General tech overview · August 2026

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

Amplifier
SUPPORTED BY GOOGLE · MOZILLA · WIKIMEDIA · FORD · OMIDYAR · MIT · COLUMBIA · UNDP arbiter.simppl.org

Americans’ trust in the media is at its lowest on record

Driven by repeated baseless claims: once people meet enough of them, they start doubting accurate reporting too.

80% 40% 0% 72% 28% 1976 September 2025 Share of Americans with a great deal or fair amount of trust in mass media

Fewer than three in ten Americans now trust mass media to report the news fully, accurately and fairly. The people who most need reliable information trust it least.

SRC · Gallup · Hoes et al., Nature Human Behaviour 2024 ARBITER · SIMPPL

Algorithms magnify fake narratives that reach people’s jobs, health and votes

False stories reach 1,500 people six times faster than true ones (MIT, 2018).

BBC News: Cambodia scams, lured and trapped into slavery in South East Asia
Jobs. Fake job ads lure workers into forced-labour scam compounds. We traced 2,251 such ads into 14 recruiting networks.
Src · BBC News ↗
The Guardian: Samoa measles outbreak, WHO blames anti-vaccine scare as death toll hits 39
Health. Online anti-vaccine fear drove Samoa's vaccination rate down to 31%. The measles outbreak that followed killed dozens of children.
Src · The Guardian ↗
Al Jazeera: Romania's top court annuls results of presidential election's first round
Elections. A coordinated TikTok operation pushed one candidate to a first-round win. Romania's top court annulled the result.
Src · Al Jazeera ↗
ARBITER · SIMPPL

Each platform only sees a fraction of the true picture

Campaigns run everywhere at once. Moderation stops at each platform's edge.

A campaign running on four platforms leaves each one holding roughly a quarter of the evidence. No moderation team sees the rest.

The tool researchers used to look across platforms, CrowdTangle, shut down in 2024. Paid replacements cost thousands a month.

OnPassive recruited investors on every platform at once, so no single platform saw the whole scheme. The SEC charged it with raising $108 million from 800,000 investors.

X YOUTUBE TELEGRAM WHATSAPP ONE CAMPAIGN · FOUR PARTIAL VIEWS
SRC · MIT (Vosoughi et al., 2018) · SEC v. Mufareh / OnPassive, LR-25809 ARBITER · SIMPPL

We catch cross-platform harms faster than any one platform

YouTube removed the channels we had already flagged

We found AI-generated channels running an all-positive PR wave for Burkina Faso’s interim leader. YouTube terminated them later, on its own. We filed no report.

Public Arbiter study

Meta removed 50 accounts and 98 Pages after our work

The network reached 3.4 million followers in Bangladesh. Meta took it down following our reporting on policy-violating hate and harassment there.

Global Voices, 11 July 2024 · Meta Q1 2024 report

GASA nominated us for its Scam Fighter Award

Working with Migrasia, we surfaced 14 networks running job scams on migrant workers in under a day. The Global Anti-Scam Alliance nominated us across its 120+ members in 90 countries.

Global Anti-Scam Alliance · nomination, August 2026

Stanford picked Arbiter to lead its data workshop

We run this year’s workshop at the Trust and Safety Research Conference. Meta, TikTok and YouTube held that slot in past years, presenting their own research APIs.

Stanford Trust and Safety Research Conference 2026
ARBITER · SIMPPL

Newsrooms, fact-checkers and public agencies signed up on their own

Deutsche Welle New York Public Radio The San Francisco Standard Rappler Chequeado Africa Uncensored NEST, Mongolia
dpa Bloomberg Maldita.es La Nación Factchequeado Jagran DeSmog Sakal Media Nepal Fact Check Migrasia The Baltimore Banner Ofcom UNDP UNHCR Internet Archive Columbia Journalism School
100+
newsrooms, fact-checkers, civil society groups and public agencies signed up
21
countries where partners are running investigations
3.7M+
posts collected across those investigations
140+
verified fact-checkers in a network that wants Arbiter for all its members
SRC · simppl.org/about · we work with a growing subset as resources allow ARBITER · SIMPPL

Under the hood: three hard problems

Find the conversations your keywords never name

Coded, multilingual wording shifts weekly; keyword lists miss it.

Map who drives which narrative, across languages

Millions of posts become a navigable map of narratives and actors.

Ask questions of millions of posts at once

An agent writes and runs the analysis, returning cited answers.

Other tools search what they already collected. Arbiter decides what to collect.

SRC · SimPPL methods talk, NYU CSMAP (Mar 2026) ARBITER · SIMPPL

We go beyond the keywords you give us and cover the unknown unknowns

A keyword list only returns posts that used your words, so the conversation you most need to see is exactly the one it misses.

>Climate skepticism and pushback on glacier-melt reporting

Your question gets expanded

Arbiter works out the actors, places and domains your question implies, then goes looking for those.

Organizations · Danish Met Institute
Locations · Antarctica · Greenland
Domains · sepp.org (skeptic) · nature.com
Platforms · X 68 · YouTube 46 · Bluesky 16

Everything collected gets ranked

Keyword search and meaning search fuse, then rerank down to the posts that answer the question.

10,231
collected
→
138
relevant

You get the posts you would have missed

Nobody labels their own post skeptical. Arbiter returns the pushback that shares none of your words.

81% → 99%
relevant results in the top 200, after reranking

Ten platforms plus 20,000+ news sites. A month of cross-platform conversation compiles in 20 to 40 minutes.

SRC · SimPPL retrieval pipeline · climate-skepticism case study ARBITER · SIMPPL

We trace each claim to its first post and everyone who repeated it

One narrative holds the X accounts and the Bluesky accounts together, so a claim that jumps platforms stays one thread.

Map who amplifies whom

Themes show which accounts drive which narrative, in which language, on which platform.

Trace the claim

Each claim links back to the post it came from and to every account that posted the same words. In one US narrative last week we traced 7 claims, and 66 pairs of accounts posted the same text.

Ask, and check the answer

The agent writes and runs the analysis, and answers come back in under a minute. Every row opens into its posts, and we hand-graded 540 agent runs against a reference analysis.

arbiter.simppl.org/narratives · what is being asserted
Arbiter claims provenance graph: the account each claim came from, the accounts that posted the same words, and the claims ranked by how many accounts they reached
SRC · US narrative, week to 16 Aug 2026 · 540 hand-graded agent runs ARBITER · SIMPPL

How partners use Arbiter to map a hidden network, and what changed after

Three investigations, each one a public study anyone can open.

Job scams aimed at migrant workers

With Migrasia, a migrant-worker rights group

Fake ads promise work abroad and deliver people into forced-labour compounds. We collected 2,251 ads across platforms and grouped them into 14 recruiting networks by the phone numbers, scripts and WhatsApp handles they shared. The whole investigation compiled in under two hours.

The Global Anti-Scam Alliance invited us to present the findings in Bangkok, and nominated us for its Scam Fighter Award.

An AI-generated PR wave for a military ruler

Public study, Burkina Faso

Pages promoted Ibrahim Traoré at 100% positive and 100% joy, a uniformity organic discourse almost never shows. The factories and railways in the videos were AI-generated and drew millions of views. A year later the network rebuilt, and we caught it again.

YouTube terminated the channels on its own. We filed no report.

A false ballot claim in a live election

With Maryland’s election officials

A printing error became a claim of 500,000 fake ballots. Across a 1,485-post study we found the same sentence posted by 50+ accounts within hours, and the top post, at 1.32M interactions, demanded a new federal voting law.

Maryland’s election administrator signed a commitment to deploy Arbiter statewide.
SRC · public Arbiter and Migrasia reports · 1,485-post Maryland study ARBITER · SIMPPL

The team, and the board behind it

Swapneel Mehta

Swapneel Mehta · Founder & Executive Director
AI safety at MIT and BU; Integrity Institute board member.

Dhara Mungra

Dhara Mungra · Co-Founder, leads engineering
Ex-Bombora data scientist; NYU M.S. in Data Science.

Six full-time engineers build and run the platform.

Board of directors

Mukund Sudarshan · Anthropic

Christian Schroeder de Witt · University of Oxford

Chirag Raman · TU Delft

Sandra Khalil · All Tech Is Human

Rebecca Harvey · UK FCDO

Karan Dhabalia · entrepreneur (exited to Stripe)

Mansi Panchamia · World Bank

Stacey Peters · ex-VTDigger

Full board and team: simppl.org/team

ARBITER · SIMPPL

What we need from you

The most useful thing you can do is make one introduction.

Platform legal and risk teams

Introduce us to legal or risk counsel at a platform. We run the cross-platform risk assessment one platform cannot run on its own.

Meta, YouTube, and X have each acted on our findings

Newsroom editors

Introduce us to an editor. We support newsgathering, digital audience work, and investigative reporting inside the newsroom.

100+ organizations in 21 countries have signed up

Funders

Grants pay for free access for journalists and fact-checkers in the countries our partners work in.

Past support: Google, Mozilla, Wikimedia, Ford, Omidyar
team@simppl.org ARBITER · SIMPPL

Try it now.

Write to team@simppl.org; we follow up within the week.

A pilot is one case study scoped to your issue. The evidence and sources are yours. The platform is live at arbiter.simppl.org.

SIMPPL · U.S. 501(C)(3) · BOOTSTRAPPED SINCE 2021 ARBITER · SIMPPL

Appendix

The deception signals we score, what institutions did about our reporting, and the three case studies in full.

ARBITER · SIMPPL

Threat intelligence

Sentiment uniformity
Discourse that agrees too much
One campaign read 100% positive everywhere; organic discourse rarely does.
Clickbait score
Bait at 5, camouflage at 1
Heavy bait scores high; polished campaigns score low to pass organic.
Amplification anomaly
Reach with no audience trail
One crypto casino drew 36.9M interactions from four identical posts.
Entity mention share
Who benefits from the story
A banned betting brand stayed most-mentioned through mirror domains.
Impersonation
Borrowed legitimacy
One page posed as 16 real manufacturers to recruit victims.
Sponsorship disclosure
Paid voice posing as organic
Dossiers flag undisclosed promotion that breaks disclosure rules.

These signals add up to tactics, tested from Hindi to Cyrillic.

SRC · public Arbiter studies and partner case reports ARBITER · SIMPPL

When partners publish, institutions act

Partners publish
19 days early

NEST caught a secret defamation bill from under 60 accounts.

Mongolia’s only internationally accredited fact-checkers · public case report
Platforms act
3 platforms

Meta removed a network; YouTube terminated flagged channels; X opened an internal investigation into pro-Russian bots.

Meta Adversarial Threat Report Q1 2024 · Burkina Faso case · X, 2023
Governments act
5 governments

Maryland, California, New York, Odisha, and the Philippines opened conversations.

Maryland committed to a statewide rollout; the rest are scoping.
Status as of July 2026 ARBITER · SIMPPL
Case study Job scams

Fake job ads promise work abroad and deliver workers into scam compounds.

With Migrasia, we traced 2,251 ads into 14 recruiting networks.

The Global Anti-Scam Alliance invited us to present in Bangkok. The investigation compiled in under two hours.

14 recruiting networks mapped
LEGITIMACY SCRIPT489
TESTIMONIALS229
URGENCY167
PHONE NUMBER154
WHATSAPP124

Recruitment tactics by post count, Indonesian study.

SRC · two public Arbiter and Migrasia reports ARBITER · SIMPPL
Case study AI-generated propaganda
01Detect

A fake PR wave for a military ruler

Pages promoted Ibrahim Traoré at 100% positive, 100% joy.

02Identify

The videos are AI-generated

Fake factories and railways drew millions of views.

03Impact

YouTube terminated the channels

We filed no report. A year later the network rebuilt; we caught it again.

ibrahim traoré · key narrative
Traore panel: 100% positive sentiment, 100% joy, Info Pulse top post
POSITIVE100%
JOY100%

Uniformly green panels are the anomaly that starts the investigation.

SRC · public Arbiter study panels ARBITER · SIMPPL
Case study Elections: Maryland's mail-in ballots

What officially happened

A printing error sent some voters the wrong primary ballot. Replacements and unique envelopes rule out duplicate voting.

state board explanation, cross-platform
Maryland State Board of Elections explanation cited across three platforms

What the feed made of it

The claim: 500,000 fake ballots. False per the election administrator and fact-checkers.
Rated false
The claim: a million ballots circulating. Originals plus replacements, recast as duplicates.
Escalation
One sentence, 50+ accounts. The same wording appeared across platforms within hours.
Identical wording

The top post, 1.32M interactions, demanded a new federal voting law. Maryland’s election administrator signed a letter of commitment to deploy statewide.

SRC · 1,485-post study · elections.maryland.gov · FactCheck.org · AP ARBITER · SIMPPL