Filter by topic and date
First Impressions from the IAB AI-CONTROL workshop
- Suresh Krishnan AI-CONTROL Workshop Program Committee Chair
- Mark Nottingham AI-CONTROL Workshop Program Committee Chair
24 Oct 2024
The Internet Architecture Board (IAB) organized a workshop on 19-20 September 2024 to discuss issues around and possibilities for practical mechanisms that publishers of data on the Internet could employ to opt out of use by the Large Language Models and other machine learning techniques used for Artificial Intelligence (AI).
Nearly 50 individuals with a broad range of perspectives considered a variety of related topics, such as use cases and requirements, possible future developments, and the policy implications of opt-out mechanisms.
The workshop was held in Washington, DC, USA under Chatham House rule. The workshop participants included AI vendors, content creators, content platforms, civil society, academia, and representatives from multiple governments. The workshop set out to answer the following questions:
- What are the success criteria for AI Control?
- What needs to be expressed?
- What technical capabilities are necessary to achieve specific objectives?
- What challenges/advantages does each approach have (opt-out, opt-in, technical blocking, etc.)?
- What are the boundaries, overlaps, and limitations in technical mechanisms and legal regulation?
The workshop chairs along with the program committee will be publishing a detailed workshop report that will provide a summary of the proceedings but there are a few early impressions and outcomes that we would like to share.
We started off by setting the stage with background and discussions regarding the history of web crawlers and opt-out mechanisms from the technical and governance viewpoints. We also heard about several regulatory initiatives and timelines that would require the definition and use of opt-out mechanisms for AI. There were also discussions about similar mechanisms from the past, including Do Not Track (DNT) and Global Privacy Control (GPC), and what kind of lessons could be learnt from them. We wrapped the first day with lightning talks providing perspectives from a cross section of attendees from different stakeholder communities.
On the second day we focused on options for the technical mechanisms for associating preferences with specific content (known as “attachment”) and for developing vocabulary for defining AI usage preferences. We also discussed how such expressions of usage preferences could be carried over IETF protocols or embedded in content metadata. There were also significant discussions on how these topics interact with licensing and copyright. We wrapped up by identifying things that would be in scope for future work at the IETF and put together a fairly focused set of potential deliverables that would need to be worked on initially. The intent is to use these as input for developing a charter for a potential BoF/WG in the IETF during the Dublin meeting timeframe.
Accepted submissions to the workshop can be found at the AI CONTROL workshop webpage.
Follow up discussions will take place on the AI Control mailing list (ai-control@ietf.org). This mailing list is open for everyone to join. Please subscribe at https://www.ietf.org/mailman/listinfo/ai-control.
There will also be a side meeting during IETF 121 in Dublin, Ireland with remote participation to explore forming a working group in the IETF to further this topic. Logistics for this meeting will be announced on the AI Control mailing list above.