legal frameworks (eg terms of service or terms of use). I’d simply say “legal terms” here.

Luis Villa

2023/11/27 05:02:14

unrestricted

Restricted by who? All AI systems will have restrictions on their usage legally required by regulators, and many will have restrictions on their training data from the same source.

Luis Villa

2023/11/27 05:03:47

discussion

OSI should clearly and explicitly say “and we support efforts to work on these issues, including through appropriate government regulation.”

The Open Source AI Definition

version 0.0.3

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; and 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)

Open Source has 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 adhere to 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 reliable and transparent.

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 any 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 or responsible, although it doesn’t prevent it. What makes an AI system ethical or responsible 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 system works and inspect its components.
Use the system for any purpose 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 with or without modifications, for any purpose.

[Provide an example, based on machine learning?]

Checklist to evaluate licenses

TODO

The Open Source AI Definition

version 0.0.3

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