January 2026, SimPPL's 2025 Wrapped and Arbiter detects something before YouTube does
2025 Wrapped: ~3.9k site visitors, 20M+ posts/week traced by Arbiter, 100+ families reached by Sakhi, awards from Google, Omidyar, Ford Foundation, and more. Plus: Arbiter surfaced a likely-AI-generated YouTube cluster two months before YouTube took it down.
2025 was a milestone year. Our site reached approximately 3,900 visitors (75% growth year-over-year). We doubled our team size. The newsletter grew to 200+ subscribers. Arbiter's beta now traces 20M+ social media posts per week. Our GenAI health-literacy pilot reached over 100 families via Sakhi in Jalgaon, India, and 200+ families via Maitri in Bangladesh. We shipped a trilingual LLM evaluation benchmark on 1,000 reproductive healthcare questions.
We expanded partnerships with the Spreeha Foundation (Bangladesh), DW Akademie (Germany, Southeast Asia, Kenya), Cohere AI (North America), and Jagran New Media (India). We won awards from Google, Omidyar Network, Ford Foundation, and others, jointly with Harvard and MIT.
Looking ahead to 2026, we are planning Arbiter case studies with journalists on AI-generated discourse in African countries, H-1B visa conversations, and digital governance debates.
Between June and August 2025, Arbiter surfaced a YouTube cluster of Ibrahim Traoré praise content with synthetic-media hallmarks. We didn't tell YouTube. We logged the detection and watched. In August, YouTube took the cluster down on its own, for what it called inauthentic behavior.
This is the kind of case we built Arbiter for. Public researchers should not need access to internal trust-and-safety systems to find this stuff. They need transparency tools. The full case study is in our 2025 Annual Report.
- AI's trillion-dollar opportunity: From Data Records to Context Graphs
Why the next generation of enterprise giants won't just store data but will capture "decision traces" — the tribal knowledge and reasoning behind exceptions that currently vanish into Slack threads and meetings.
- LACUNA: Cross-Market Data Fusion for Prediction Trading
An experimental reinforcement learning agent designed to trade on Polymarket's 15-minute crypto prediction markets by fusing fast and slow market data.
- Scaling Laws for the Training Horizon
Clarifying a common misconception in LLM scaling: while muP handles hyperparameter transfer across model widths, it does not account for the training horizon or batch size.
- The Impact of LLMs on Online News Consumption and Production
How LLMs are reshaping the news industry — analyzing consequences for publishers who block AI crawlers.
- The AI Newsroom: Widespread Integration in American Journalism
AI usage is no longer experimental but has become widespread across American newspapers.
- NATO's Cognitive Warfare (CogWar) Framework
Research into Cognitive Warfare: a new domain of conflict where the human mind is the primary target.
- Show Me The Data: A Guide to Platform Data & Research Access
A technical manual for researchers navigating data warehouses of Very Large Online Platforms under the EU's Digital Services Act.
- Mapping the Online Manipulation Economy
Science article analyzing industrial-scale infrastructure supporting inauthentic activity online — over 60,000 bots tracked.
- The $200k Esports Arbitrage: TeemuTeemuTeemu's Winning Play
A Polymarket bot that turned $900 into $208,500 in profit within a single quarter.