[ai-control] Re: Considerations for AI Opt-Out

Mark Nottingham <mnot@mnot.net> Sat, 18 May 2024 03:22 UTC

Feedback-ID: ie6694242:Fastmail
Content-Type: text/plain; charset="utf-8"
Mime-Version: 1.0 (Mac OS X Mail 16.0 \(3774.600.62\))
From: Mark Nottingham <mnot@mnot.net>
In-Reply-To: <AM5PR01MB299431A3FBD50BD1F64E64F6A8EE2@AM5PR01MB2994.eurprd01.prod.exchangelabs.com>
Date: Sat, 18 May 2024 13:21:56 +1000
Content-Transfer-Encoding: quoted-printable
Message-Id: <3C9CA076-63D2-4E82-A13B-A02C6921B5F0@mnot.net>
References: <66D05E57-6474-414D-AF4C-409B063E970F@mnot.net> <AM5PR01MB299431A3FBD50BD1F64E64F6A8EE2@AM5PR01MB2994.eurprd01.prod.exchangelabs.com>
To: Chris Needham <chris.needham=40bbc.co.uk@dmarc.ietf.org>
Message-ID-Hash: H2YRLFOYICDZR22SSVVYZPQPVPLBT7U5
CC: "ai-control@ietf.org" <ai-control@ietf.org>
Precedence: list
Subject: [ai-control] Re: Considerations for AI Opt-Out
Archived-At: <https://mailarchive.ietf.org/arch/msg/ai-control/TKaTOuM6gu6KpvGiPL7uS1HLAbY>

Hi Chris,

> On 18 May 2024, at 12:31 AM, Chris Needham <chris.needham=40bbc.co.uk@dmarc.ietf.org> wrote:
[...]
> There needs to be a solution for a global opt-out like this, and I'd suggest also more fine grained controls for particular "AI" use cases like generative AI training LLMs, detection and classification, or possibly for allowing some crawlers and not others.

Yes! Part of the problem here is that AI is still evolving rapidly, so techniques and applications may emerge that test the assumptions of any framework established. 

At the same time, we should be wary of a framework that's so detailed that it creates friction against adoption (e.g., P3P), or one that's so extensible that it can be subverted by powerful implementers (e.g., SOAP).

> For specs like IPTC and CAWG's that embed the opt-outs in image or video assets, when content is voluntarily uploaded to such sites (rather than being scraped by them), would the social site's T&Cs apply, overriding any assertions in the content itself?

Good question. I suspect precedence needs to be specified, or each site will decide however is most advantageously to it. That might mean the most restrictive policy available always wins.

It also hints at another issue -- content on platforms that intermediate and control, like what is being seen now on StackOverflow and Slack:
  https://arstechnica.com/tech-policy/2024/05/slack-defends-default-opt-in-for-ai-training-on-chats-amid-user-outrage/

In other words, easy content creation tools that live within walled gardens controlled by platforms whose interests often don't always align with those of their users. 

That content is often textual (which isn't amenable to embedded policy), and/or lives entirely within the platform, which can choose what affordances for expressing preferences are available to users (with attendant dark patterns).

My initial thought is that technical controls aren't really effective in these situations; they require policy intervention and oversight (because of the likelihood of dark patterns being used to steer users' actions). That said, we should consider how those cases might interact with technical controls.

Cheers,

--
Mark Nottingham   https://www.mnot.net/

[ai-control] Considerations for AI Opt-Out Mark Nottingham
Re: [ai-control] Considerations for AI Opt-Out Michael Prorock
Re: [ai-control] Considerations for AI Opt-Out Mark Nottingham
[ai-control] Re: Considerations for AI Opt-Out Chris Needham
[ai-control] Re: Considerations for AI Opt-Out Mark Nottingham
[ai-control] Re: Considerations for AI Opt-Out Martin Thomson