3.7M+
posts collected across investigations
Arbiter
See the accounts shaping your feed before they shape your views.
An AI copilot for public discourse · by SimPPL, a U.S. 501(c)(3) nonprofit
SUPPORTED BY GOOGLE · MOZILLA · WIKIMEDIA · FORD · OMIDYAR · MIT · COLUMBIA · UNDP
arbiter.simppl.org
Trust in what we read online has collapsed
28%
of Americans trust the mass media. The lowest reading in Gallup's five-decade trend (Sept 2025).
58%
of people worldwide are worried about what is real and what is fake online when it comes to news (Reuters Institute, 2025).
54%
of Americans now get news via social media and video networks, ahead of TV for the first time (Reuters Institute, 2025).
AMERICANS WITH A GREAT DEAL OR FAIR AMOUNT OF TRUST IN MASS MEDIA · GALLUP
SRC · news.gallup.com/poll/695762 · Reuters Institute Digital News Report 2025
ARBITER · SIMPPL
Campaigns cross platforms. Moderation stops at each platform's edge.
On Twitter, true stories took about six times as long as false ones to reach 1,500 people (MIT, Science, 2018). By the time a claim is checked, the campaign that planted it has moved on.
Each platform moderates only its own feed. A campaign that runs on X, YouTube, Telegram, and WhatsApp at once shows each moderation team a quarter of the picture, and no one is looking at the whole.
OnPassive, an "AI-powered" tech scheme, ran its promotion on five platforms at once. The SEC charged it with fraudulently raising over $108 million from more than 800,000 investors. Each platform saw only its own slice.
SRC · news.mit.edu (Vosoughi, Roy & Aral 2018) · SEC LR-25809 (2023)
ARBITER · SIMPPL
Arbiter reads public conversation across platforms
01 · COLLECT
Public data, collected within platform rules
We help collect public social data from sources within platform terms of service, including platform-granted APIs, across X, YouTube, Reddit, Bluesky, Instagram, Facebook, TikTok, LinkedIn, Telegram, and 4chan, plus 20,000+ newspapers and Wikidata.
02 · ORGANIZE
Themes, actors, networks
Posts cluster into themes and sub-themes. Each theme surfaces the accounts moving it, with per-actor dossiers and influence maps.
03 · INVESTIGATE
Agentic investigations
Deep research agents answer your questions across the corpus and cite every post they draw on. You keep the judgment.
248B
interactions on content from the accounts we monitor
150+
organizations with journalists signed up
12
countries with newsroom and civil-society partners
SRC · arbiter.simppl.org live counters (May 2026) · simppl.org/about
ARBITER · SIMPPL
Platforms acted on what Arbiter surfaced
X 2023
Our 2023 reporting prompted an internal investigation at X into pro-Russian bot networks.
Investigation opened
Meta Q1 2024
We shared findings with Meta's team; their public Q1 2024 Adversarial Threat Report details the takedown of a Bangladesh network: 50 accounts and 98 Pages with 3.4M followers.
Network removed
YouTube 2025
Channels we flagged for AI-generated propaganda were terminated by YouTube independently. We had filed no report; enforcement confirmed the detection.
Channels terminated
SRC · transparency.meta.com/sr/Q1-2024-Adversarial-threat-report
ARBITER · SIMPPL
Investigations start from a live case-study library
Pick a topic or a set of accounts, set a date window, and Arbiter builds the corpus. 31 public case studies are open to read today, from the Online Safety Act to Air India 171.
XYouTubeRedditBlueskyTelegramInstagram
Every study keeps its sources attached: each number traces back to the posts behind it.
SRC · arbiter.simppl.org/studies · product screenshot
ARBITER · SIMPPL
Themes surface the actors moving them
Posts cluster into themes, themes into sub-themes, and each cluster names the accounts driving it. The graph turns a feed into a map: who amplifies whom, in which language, on which platform.
Reports built this way sit behind both platform outcomes on the track-record slide: X's internal investigation and Meta's Bangladesh takedown.
SRC · product screenshot
ARBITER · SIMPPL
Ask the agent a real question
We asked this on the Traoré study, live, on July 8. The agent read the 191 posts, ranked the accounts by engagement, and checked where each one sends its audience.
The answer came back with a sortable table and a finding we could act on: the top posts carry no outbound links at all. This engagement lives in native video, so link-tracing tools would miss it.
Every row ties back to the posts behind it. A question like this costs 5 credits and takes under a minute.
traoré study · 191 posts · 3 platformsMAY 26 – JUN 2
Which accounts drive the most engagement on Ibrahim Traore content in this window, and where do their outbound links point?
ANALYZING PATTERNS…
Engagement is driven by a small set of accounts, mostly via high-performing posts that don't include outbound links: native video and text rather than link-sharing.
| Account | Posts | Interactions | Outbound links |
|---|---|---|---|
| zoomafrika1 | 1 | 56,860 | none detected |
| SahelAlerte | 1 | 29,721 | none detected |
| WelcomeTheGulag | 1 | 23,175 | none detected |
| kelevitch | 1 | 22,426 | none detected |
| WithoutHistory | 1 | 10,499 | none detected |
| kwadwosheldon | 1 | 8,725 | youtube.com |
"Driver" status here is concentrated in single spikes rather than sustained volume.
Which accounts are amplifying this narrative the mostHow did this discussion spike and evolve over time
SRC · live agent answer, Jul 8 2026 · question, table, and takeaway verbatim, prose condensed (research/09)
ARBITER · SIMPPL
Under the hood: three problems we solve through rigorous science
01 · SEARCH
Find the right posts among millions
Platforms do not share data with each other, so cross-platform patterns stay invisible. Retrieval has to rebuild the picture.
02 · ORGANIZE
Theme labels collapse into duplicates
Ask an LLM to label clusters one at a time and it repeats itself. The failure is documented across the field (Ziems et al. 2024).
03 · INVESTIGATE
Agent answers arrive ungraded
An agent that picks tools and writes code can be wrong quietly. We grade every answer against hand-written baselines.
Most listening tools answer with RAG over whatever their feeds already collected. Arbiter plans the collection itself and surfaces conversations you did not know to ask for. The next slides show the method; the appendix lists every model.
SRC · SimPPL methods talk, NYU CSMAP (Mar 2026) · aclanthology.org/2024.cl-1.8
ARBITER · SIMPPL
Search: a question becomes a typed retrieval plan
01Plan
Typed facets from structured sources
The Context Expansion Engine breaks a question into actors, organizations, places, and phrases. It draws on a daily-refreshed knowledge graph: Wikipedia, GDELT, national and regional news.
02Retrieve
Two searches, fused
Keyword search catches exact strings like #BurkinaFaso that embeddings miss. Semantic search over 1024-dim embeddings catches paraphrases with no keyword overlap. Reciprocal rank fusion merges the two rankings.
03Rerank
One calibrated threshold
A neural reranker scores each candidate against the original question, 0 to 1. Scores stay comparable across topics, so a single threshold drops 60 to 80% of candidates.
04Measure
Graded against hand-labeled posts
We grade the whole funnel against posts we labeled by hand. On that benchmark, reranking lifted the share of relevant results in the top 200 from 81% to 99%.
>
Climate change narratives and glacier melting discourse
ORGANIZATIONS · Danish Met Institute · Austrian Alpine Club · Columbia Climate School
LOCATIONS · Antarctica · Greenland · Pyrenees · Andes
DOMAINS · sepp.org (skeptic) · wattsupwiththat.com (skeptic) · nature.com
PLATFORMS · X 68 · YouTube 46 · Bluesky 16
LOCATIONS · Antarctica · Greenland · Pyrenees · Andes
DOMAINS · sepp.org (skeptic) · wattsupwiththat.com (skeptic) · nature.com
PLATFORMS · X 68 · YouTube 46 · Bluesky 16
The public walkthrough at arbiter.simppl.org/features: one question became 130 posts on three platforms, including skeptic domains the user never typed.
10,231collected
→
138relevant · white-monkey study
"White monkey" is coded slang for foreigners hired as props by Chinese firms. The wording moves too fast for keyword lists, so the engine reads the intent of the question and retrieves posts that share none of its words. Expansion casts wide; the whole funnel, plan through rerank, keeps what answers.
SRC · RRF: Cormack, Clarke & Buettcher 2009 · a deterministic entity prefilter cuts rerank GPU ~99% · white-monkey study, arbiter.simppl.org
ARBITER · SIMPPL
Arbiter maps the conversation beyond your search terms
A keyword tool returns matches. We asked one question about Ibrahim Traoré, collected 22,719 posts, and kept the 19,532 relevant ones. Arbiter organizes them into a map you can walk, concept to sub-concept to post, including conversations the query never mentioned. click any row below
Concepts 9 of 51 shown
Governance Challenges in West Africa1253
Cohesive National Development883
Sovereignty and Cultural Resilience395
Strategic Governance and Economic Transformation370
Political Discourse and Rights310
Political Authority and Resistance Dynamics264
Cultural and Political Empowerment258
International Relations and Reconciliation60
Judicial Inquiry Narratives24
Sub-concepts 7 of 41
National Unity and Progress719
Bobo-Dioulasso Event Highlights68
Historical Prison Visit Impact58
Western Criticism Response12
Unfiltered Interview Insights12
Leadership and Patriotism8 ▾
April Ministerial Resolutions6
Posts 3 of 8 · Leadership and Patriotism
TrueMamle YOUTUBE
Captain Ibrahim Traore Speech 1st Of April : Revolutionary Spirit Of Burkina Faso (English)
504 interactions
Bridge24 YOUTUBE
Interview of Captain Ibrahim Traoré | President of Burkina-Faso | (Part1)
9 interactions
FLPLOUZANE TV YOUTUBE
News : President for life' Inside Traoré's Burkina Faso (UK, 13 april of 2026)
2 interactions
The YouTube tab of this study holds 51 concepts, 41 sub-concepts, and 3,817 posts from 1,115 users; this slide walks one path. The rest open in the live study.
SRC · Ibrahim Traoré study, Apr 1–30 2026 window, YouTube concept web · captured Jul 2026
ARBITER · SIMPPL
Organize: geometry stops theme labels from repeating
Say a search comes back with 100,000 posts. Point an LLM at them naively and the bill explodes while the labels collapse into duplicates: 28 of 53 clusters in one corpus came back with the same label.
We organize before any LLM reads the corpus. UMAP and HDBSCAN group posts by meaning, the combination that scored highest for cluster quality across the five clustering methods we tested.
The LLM then labels whole clusters, so cost scales with a few dozen clusters instead of a hundred thousand posts. It proposes several labels per cluster, and any that lands too close to an existing label is thrown out.
The LLM provides the vocabulary. The geometry enforces distinctness.
NAIVE Climate Change Discussion · Narratives · Debate · Discourse (all redundant)
OURS Science Explainers · Extreme Impacts · Climate Policy · Indigenous Livelihoods
OURS Science Explainers · Extreme Impacts · Climate Policy · Indigenous Livelihoods
Average pairwise label similarity, lower is better. We benchmarked eight strategies; ours keeps over 98% of labels unique.
The same embedding space powers entity extraction: nine entity types, resolved to Wikidata, name variants merged automatically. Each actor gets sentiment, emotion, and a clickbait score on a five-point scale.
SRC · SimPPL labeling benchmark (2026), silhouette 0.71 vs 0.55 K-Means · Ziems et al. 2024 · UMAP: arxiv.org/abs/1802.03426
ARBITER · SIMPPL
Deception signals Arbiter reads in every corpus
Sentiment uniformity
Discourse that agrees too much
The Traoré PR campaign read 100% positive, 100% joy across coordinated pages. Organic discourse rarely does.
Clickbait score
Bait at 5, camouflage at 1
The top jailbreak seller scores 4.2. The Traoré operation scores 1, polished to pass as organic praise.
Amplification anomaly
Reach with no audience trail
One crypto casino drew 36.9M interactions from four identical posts; one Traoré post drew 88.8K alone.
Entity mention share
Who benefits from the story
1win led the betting corpus at 90 mentions while mirror domains kept the banned brand alive.
Impersonation
Borrowed legitimacy
Entity resolution surfaced one page posing as 16 real manufacturers, Indofood to Yamaha, recruiting into scam compounds.
Sponsorship disclosure
Paid voice posing as organic
Influencer dossiers separate disclosed from undisclosed promotion and flag SEBI-violation-shaped patterns.
Arbiter sits as an analytical layer over social platforms and online sources. Signals compose into tactics: manufactured consensus, coordinated amplification, borrowed legitimacy, undisclosed promotion. We audit the classifiers in non-Latin scripts, Hindi through Cyrillic, so a campaign cannot hide behind an alphabet.
SRC · public Arbiter studies + partner case reports · arbiter.simppl.org/studies
ARBITER · SIMPPL
Investigate: every agent answer gets graded
The agent picks from an analysis toolkit of 30+ functions across six families: actor profiling, topics, themes, claims, retrieval, and visualization. It answers by writing code against the corpus in a sandbox.
Each answer is a tool-calling loop capped at eight steps, and the executor writes and runs TypeScript in a sandbox. An answer arrives as an artifact you can verify: the code it ran and the posts it cites.
We graded 540 agent runs by hand across nine case studies. An LLM judge scores the code the agent writes against implementations we built ourselves, on everything from tool choice to error recovery.
That grading runs continuously in production. Flagged errors become the input to prompt-optimization runs (GEPA, appendix table), with no person inside each loop. Our engineers review the surfaced error patterns at the system level; each fix propagates to every future query.
Top failure codes across 540 graded runs; the most common defect is answering in prose when code was needed. Partners see what the agent gets wrong before they rely on it.
SRC · SimPPL agent evaluation framework · LLM-as-judge: Zheng et al. 2023, arxiv.org/abs/2306.05685
ARBITER · SIMPPL
Accessing social data isn't the hardest part
Getting the posts is the straightforward part. The layer that turns them into evidence is what we build, and it carries five problems at once.
Vendor sampling
Every feed is a sample, and the vendor's sampling choices ride into every count computed downstream.
Continuous analysis
Pipelines have to keep running over new posts and their images, video, and audio, long after the first export.
Cross-platform patterns
Campaigns move across platforms, languages, and countries. The analysis follows them or misses the pattern.
Evidence of harm
Harm gets labeled, flagged, and documented to a standard platforms and regulators will act on.
Near real time
All of it lands faster than the harm reaches a larger audience, or the finding arrives too late to matter.
SRC · SimPPL data-pipeline practice · arbiter.simppl.org
ARBITER · SIMPPL
Case file 01
Burkina Faso · AI-generated PR campaign
Jun 30 – Aug 30 2025 · X · YouTube · Reddit
01Detect
A cluster of accounts pushes a flawless president
Coordinated pages promote Ibrahim Traoré with sentiment reading 100% positive, 100% joy. Organic discourse rarely reads this way.
02Measure
Engagement without an audience trail
One account lands 88.8K interactions from a single post (views, comments, and likes combined). Its clickbait score is a clean 1 of 5. The campaign is built to look like ordinary fan praise, and to a casual reader it does. Arbiter flags it anyway: organic audiences disagree with each other, and this one never does.
03Identify
The videos are AI-generated slop
Fake factories, fake railways, fake tributes, drawing millions of views before any platform labels them.
Uniformly green panels are the anomaly that starts the investigation.
SRC · public Arbiter study panels · arbiter.simppl.org/studies
ARBITER · SIMPPL
Case file 01
The campaign's claims, checked
Verdicts · enforcement · 2026 follow-up
The claim: new factories, railways, and mega-projects across Burkina Faso. The flagship videos celebrating them are synthetic.
AI-generated
The claim: an organic pan-African support wave. Uniform 100%-positive sentiment and single-post interaction spikes match coordination patterns, and platforms later labeled videos as altered.
Coordination pattern
The claim: harmless fan content. YouTube terminated the flagged channels on its own; we had filed no report.
Enforced against
Early detection · validated by YouTube
One year on (May 26 to Jun 2, 2026), a rebuilt network of 30 channels cut from one naming template works the same beat, now with tip jars, memberships, and merch links attached. One of the 2025 actors is back under the same name.
"This video is no longer available because the YouTube account associated with this video has been terminated."
SRC · study panels + subsequent platform enforcement
ARBITER · SIMPPL
Case file 02
Fake overseas jobs, Indonesia and the Philippines
Dec 2025 – Apr 2026 · Facebook · Migrasia
Two studies built with Migrasia read 2,251 recruitment posts from 1,444 accounts. Most scam posts work hard to look legal: 489 Indonesian posts open with words like "resmi" (official) and "no upfront fee", and 183 Filipino posts print a recruitment licence number.
The contact details give the networks away: one shortened link posted by 21 regional job pages under different names; six profiles that all route to a single WhatsApp number; in the Philippines, 34 shared phone numbers and emails tie the pages into 14 networks.
One agency kept posting vacancies under licence DMW-011-LB-051822-R after that licence had expired, so the licence number itself was part of the sales pitch.
14 recruiting networks mapped
Recruitment tactics by post count, Indonesian corpus (1,492 posts); bars scaled to the leader.
A quarter of the Filipino corpus is the aftermath: airport interdictions, arrests for illegal recruitment, repatriations of trafficking victims.
SRC · two public Arbiter × Migrasia reports, linked on the partners slide · arbiter.simppl.org/studies
ARBITER · SIMPPL
Case file 03
Manosphere narratives, packaged for young audiences
Jan 1 – Feb 28 2026 · X · 3,361 posts
Two months of high-engagement posts drew 127.3M interactions. The content is sold as self-improvement: "looksmaxxing" routines and testosterone regimens, with verified health information pushed aside.
Engagement runs on a star system: the top 3 accounts take 14.6% of all engagement, while the highest-volume cluster (1.8K posts) registers 0.0%. Culture-war framing carries 57% of interactions; belonging and dating discourse carries 56.8%.
The report maps outbound audience redirection: the links actively shared to move audiences off-platform, and the content driving them there.
3 accounts · 14.6% of engagement
Share of total interactions carried by each theme (themes overlap, so shares exceed 100% combined).
Inside the conversation, a few accounts like @DrewPavlou and @MikhailaFuller carry it, each turning one or two posts into millions of interactions. The recurring claims run from testosterone decline to dating advice, wrapped in culture-war and media-accountability framing; Netflix's documentary was one entry point.
SRC · Understanding the Manosphere report (Arbiter, Mar 2026) · arbiter.simppl.org/studies
ARBITER · SIMPPL
Case file 04
Maryland's mail-in ballots: record vs feed
May 1 – 19 2026 · X · YouTube · Facebook · Bluesky
What officially happened
A vendor printing error sent some voters the wrong party's primary ballot. Unable to identify who was affected, the state re-sent replacements to everyone mailed a ballot before May 14, covering 500,000+ requests. Originals were voided; every return envelope carries a unique identifier, so officials state there is no risk of duplicate voting.
The official explanation, cited in the study across X, Facebook, and Bluesky.
What the feed made of it
The claim: 500,000 fake or illegal ballots revealed. False per the state's election administrator and independent fact-checkers.
Rated false
The claim: a million ballots now circulating. Originals plus replacements, recast as duplicates; the study's claims review found the leap unsubstantiated.
Escalation
One templated line, 50+ accounts. The same wording burst across platforms within hours, overwhelmingly from accounts that are commentary rather than news outlets.
Template bursts
The stakes were legislative: the study's top post (1.32M interactions) demanded the Senate attach the SAVE America Act, which would require in-person documentary proof of citizenship to register and would replace automatic mail-ballot delivery with application-only mail voting. The false ballot claims circulated as the Senate weighed it; the act passed the House 218–213 and failed in the Senate.
SRC · 1,485-post study, saved claims + coordination runs · elections.maryland.gov · FactCheck.org · AP · congress.gov (H.R. 7296)
ARBITER · SIMPPL
Where Arbiter fits in your workflow
01 · SCOPE
You bring the question
A tip, a beat, a policy window. We scope the corpus, platforms, and dates together.
02 · RUN
The pipeline does the sweep
Retrieval, themes, and actor dossiers land in a case-study library. Your analysts question it through the agent, with a citation on every answer.
03 · PUBLISH
You keep the evidence
The findings are yours to publish. Evidence stays attached, so editors and reviewers can audit every number.
Monitoring and OSINT desks
One scoped corpus replaces manual sweeps across four platform search bars.
Fact-checking desks
Claim clustering scales one verified check to thousands of grouped posts.
Trust & safety teams
Coordinated campaigns surface before they reach your abuse reports.
Narrative-intelligence products
Our scored corpora and dossiers slot in as an analytical layer under your platform.
Model stages are commodity parts; the retrieval planning, the thresholds, and the grading harness are ours. When a better model appears, we swap the part and your workflow does not change.
SRC · pilot scoping: team@simppl.org · platform: arbiter.simppl.org
ARBITER · SIMPPL
Newsrooms, civil society, and agencies in twelve countries
Newsrooms
RapplerPH
Jagran New MediaIN
Gothamist · NYPRUS
FactchequeadoUS · LATAM
Africa UncensoredKE
Agencies & platforms
UNDP AI Trust & SafetyUN
UN peacekeeping briefingsUN
U.S. state elections officialsUS
State attorney general officesUS
National regulatory agenciesINTL
Journalists from 150+ organizations in 12 countries have signed up, spanning newsrooms, civil society, and policymakers. YouTube grants us 1.2M API calls a day for data access. The harms repeat across borders, so investigations built in one country become templates for the next. Inbound users include existing Dataminr and Meltwater customers asking to switch. Stanford picked Arbiter for this year's platform workshop with trust and safety leaders; Meta, YouTube, and TikTok held that slot in past years.
SRC · arbiter.simppl.org partner strip · simppl.org/about
ARBITER · SIMPPL
Built by researchers from Twitter, Adobe, and Slack
Swapneel Mehta
Founder & President
Former postdoc in AI safety at MIT and Boston University. Board member at Integrity Institute. Built AI systems at Twitter, Adobe, and Slack. Google Research Innovator.
Dhara Mungra
Co-founder · Engineering lead
Led audience-analytics ML as Data Scientist III at Bombora. M.S. in Data Science at NYU. Fellow with the Center for AI and Digital Policy.
Six full-time engineers
Utkarsh Verma · Atmik Shetty · Dev Bhut and team
System design and infrastructure, NLP and deep research agents, platform backend and APIs.
Why trust us
Credibility
14 peer-reviewed papers (NeurIPS, ICML, AAAI, ICWSM). Awards from Google, Mozilla, Wikimedia, Ford, and Omidyar. Bootstrapped as a nonprofit since 2021.
SRC · simppl.org/team · simppl.org/research
ARBITER · SIMPPL
What Arbiter does, and what stays human
The agent
Collects public social data within platform terms of service, clusters themes, maps actors and networks, and answers questions with a citation for every claim.
The human
Draws the conclusions. Journalists, researchers, and trust & safety teams keep the judgment; Arbiter keeps the evidence attached.
Scope
We fact-source: the system surfaces claims and groups related ones, so one verified fact-check scales to thousands of posts. We spent a year building this with Wikimedia.
The case files in this deck are public studies from 2025 and 2026 windows. A study scoped to your issue tracks your actors, in your languages, on your timeline.
SRC · arbiter.simppl.org/studies · methods grounded in peer-reviewed research
ARBITER · SIMPPL
See it on your issue.
Request a demo: write to team@simppl.org
and we follow up within the week.
A pilot is one case study scoped to your issue: we set the date window together, run the investigation with you, and you keep the evidence and the sources. Our grants subsidize free access for independent journalists and fact-checkers, and paid subscriptions with larger newsrooms and civil society clients generate growing revenue each year. Free investigation reports land in your inbox via simppl.org/newsletter. The platform is live at arbiter.simppl.org.
SIMPPL · U.S. 501(C)(3) · BOOTSTRAPPED SINCE 2021
ARBITER · SIMPPL
Appendix: models in the pipeline, and why
Query planning
GPT-5-nano
Constrained decoding for typed facet decomposition
Post embeddings
Voyage 3.5-lite (1024-dim)
Cost and quality tradeoff at corpus scale
Neural reranking
Voyage rerank-2.5
Absolute relevance scores that hold across topics
Entity extraction
gpt-oss-120b via Groq
Fast open-weights inference; 20b fallback
Theme labels
GPT-4o-mini
Best quality at lowest cost in our labeling benchmarks (GPU theme service)
Entity narratives
GPT-5-nano
Lightweight generation with a rule-based fallback
Agent task LLM
GPT-5.2
Tool selection and code generation
Agent judge
Model-agnostic (OpenRouter)
Six-dimension rubric; Claude Sonnet 4.6 judged the 540-run campaign
Prompt optimization
GEPA + Claude Sonnet 4.6
Prompt evolution via natural-language reflection (arxiv.org/abs/2507.19457)
APPENDIX · MODELS AS OF JULY 2026 · PARTS SWAP AS BENCHMARKS MOVE
ARBITER · SIMPPL