An open research project
Misinformation is a system. We're studying how it works.
Misinfo Trace detects factual narratives as they emerge, examines the evidence around them, and studies how misinformation originates and spreads — holding every claim to the same standard, whichever side it favors.
The problem
False claims don't travel alone.
Misinformation is usually treated as a string of isolated false statements, each checked on its own. In practice a claim starts somewhere, moves through posts, articles and reposts, gets picked up by recurring sources and amplifiers, and becomes part of a larger narrative.
Studying it as a system means following that whole path — not just ruling on the last sentence someone happened to read.
- CLAIM
- SOURCE
- AMPLIFICATION
- NETWORK
- IMPACT
Emerging
A factual narrative beginning to spread before evidence or correction has caught up with it.
Persistent
A narrative already checked or corrected that keeps resurfacing and reaching new audiences.
Context-distorting
A statement that may be accurate, but misleads because important context is left out — like a short window that hides the long-term trend.
What we're building
Identify. Investigate. Trace.
Research direction — not all of this exists yet
Identify
Detect factual claims and narratives worth investigating — including the same claim in different words — before judging whether they are true.
Investigate
Determine what available evidence supports or contradicts those claims.
Trace
Study where misinformation originates, how it spreads, and whether recurring patterns emerge.
How this differs from fact-checking
Fact-checking is essential work, and it usually begins once a claim is already circulating widely. We are researching whether narratives can be found earlier and routed to human verification sooner.
Detection comes before judgment.
| Conventional fact-checking | What Misinfo Trace aims to do | |
|---|---|---|
| Starts with | A claim someone has chosen to check | Narratives detected as they emerge |
| Timing | Often after the claim has spread widely | While the narrative is still emerging |
| Wording | One phrasing of the claim | The same claim, however it is worded |
| Afterwards | The check is published | Watch for debunked narratives resurfacing |
How it works
From first sighting to published method.
Seven steps, in order. Each is designed to leave a record someone else can check.
Research direction — not all of this exists yet
-
1
Observe
Collect public posts and published articles from open sources, recording when each was published and when we saw it.
-
2
Classify
Keep public-interest factual claims and record what kind of statement each is — directly verifiable, contested, interpretive, normative — without judging whether it is true.
-
3
Investigate
Gather the evidence that supports or contradicts the claim, from sources anyone can inspect.
-
4
Evaluate
Produce a structured assessment with the evidence and confidence behind it. No single AI model decides whether a claim is true, and consequential conclusions are human-reviewable.
- Supported
- Contradicted
- Disputed
- Insufficient evidence
- Unknown
-
5
Trace
Follow how the claim and its rewordings spread over time and across sources.
-
6
Connect
Link related claims into larger narratives, and notice when an old narrative resurfaces.
-
7
Publish
Release methods, evidence and corrections so others can check the work.
Same standard, regardless of side
Same claim. Same evidence. Same standard.
Our goal is to filter, classify and assess every claim by the same public-interest, classification and evidentiary rules — whichever side its political or ideological implications favor.
We don't just assert that. We test it, by running mirrored pairs of claims through the method and checking that they're treated alike.
- Supported
- Contradicted
- Disputed
- Unknown
Research principles
What we hold ourselves to.
- Evidenceover authority
- Transparencyover trust
- Consistencyover ideology
- Uncertaintyover false certainty
- Correctionover defensiveness
- Investigationover accusation
What this project is not
The concerns you should have. Our answers.
- Not a system for policing opinions.
- It evaluates factual claims.
- Not a partisan fact-checking project.
- The same methodology applies regardless of viewpoint.
- Not an automated truth machine.
- No single AI model decides whether a claim is true. Consequential conclusions are human-reviewable.
- Not an authority asking you to accept its conclusions.
- Evidence and methodology are meant to be inspectable.
- Not a system for labeling someone a bad actor because they were wrong.
- Intent and coordination require additional evidence.
Current research
Where the work stands.
Each area is labeled by how far it has actually come. Planned means planned.
- ResearchingUnder active study
- TestingMethod built, being tested
- PlannedOn the roadmap; no results yet
- Researching
Early Warning
Testing whether emerging factual narratives can be detected before they are widely recognized, including the same claim in different words.
- Testing
Defining the Boundary
Determining which kinds of content should qualify for misinformation analysis.
- Testing
Testing for Bias
Testing whether opposed claims receive equivalent treatment under the methodology.
- Planned
Persistent Narratives
Finding already-debunked narratives that keep resurfacing, and tracing where they reappear.
- Planned
Evaluating Claims
Developing repeatable methods for comparing factual claims against available evidence.
- Planned
Tracing Provenance and Networks
Investigating where claims originate, and how recurring sources and amplification relationships can reveal larger patterns.
Get involved
Follow the work, or help with it.
Follow the Research
Get project updates by email: new methods, results and corrections as they're published.
Contribute
Researchers, technologists, journalists and others interested in collaborating: tell us what you work on and where you'd like to help.