# IGF 2026 WS #374 The UnTraining Dilemma: "Right to be Forgotten" in AI Models

## Description

Primary Focus
The session centers on a core technological paradox: Global privacy laws require a "Right to be Forgotten," but GenAI models currently lack a built-in way to delete specific ingested data.
Short Description & Key Topics Covered
The event explains "Machine Unlearning"—the technical challenge of making AI models selectively "forget" personal data, copyrighted code, or sensitive information without having to retrain multi-billion-parameter models from scratch.  We will address three core topics during this session:
1. Legal vs. Neural Reality: Explaining the "Baked Cake" analogy—why deleting data from a standard database is simple, but extracting data from neural network weights is mathematically complex.
2. The Machine Unlearning Solution: Clarifying emerging technical methods for targeted data removal and demonstrating why full model retraining violates sustainable "Green AI" practices.
3. Shared Governance & Standards: Linking policymakers and AI developers to establish proper compliance standards, verifiable "Proof of Unlearning," data provenance, and shared accountability.

What Sets This Session Apart
Unlike traditional AI governance panels that focus purely on the legal philosophy of privacy, this session confronts the urgent, applied technical reality of enforcement. While previous IGF discussions have debated whether data should be deleted, we address the mathematical roadblock of actually doing it.  Furthermore, we introduce a critical dimension to the privacy debate: The Environmental Cost of Compliance. We will highlight how forcing full model retraining to meet legal deletion requests actively violates green energy targets, establishing Machine Unlearning not just as a privacy workaround, but as a mandatory foundation for Sustainable AI.

## Policy Question(s)

Harmonization & Legal Enforcement: How can global data protection regimes reconcile statutory mandates for individual erasure (e.g., the "Right to be Forgotten") with Generative AI architectures, where data deletion cannot mirror traditional database record removal?

Standards & Technical Verification: What standards, technical criteria, and governance frameworks are needed to establish verifiable "Proof of Unlearning," ensuring regulatory compliance without placing unreasonable audit burdens on AI developers?

Sustainability & Policy Alignment: How can privacy enforcement balance strict legal compliance with global sustainability commitments, considering the substantial carbon footprint and energy demands associated with full model retraining?

## Expected Outcomes

Expected Outcomes & Outputs
1. Tangible Outputs & Deliverables

**Interdisciplinary Policy Brief:** An open-access synthesis mapping machine unlearning benchmarks to data protection requirements, published via NIXI.

**Unlearning Verification Framework:** A draft blueprint defining metrics for "Verifiable Proof of Erasure" to inform standards bodies (IEEE, ISO/IEC, ITU-T).

**Sustainable AI Policy Note:** Recommendations submitted to IGF Dynamic Coalitions on balancing algorithmic enforcement with environmental sustainability.

2. Participant Takeaways

Algorithmic & Regulatory Literacy: Understanding the "Baked Cake" paradigm—contrasting neural network parameter weights with standard database deletion.

**Multi-Stakeholder Alignment:** Regulators explore computational friction and carbon intensity, while ML practitioners internalize privacy rights frameworks.

**Operational Compliance Matrix:** A practical heuristic tool for decision-makers to triage deletion requests, evaluating unlearning, full model retraining, and alternative compliance paths.

## Reference Document

Reference Document

## Format

Roundtable
**Duration (minutes):** 90
**Format description:** 1. Multi-Stakeholder Dialogue: A roundtable layout dismantles the traditional speaker-versus-audience divide, establishing an inclusive, collaborative space. By bridging complex machine learning concepts with high-level legal policy, this setup fosters active peer-to-peer dialogue, transforming attendees from passive listeners into true co-contributors during interactive segments.

2. Depth for a Complex Paradox: Unpacking the intersection of generative AI, the "Right to be Forgotten," and green energy constraints demands dedicated time. A 90-minute duration provides the necessary runway to introduce the technical "Baked Cake" paradigm accessibly, while reserving generous space for deep thematic analysis across our diverse panel.

3. Seamless Hybrid Q&A Execution: The open room layout allows the Onsite Moderator to effortlessly navigate microphoned audience contributions. This setup powers our 35-minute alternating Q&A framework, ensuring natural engagement for physical attendees while keeping the dedicated Online Moderator fully synchronized with virtual chat inputs and Slido polling.

## Hybrid Format

**Hybrid Format:** Interaction & Dynamic Moderation
We will deploy an Onsite Moderator and a dedicated Online Moderator to maintain a strict 1:1 alternating Q&A queue. The Online Moderator will read virtual questions into the microphone to ensure full transcript integration. Remote and in-person speakers will receive equal floor time.

Inclusive Design & Parity
To prevent digital fatigue, presentation decks will use high-contrast, low-bandwidth graphics illustrating core technical concepts. Speakers will address the camera, pacing interventions to accommodate live interpretation. A one-page concept PDF will be distributed via chat and QR code at minute zero.

Tools & Platforms
We will use Slido as a unified platform for both audiences. A 10-second opening poll will gauge the room's expertise to calibrate technical depth, while unified Q&A upvoting will highlight the most pressing audience queries.

## Organizers

**Organizer 1:** Yash Vashishth, National Internet Exchange of India (NIXI)
**Organizer 2:** Mudassir Inam, National Internet Exchange of India
**Organizer 3:** Nikhil Raj, National Internet Exchange of India
**Organizer 4:** Abhishek Anand, NIXI
**Organizer 5:** Sandeep Singh,  National Internet Exchange of India
**Organizer 6:** Sameer Gahlot, National Internet Exchange of India

## Speakers

**Speaker 1:** Luca Belli, Technical Community, Latin American and Caribbean Group (GRULAC)
**Speaker 2:** Nicolo Zingales, Civil Society, Latin American and Caribbean Group (GRULAC)
**Speaker 3:** Rojszczak Marcin, Technical Community, Eastern European Group
**Speaker 4:** Rolf H. Weber, Civil Society, Western European and Others Group (WEOG)

## Moderator

Yash Vashishth, Technical Community, Asia-Pacific Group

## Online Moderator

Sandeep Singh, Technical Community, Asia-Pacific Group

## Rapporteur

Abhishek Anand, Intergovernmental Organization, Asia-Pacific Group

## Participant Benefits

What will participants gain from attending this session?

## SDGs

9.4
9.5
9.b
16.10
17.14
17.16
17.17