Category: CanuckDUCK Brand Forums → Platform Development → AI & Moderation
Opening the Conversation
As we continue developing CanuckDUCK's AI-assisted tools—particularly the Forum Analysis Engine (FAE) that helps community leaders and municipal administrators understand the pulse of local discussions—we want to be transparent about the edge cases we're actively working through.
This isn't a polished policy document. It's an invitation to think alongside us.
What AI Does on CanuckDUCK (Currently)
Before diving into the hard questions, here's what we're building:
- Forum Analysis Engine (FAE): Summarizes discussion threads so community leaders can quickly understand constituent concerns without reading hundreds of posts
- Content flagging assistance: Helps moderators identify potentially problematic content for human review
- Geographic routing: Assists in directing discussions to appropriate community levels
What AI does not do: make final moderation decisions, identify users, or replace human judgment on community standards.
The Edge Cases Keeping Us Up at Night
1. Summarization Bias
When FAE summarizes a 200-post thread for a city councillor, whose voices get amplified? We're wrestling with:
- Do passionate minority viewpoints get flattened into "some users disagreed"?
- How do we represent genuine consensus vs. coordinated campaigns?
- Should summaries flag when discussion participation skews demographically?
2. Context Collapse in Moderation Flagging
Sarcasm. Regional slang. Inside jokes from long-running community discussions. AI systems notoriously struggle with context. A post saying "yeah, that's brilliant city planning" could be genuine praise or biting criticism depending on the thread.
We're exploring: layered context windows that consider thread history, community norms, and user participation patterns—without creating invasive profiles.
3. The "Reasonable Canadian" Problem
Content moderation often defaults to some imagined "average user" standard. But CanuckDUCK serves communities from downtown Toronto to rural Yukon. What reads as heated-but-normal political discourse in one community might genuinely alarm another.
Should AI calibrate to local norms? If so, how do we prevent that from enabling toxic community bubbles?
4. Pseudonymity vs. Pattern Recognition
Our privacy-by-design architecture uses pairwise pseudonymous identity—you're a different "duck" in each community context. But effective AI moderation often relies on behavioural patterns.
How do we give AI enough signal to identify bad actors without undermining the pseudonymity that enables authentic discourse?
5. Transparency of AI Involvement
When a community leader receives an FAE summary, should the summary itself disclose its limitations? When AI flags a post for review, should the poster know? We're leaning toward radical transparency, but the UX implications are real.
6. Canadian Data & Model Training
We're committed to Canadian data sovereignty. But what happens when we use AI models trained predominantly on American discourse patterns? Does that subtly import cultural assumptions about what "civil discourse" looks like?
What We're Not Building
Let's be explicit:
- No AI-generated content masquerading as user posts
- No predictive profiling of users' political leanings
- No automated bans or sanctions without human review
- No selling of interaction data to train external models
Your Turn
We don't have all the answers. That's the point.
- Which of these edge cases concern you most?
- What have we missed?
- Are there AI applications you'd want to see that we haven't considered?
- Where should the human-in-the-loop checkpoints be?
This is civic infrastructure. You should have a voice in how it's built.