AI training setting problematic

[qubodup]

Freesound did the right thing when it was time to change licenses and sunset Sampling+ by asking each user to update their license choice.

When AI training agreement was added, it was done quietly without opt-out, without TOS update or notification (as far as I can tell). The setting is wordy and seems unclear on first look.

The FAQ states "not legally binding" but if so, why does it even exist? If it's "not legally binding" then it doesn't give rights to those training AI and it does not create clarity to uploaders what they are allowing or restricting.

This all appears secretive, obfuscated and forced.

Without the user knowing anything changed, Freesound put the following words into each uploaders mouth:

I agree with my sounds being used to train generative AI models provided that: My sounds are used to train open source models that are freely available to the public and that do not allow a commercial use.

I for one would not put these words in my mouth because I do not know what they mean. Most likely I will never read the full blog post and will never fully know what this means.

At a time where cookie banners are either bad because they make it hard to opt out or ok because they have a 'only required ones' option, this step appears to be worse than bad.

---

I suspect:

1. This has no effect yet, it is a setting the result of which is not (yet) visible on the site or via API, and Freesound probably does not provide datasets in other ways where this setting is being treated as truth.
2. The impression that Freesound forces uploaders to give more permission might be accidental, based on wording.
3. Freesound might not be aware that the UI communicates that it forces a choice on the user. To illustrate: When I open https://freesound.org/home/edit/ there is no option to opt out and "My sounds are used following Freesound's recommendations for interpreting Creative Commons licenses in a generative AI training context" is selected by default without me knowingly selecting it, as far as I remember.

---

The current section looks like this:

I agree with my sounds being used to train generative AI models provided that:

[dropdownmenu]

1. My sounds are used following Freesound's recommendations for interpreting Creative Commons licenses in a generative AI training context
2. My sounds are used to train open source models that are freely available to the public
3. My sounds are used to train open source models that are freely available to the public and that do not allow a commercial use

[/dropdownmenu]

Use the setting above to express a preference regarding the usage of your sounds for training generative Artificial Intelligence models. This preference applies to all your uploaded sounds. Please, read the Usage of my sounds for training generative AI models help section to learn more about the details and implications of the available options.

It is not clear whether No 1 or No 2 is the more strict one. It is not clear whether selecting No 2 also allows No 1 or not, or whether selecting No 1 also allows No 2 or not.

---

I suggest to completely rework the option as checkboxes and give it a header:

Additional AI training permission:

In addition to the permissions given by my license choices, I permit:

[ ] No additional permissions
[ ] AI training based on Freesound's recommendations
[ ] AI training of free open source models
[ ] AI training of non-commercial free open source models

If No 1 is selected, all others get deselected. If No 3 is selected, it also selects No 4. If No 3. is deselected, it also deselects No 4. Selecting No 2-4 deselects No 1. This needs to happen client side but more importantly server side, so that the result is logical after the page is saved and reloaded.

---

Mockup:

---

Code block:

<div>
  <p>Additional AI training permission:</p>
<p>In addition to the permissions given by my license choices, I permit:</p>


  <div>
    <span class="bw-checkbox-container">
      <span class="bw-icon-checkbox" role="checkbox" aria-checked="false"></span>
    </span>
    No additional permissions
  </div>

  <div>
    <span class="bw-checkbox-container">
      <span class="bw-icon-checkbox" role="checkbox" aria-checked="false"></span>
    </span>
    AI training based on 
    <a href="https://blog.freesound.org/?p=2082" target="_blank">
      Freesound's recommendations
    </a>
  </div>

  <div>
    <span class="bw-checkbox-container">
      <span class="bw-icon-checkbox" role="checkbox" aria-checked="false"></span>
    </span>
    AI training of free 
    <a href="https://opensource.org/ai/open-source-ai-definition" target="_blank">
      open source models
    </a>
  </div>

  <div>
    <span class="bw-checkbox-container">
      <span class="bw-icon-checkbox" role="checkbox" aria-checked="false"></span>
    </span>
    AI training of non-commercial free open source models
  </div>
</div>

---

Note: https://blog.freesound.org/?p=2082 is overwhelming. Freesound might want to create a short post with only the recommendations.

---

Furthermore, it was discussed in https://freesound.org/forum/freesound-project/45363/ that the setting is not visible to users without uploads. In its current state it should definitely be visible to any user. It makes no sense to wait until a user uploads 1 sound to tell them "actually, you agree to specific AI training terms".

I would suggest just showing it to all users and keep the warning visible to all users if they haven't set it and try to upload something.

[ffont]

Hi @qubodup,

Thanks for sharing the impressions in such a clear way. Our intention was to make things as clear as possible and as transparent as possible, but it seems clear that we did not manage well.

The feature has not been announced yet, and this is because after the discussions in the forums we paused the announcement in case we want to make further changes. In practice the feature has no effect yet, because it will most likely only have effect when we add further documentation to research datasets about it, or when we get contacted by 3rd parties wanting to train models (see below).

We add this features to give more options to uploaders for expressing their preference, and these options are basically adding further restrictions on top of CC licenses (so this is not to give more permission, on the contrary, to provide different ways to restrict some). Our default option is to follow the Freesound interpretation of CC licenses as applied to AI. This is like that because as you know CC was not designed for AI, so it is not always clear how they should be interpreted in the context. This is all discussed and explained in the blog post (https://blog.freesound.org/?p=2082). We need to provide an interpretation because when 3rd parties approach us asking if they can use the sounds for training AI, we want to provide a clear answer and not let it to their own interpretation of CC (which of course they can still do if they want). We thought that having this option as default was good to provide some common ground. We thought that the alternative, which would be having an option that says "no preference", is basically the same in practice, because it basically falls back to a CC interpretation.

I agree however that the wording might not be the best. Also a checkbox system similar to the proposed in this issue might be better. We can work on improving all that.

Finally, a relevant point to discuss is the feature not being "not legally binding". Unlike Creative Commons licenses which have both the simplified license text and the deed, we don't have a "deed" for this preference. That is what we mean. It is unclear in which ways the preferences that we are offering could/should be legally enforced. This is not a Freesound thing only, simply there's no CC-like initiative providing it. That said, we still believe it is relevant. Similarly to how there's no unanimous CC interpretation in relation to AI, but when 3rd parties approach us we can make our recommendation, we can also indicate what sounds can be used (or not) based on such user preferences. Sure some 3rd parties might decide not to respect that (and some others will not even ask), but that's almost the same thing with the CC licenses and their unclear interpretation. Our will would be to update our preferences to match any sort of more standardised system in the future.

[ffont]

FYI, I'm deploying a change that fixes the wrong link issue, and also makes the setting visible even for users that have uploaded no sounds.

[qubodup]

https://freesound.org/forum/freesound-project/45385/
there's no unambiguous way to address that
we did not include a "full opt out" option for reasons explained in the conversation that I linked ( https://freesound.org/forum/freesound-project/44672/?page=3#post105846 )

(I have not read the forum thread) Freesound doesn't have to provide an 'opt out' option. It very much should let the user not make any additional statement about gen ai.

Currently a statement is chosen for users without explanation or warning.

these options are basically adding further restrictions on top of CC licenses

The options area sounds like these options are adding circumventions around the limitations of CC licenses. One would have to read https://blog.freesound.org/?p=2082 and try to make heads or tails of https://opensource.org/ai (or perhaps just https://opensource.org/ai/open-source-ai-definition ) to know what one agrees to.

The situation seems to continue to be a misunderstanding. The following video with voice audio (muted by default, unmute icon at bottom right) attempts to illustrate the situations users are put into:

freesound-ai-agreement-small.mp4