Briefing Notes – Ethics of AI Society Meeting

General Impressions

• Positive feedback regarding variety of topics and quality of discussions.
• Appreciation for including early-career presenters (students, emerging scholars) to encourage skill development.

Keynote Discussion

• Keynote featuring the MPP was widely praised.
• The MPP provided practical, policy-focused insights that complemented philosophical perspectives.
• Participation arose from prior consultation regarding an AI-related bill.
• Conversation was viewed as balanced, nuanced, and constructive.

Content Themes

• Last symposium was broader (medical AI, governance, opacity, social issues).
• This symposium leaned more heavily into education and AI literacy, reflecting its growing urgency.
• Group members emphasized the value of maintaining a balanced, non-alarmist tone toward AI.

Emerging Concepts

• “34%” suggested as shorthand for individuals strongly opposed to AI (a gentler alternative to “Luddite”).
• Awareness of new methods for AI-assisted cheating (e.g., faked LMS activity) raised concerns.

AI-Assisted Assessment & Cheating

• AI enables students to fake not only writing but also typing behavior.
• Discussion of combining AI-generated personalized exam questions with in-class written responses to ensure authentic understanding.

Grade Inflation

• Worsening grade inflation contributes to unhealthy academic competition.
• Suggestions included:
  • Switching to a points-based system spanning a student's entire degree.
  • More emphasis on engagement, persistence, and learning outcomes rather than pure percentages.
• Observations that Laurier’s letter-grade system reduces grade-chasing pressures.

Motivation & Learning

• Students increasingly focus on grades rather than learning.
• Group suggested reforms to increase:
  • Collaboration
  • Community learning
  • Peer accountability
  • Reduction of individualistic competition

Collaborative Learning Debate

• Some advocated for collective classroom learning models to build communication and problem-solving skills.
• Others warned of:
  • “Classroom communism” (free-riders, uneven contribution)
  • Structural incentives that still reward individual competition
  • Systemic problems (job scarcity, societal pressures) that schools alone cannot solve

University & Inter-Institutional Engagement

• Efforts underway to connect with:
  • University of Waterloo departments
  • Laurier philosophy & AI ethics faculty
  • University of Guelph AI initiatives
  • Conestoga College (new member joined)
• Progress uneven; varying degrees of institutional buy-in.

Possible Name Change

• Consider renaming the group to:
  • Remove “Waterloo” to appear less restricted
  • Adopt terms like Institute instead of Society
• Decision deferred; URL purchase postponed until name is finalized.

Municipal / Political Engagement

• Potential for further collaboration with MPPs or municipal bodies.
• Interest in connecting with emerging national ethics-of-AI organizations.

Public Presence

• Symposium videos to be posted on YouTube.
• Website now active; provides a public-facing home for resources and events.

Core Problem

AI models often appear accurate while still being wrong for hidden reasons—due to opacity, confirmation bias, and flawed assumptions.

Key Concepts

Training vs. Test Data:
Even “successful” test performance may mask underlying conceptual errors.

Confirmation Bias in AI Models:
If a model learns the wrong correlations but appears accurate on test data, users may be misled.

Feedback Loops:
Especially relevant in predictive systems (e.g., predictive policing), where model outputs actively shape the world being measured.

Predictive Policing Case Study

• Models predict crime detection, not crime itself.
• Finite police resources create self-fulfilling prophecies:
  • More police in one neighbourhood → more detected crime → model “validated”
  • Under-policed neighbourhoods generate no data → model never corrected
• Hidden assumptions (e.g., infinite police resources, unbiased detection) rarely hold.

Need for Conceptual Modeling

• The “model in the head” (human assumptions) is as important as the algorithmic model.
• Errors arise when conceptual assumptions are never challenged.

Proposed Solutions

• Independent data collection and external audits to break confirmation loops.
• Multidisciplinary collaboration (philosophers, data scientists, policy experts).
• Development of falsifiability criteria and counterfactual reasoning tools.

Certification Proposals

• AI models as a form of “modern snake oil” needing regulation.
• Possible creation of third-party accreditation bodies akin to:
  • Food safety inspections
  • Energy-efficiency certifications
  • Device standards compliance

Hybrid Governance Model

• A purely governmental or purely market-driven system is insufficient.
• Proposal:
  • Government establishes baseline standards
  • Private organizations compete to certify models
  • Public looks for “trusted AI” labels
• This would create:
  • Profit incentives for improving certification
  • Nonprofit oversight to avoid regulatory capture

Critical Need

• Strong consensus: widespread AI literacy is now a societal necessity.
• Public must understand:
  • False positives/negatives
  • Thresholds
  • Trade-offs in model accuracy
  • Limitations of AI outputs
• Without this, people misinterpret inevitable imperfections as catastrophic failures.

Education Recommendations

• Integrate AI literacy as early as elementary school.
• Teach:
  • Difference between AI and non-AI tools
  • Critical evaluation of AI outputs
  • Ethical considerations
  • Proper use vs. misuse
• Emphasize that completely refusing to use AI impedes literacy.

Membership & Continuity

• Group now includes >100 members across Waterloo, Laurier, Guelph, and Conestoga.
• Concerns about long-term sustainability; may eventually need faculty leadership.

Technology & Infrastructure

• Teams group remains under UW infrastructure; independence would require alternative hosting.
• Website now provides a public-facing platform; symposium videos being published.