-
Notifications
You must be signed in to change notification settings - Fork 58
Feedback from UNGA80
ROOST co-hosted a side event with The Christchurch Call Foundation at the UN General Assembly in 2025 (UNGA80). We facilitated a rose-bud-thorn[^1] exercise to examine and collect feedback about Osprey. To promote the culture of openness and transparency, these are the raw notes from the session and some insights we drew based on the discussion.
The session brought together participants from diverse backgrounds, including civil society organizations, policy teams from tech companies, and policymakers. Most attendees were not software engineers or data analysts, which shaped the nature of the discussion and feedback we received. We noticed a significant knowledge gap around open source software and how it differs fundamentally from OSINT (Open Source Intelligence) tools. This distinction became important as the conversation developed. Osprey is designed as internal platform tooling that helps tech companies monitor and respond to harmful content on their own platforms. The data processed by Osprey stays within each platform and is only accessible to that platform's trust and safety teams. Unlike OSINT tools or signal-sharing systems, Osprey does not expose platform data to external parties.
Some concerns raised during the session were thoughtful and valid for certain types of tools, but may not apply to Osprey specifically given its design as internal infrastructure. Understanding this distinction helps contextualize the feedback below.
Positive highlights or successes
- Simple interface for non-data driven people to use/understand
- User friendly + access to User ID
- keyword/threat signal detection
- Simplicity
- Rules can be created easily
- Clean user interface
- Picking up positive/negative sentiments around key topics
- Easy monitoring/requires not much technical skill/expertise
- Clean simple interface
- Seems quite easy to learn how to use even for non-tech people
- Enhances access to data for small/medium platforms
- SME access
- Open access
- Helps small company
- In-built privacy
- It’s free, sharable and scalable to apply for small to med companies and different types not limited to gov
- Brings together different aspects in TFGBV prevention
- One tool does it all which is great
- Great that it’s open source - could be a really helpful moderation resource for small developers committed to SbD
- Open source / accessibility
- Osprey could bring real time evidence into policy, enforcement conversations
- Seems to have great options for customization of queries and keywords
- Tracking TVEC perpetrators
- More granular
- Time parameters for ongoing monitoring
- Identifying bad actors feasibly
- Effective content moderation at scale
- Moderation @ scale
- Visibility of concurrence TVEC → Reviews high volume of content
- Tracking (very specifically) can be very useful to spot early patterns as a human (so that not only AI is measured in detection)
- Find online links to off-platform TVEC + Tags of known extremists
- Allows platforms to tailor their search to their needs/situation
- Layered signals can be investigated
- Logging, technical but opinionated
- Focused semantics
- As a rules engine, it could be effective for TVEC in theory, to support companies beyond detention
- Allows for real-time identification and analysis that’s accessible
- [redacted] research team sees value for our rhetorical qual/quant analysis
- Increased capabilities for researchers/company insights
- How it might interact with other platforms when layered
Opportunities for growth or future goals
- DIRE could become DIRES for “support”: create space for positive interventions
- Ability to customize investigation timing/query time frames (eg to specific multiple dates)
- If fine-tuned opportunities for broad use cases, requiring little training or specific expertise in many sectors
- Feasibility adaptable
- Potentially useful features
- Would this be applicable to legal but harmful content, should it be?
- Move beyond content signals to ID larger network and behavioral signals for prevention
- Leveraging at scale with automation
- Safeguarding youth and extending to smaller tech companies
- Roll in more platform interfaces
- Draws out anonymised user level data to enable targeted prevention interventions
- Could support small medium (waiting for data)
- Expansion - more tech platforms on board who can explain and promote, great potential for policy makers to evidence to form policy recs
- Audio and image content → can this be added
- Signal sharing
- Opportunities for insight analysis from more diverse/varied perspectives
- Continued feedback loops with the community. Diversity opinions based on expertise
- How to ensure collective action across platforms
- Evidence based safeguards ie what can be systematically enforced
- You could add an auditable tool to publish transparency reports
- Pilot it on platforms of different types! You might end up needing different versions
- Potential to be used across multiple platforms by user type and content flags (lingo, specific words) to track migration cross pollinations
- Opportunity for use by small tech that is committed to safety
- Potential for shared learning and data analysis
- Could this tool also be used for detection done by competent authorities? They don’t look for the accidents but do want to find
Challenges or difficulties faced
- Boring interface
- User interface is a little challenging to navigate for non-tech people
- Too much info?
- History layout could be improved for readability and digestibility (not that readable/easily digestible)
- Requires a lot of human input and decision making to set up
- Not sure if it can capture nuanced contexts, languages, different formats
- English language exclusivity
- Limited to certain platforms
- If not totally open source
- Will SM companies play ball?
- How to raise awareness and adoption among small, medium platforms?
- Pushback from certain tech or platforms not using could mean less comprehensive data
- Single platform scope
- Could be used by nefarious actors
- Misuse of built in safeguard mechanisms
- Dual-use: can probably be used to deeper polarization conflict
- False flags (false positives and/or negatives) → based on the rigidity or limitations of the rules
- Evolving content → constant need to reassess efficacy of program (evolving threat landscape)
- Potential misuse by governments to track human rights defenders
- Abuse and chance for bias
- No checks and balances on how searches are done ie words searched might not be illicit content
- Malign govt interpretations of TVEC → eg trans as nihilistic violent extremism
- Too database query-ish
- Competes with other log viewing tools
- Lots of “osprey”s on github
- How can this meet the needs of most/all platforms? Don’t they already have diversified ideas about the tools they would need and/or already have tools?
- Whether this can properly layer in existing detection methods and cooperation from GIFCT/TAT
[^1]: Rose: positive highlights or successes; Bud: opportunities for growth or future goals; Thorn: challenges or difficulties faced