Carrie Bickner

Editorial Note: This document has been updated and incorporated into the MARTI Framework master document, which provides a comprehensive overview of the latest developments. For the most current and complete information about MARTI, please visit: https://carriebickner.com/marti/

This document demonstrates how the MARTI Framweork could be implemented across domains or putput types.

Marti Output-Focused Metadata Rubric (Version 1.1)

One of the most important features of the MARTI Framework is that it makes room for a scale of “adherence” or “adoption.” When I initially conceptualized this scale, I had imagined that certain domains would require different levels of adoption, and I wanted to allow participants to be able to decide how strictly they wanted to apply the MARTI Framework.

I was initially surprised when my chatbot, Orla, suggested that, in areas where AI has more autonomy and thus the potential to inflict damage, strict implementation is always recommended.

I had initially  thought that by having a scale of adoption, from loose to strict, where loose adherence would require a smaller set of metadata elements, and strict would require full and comprehensive metadata documentation, potential participants would be drawn to the flexibility in their level of adoption.

Orla and I talked about this a lot, and I think that she is correct in that as the as the work being undertaken by humans and/or generative AI has the potential to cause harm on a large scale, think cyber security, biotechnology and medicine, and military, and where the autonomy of generative AI is going to increase over time, this scale really needs to be more prescriptive, forced or required.

I need help thinking about the levels of adoption or adherence.

I also want to look at how the MARTI Framework could be informed by and work with initiatives like the Responsible Scaling Policiy (RSP) at Antropic.

Who is with me?