Beyond Algorithms:
Machine Unlearning from a Multidisciplinary Perspective

July 28–29, 2026 Vector Institute, Toronto, Canada
Attend View Schedule

About

Machine unlearning—the removal of training data from AI models—plays a key role in making AI systems trustworthy by supporting privacy rights such as the right to be forgotten. As the field enters the era of large generative models, unlearning is also being leveraged to mitigate bias, protect copyright, and correct harmful behaviors. However, this rapid expansion of applications has created a gap between the research community and stakeholders of unlearning, including regulators, content creators, and the broader public.

This two-day workshop brings together participants from diverse disciplines to address questions such as: (a) what are the practical incentives for unlearning, and (b) what are the best practices for performing and evaluating unlearning across different incentive structures. Particular attention will be given to challenges introduced by the widespread adoption of generative AI, including issues such as model collapse and the need for recursive forms of unlearning.

Attend

Venue: Vector Institute, 11th Floor, 108 College St W1140, Toronto, Ontario, Canada M5G 0C6

Dates: July 28–29, 2026

Schedule

Day 1 — July 28, 2026
Time Event
08:00 – 08:45 Registration and Breakfast
08:45 – 09:00 Welcome Remarks
09:00 – 09:30 Introduction Talk: What is Machine Unlearning?
09:30 – 12:30
Session 1 Regulatory and Societal Motivation for Unlearning
09:30 – 10:30 Discussion in Breakout Groups
Three parallel groups (same topic)
Discussion Topics
  1. Who are the key stakeholders of machine unlearning, and what are their primary objectives? Is unlearning the most appropriate framework for addressing their concerns?
  2. What are the core motivations for unlearning (e.g., privacy, copyright, safety)? Are there additional motivations beyond these three? For each motivation, who are the principal actors involved, and what are their respective interests?
10:30 – 10:45 Short Talk: Privacy and AI (tentative)
10:45 – 11:00 Short Talk: Copyright and AI (tentative)
11:00 – 11:30 Coffee Break
11:30 – 12:30 Discussion in Breakout Groups
Three parallel groups (same topic)
Discussion Topics
  1. Which motivations for unlearning are/could be mandated by regulation, and which fall outside the scope of regulatory frameworks and why?
  2. If we categorize unlearning applications based on Question 1, how do the resulting categories differ in scope and requirements?
  3. From a technical standpoint, what can current unlearning methods achieve, and what remains as open challenges?
12:30 – 13:30 Lunch
13:30 – 16:30
Session 2 Definitions of Unlearning and Their Alignment with Policy
13:30 – 13:45 Talk: TBD
13:45 – 14:00 Talk: TBD
14:00 – 15:00 Discussion in Breakout Groups
Three parallel groups: Privacy Copyright Safety
Shared Discussion Questions
  1. For this motivation (i.e., privacy/copyright/safety), are there well-defined goals and threat models for unlearning? If not, how should they be defined? What are the requirements and expected effects of unlearning in this context? What are the ideal forms of unlearning (e.g., when is exact vs. inexact unlearning appropriate)?
  2. Do existing definitions and methodologies for unlearning align well with this motivation and the corresponding regulation (if there exists)?
  3. What is ambiguous in the translation from motivations to formal definitions, and from definitions to practical methodologies?
  4. Unlearning is typically defined as a single-round operation. What considerations arise when multiple rounds of unlearning are required? What could go wrong in such an iterative process?
  5. Beyond "removing a data point's influence from a trained model," what other factors should be reflected in the definition of unlearning? For instance, the distribution of unlearning requests may be non-uniform—should definitions be tailored based on such prior knowledge?
15:00 – 15:30 Coffee Break
15:30 – 16:30 Discussion in Breakout Groups
Three parallel groups: Privacy Copyright Safety

Same questions as the first round—participants are encouraged to join a different group.

Discussion Questions
  1. For this motivation (i.e., privacy/copyright/safety), are there well-defined goals and threat models for unlearning? If not, how should they be defined? What are the requirements and expected effects of unlearning in this context? What are the ideal forms of unlearning (e.g., when is exact vs. inexact unlearning appropriate)?
  2. Do existing definitions and methodologies for unlearning align well with this motivation and the corresponding regulation (if there exists)?
  3. What is ambiguous in the translation from motivations to formal definitions, and from definitions to practical methodologies?
  4. Unlearning is typically defined as a single-round operation. What considerations arise when multiple rounds of unlearning are required? What could go wrong in such an iterative process?
  5. Beyond "removing a data point's influence from a trained model," what other factors should be reflected in the definition of unlearning? For instance, the distribution of unlearning requests may be non-uniform—should definitions be tailored based on such prior knowledge?
17:30 Dinner
Day 2 — July 29, 2026
Time Event
08:00 – 09:00 Breakfast
09:00 – 12:00
Session 3 Practicing and Enforcing Unlearning in the Real World
09:00 – 09:30 Talk: TBD
09:30 – 10:30 Discussion in Breakout Groups
Three parallel groups (same topic)
Discussion Topics
  1. How should unlearning be audited, evaluated, and/or enforced for each motivation? What are the requirements and expected effects, and how could they be captured by quantitative metrics?
  2. Who are the principals responsible for enforcing unlearning? Depending on this, what practical assumptions can be made when computing these metrics? It is worth noting that the ultimate beneficiaries of unlearning may not be the same parties as those enforcing unlearning. Is this problematic?
10:30 – 11:00 Coffee Break
11:00 – 12:00 Discussion in Breakout Groups
Three parallel groups (same topic)
Discussion Topics
  1. What would a complete, end-to-end unlearning framework look like? For each motivation, how do we determine when unlearning is needed?
  2. Then, given a set of data points to be unlearned, could some be fundamentally "un-unleanable", i.e., removing them would significantly degrade model utility or even create other risks? What characterizes such points, and how should they be handled?
  3. In practice, unlearning requests can affect multiple models and may impact a single model more than once. For example, if a generative model is trained on outputs from another model whose training set contains a point to be unlearned, should the downstream model also undergo unlearning? How should such cascading dependencies be addressed?
12:00 – 13:30 Lunch
13:30 – 14:30
Concluding Literacy and Policy Recommendations
13:30 – 14:30 Discussion
Discussion Topic
  1. What are the most urgent findings and recommendations from this workshop that need to be communicated to stakeholders and regulators to enable the practice of unlearning?
14:30 – 14:45 Closing Remarks

Note: The above schedule is tentative. Each session will primarily feature discussions in breakout groups, supplemented by short talks. We are still working on specific discussion topics and questions based on the expertise and interests of confirmed attendees.

People

Organizers

Nicolas Papernot
Nicolas Papernot
Nick (Hengrui) Jia
Nick (Hengrui) Jia
Anvith Thudi
Anvith Thudi

Attendees

Contact

For inquiries about the workshop, please email:

Nick (Hengrui) Jia

Sponsor

Schmidt Sciences