Suggestions
Bogdana Rakova
Responsible AI at DLA Piper // GenAI socio-technical evaluation and safety // ex-Mozilla Fellow
Bogdana Rakova is a seasoned machine learning research engineer with 8 years of industry experience and a track record of 4 patents for innovative ML products.
With a focus on ML risk assessment, Bogdana has led projects in various industries and conducted applied research in ML, multi-stakeholder engagement, data, and AI governance.
At the Mozilla Foundation, Bogdana advocates for a community-led approach to algorithmic and artificial intelligence systems with research in multi-stakeholder engagement, legal innovation, and AI literacy.
Her background includes roles at organizations like Accenture, IEEE, Partnership on AI, and Samsung's Think Tank Team, with a degree in computer science from the University of Sofia St. Kliment Ohridski and additional studies at Harvard Kennedy School and Singularity University.
Bogdana has contributed as a Senior Data Scientist, Research Affiliate, Board Member, Founder, Senior Fellow, Research Manager, and Lead Guest Editor in various prestigious organizations and initiatives.
Highlights
What does the right-to-repair movement have to do with AI and algorithms?
It was a great privilege to organize a panel discussion on “The Symbiotic Right to Repair an Algorithm: Trust, Transparency, and Democratic Design” at the Internet Archive #DWebCamp conference.
Aza Raskin, B Cavello, Zachary Schlosser, and I engaged with the audiance in an active conversation, collectively imagining pathways for repair.
The key questions included - What if “the right to repair” was already here in the context of how we interact with AI-based systems? What would it mean, how would it look like, what would it take, what needs to exist? What would it look like if we could “repair” trust relationships in the context of the way we interact with algorithms? What does meaningful voice and responsibility mean in the context of AI innovation?
There's a lot to learn from the right-to-repair movement in software, the challenges with existing copyright laws, fair use, ownership models, legal design interventions, user-centered innovation in the intersection of AI and Law, and meaningful transparency artifacts for AI systems. I argued that, we need to acknowledge and consider the limitations of existing models of consent and think about what new kinds of consent and contestability infrastructure could enable such a right-to-repair. Read our recent paper on this together with Dr. Megan Ma and Renee Shelby, PhD here: https://lnkd.in/gJfYGD7n
Learn more about recent Generative AI copyright legal battles in this excellent article by Julia Barnett: https://lnkd.in/gyGvcRah
By far, one of my favorite #DWeb moments at the conference was a panel on "Women Shaping the Decentralized Future", by Wendy Hanamura, Mary Lou Jepsen, Mei Lin Fung, Tracey Jaquith, and Jay Graber.
It felt so inspiring to be among a powerhouse of women spearheading world-changing technologies and making AI work for everyone.
Grateful for old and new friends and community! Thank you Wendy Hanamura, Joy Chesbrough (she/her), Liz Henry, Silona Bonewald, Brian Behlendorf, Boris Mann, Aza Raskin, B Cavello, Zachary Schlosser, Joshua Tan, Danny O'Brien, Metagov, Internet Archive, and many others for helping to make this all possible.
How do you imagine a relational, participatory, and socio-technical approach to the evaluation of emerging #generativeAI applications?
Evaluation can take many forms, for example, asking experts to label data in a test set, red teaming, bug bounties, focus groups, citizen juries, and others.
🎉 In a new Mozilla article, I share the outcomes of a 4-month collaboration sprint in partnership with LLM startup Kwanele South Africa and Meg Young, PhD and Tamara Kneese, Ph.D. from Data & Society Research Institute, where we evolve a federated scenario-writing approach to designing an evaluation protocol - https://lnkd.in/gHsBeEEn
Kwanele builds a chatbot that provides legal assistance and social support to survivors of gender-based violence in South Africa. Our goal was not to conduct an evaluation but to take a step back and understand key strengths and weaknesses that need to be evaluated, centering the lived experience of Kwanele’s key stakeholders and community. The outcomes are a series of evaluation objectives and scenarios that contribute to an evaluation protocol with the core tenet that when we design for the most vulnerable, we create better futures for everyone.
🛑 Existing work on evaluations (in the summary diagrams below) provides an overview of what needs to be evaluated, how and where in the AI system (see https://lnkd.in/gKNKk6Ez). Scholars have argued for public accountability (see https://lnkd.in/g8AnsE9k) and structured access for third-party evaluations (see https://lnkd.in/gdxivtif)
🛑 In our work, we question existing power structures in the context of who gets to define what goes into an evaluation protocol. This has implications for essential questions about the validity and reliability of the outcomes of an evaluation process.
In conversation with recent research, we argue for the need to examine evaluation protocols and methodologies through axes of empowerment and reflexive assessment. See:
- Meg Young, PhD et al’s https://lnkd.in/gpPfTV9E - for an overview of the relationship between power structures and participatory methods
- Mona Sloane, Ph.D. -https://lnkd.in/g_HB2iEd Controversies, contradiction, and “participation” in AI
- Fernando Delgado, PhD et al https://lnkd.in/gSVF96Gg - The Participatory Turn in AI Design: Theoretical Foundations and the Current State of Practice
💜 Kwanele South Africa are doing fascinating work at the forefront of #generativeAI innovation! Thank you Leonora Tima and Ronel Monet Koekemoer for your partnership and this opportunity to contribute to your work. Huge thank you Hanlin Li, Briana Vecchione, Ph.D., and Renee Shelby, PhD!
