About
Squash is a pioneering, research-grade fact-checking product built by the Reporters' Lab as part of its Tech & Check Cooperative. Designed for live political events such as debates and State of the Union addresses, Squash listens to what politicians say, converts their speech to text in real time, and then queries a large database of previously published fact-checks to find relevant matches. When a match is found, a summary of the fact-check is displayed on screen — giving viewers immediate context. Under the hood, Squash leverages Elasticsearch enhanced with two key filters: ClaimBuster, an algorithm that identifies sentences containing factual claims (filtering out non-claim filler), and a curated set of common synonyms to improve matching quality. The team is also developing subject-based NLP matching, supported by a crowdsourcing game called Caucus that lets human contributors tag thousands of fact-checks with subject labels. To reduce false positives and improve editorial quality, Squash now incorporates a human-in-the-loop model: an editor reviews AI-recommended matches and decides which fact-checks actually appear on screen. This hybrid approach balances speed with accuracy. Squash is primarily aimed at journalists, fact-checkers, newsrooms, and media researchers who want to bring automated accountability to live political coverage. As a research project, it represents a significant step toward scalable, real-time fact-checking for broadcast and streaming environments.
Key Features
- Real-Time Speech-to-Text: Listens to live political speeches and debates, converting spoken words to text instantly for downstream fact-check matching.
- Automated Fact-Check Matching: Uses Elasticsearch with synonym expansion to search a ClaimReview database and surface the most relevant previously published fact-checks.
- ClaimBuster Integration: Filters out non-factual sentences using the ClaimBuster algorithm, focusing processing power only on checkable factual claims.
- Human Editor Override Interface: An editorial dashboard lets human journalists review AI-recommended fact-checks and approve or reject them before they appear on screen.
- Subject-Based NLP Tagging: A crowdsourced game (Caucus) lets contributors assign subject tags to fact-checks, enabling more accurate topic-based NLP matching.
Use Cases
- Displaying relevant fact-checks in real time during televised presidential debates for viewer context.
- Monitoring State of the Union addresses and surfacing prior fact-checks on claims as they are made.
- Helping newsroom editors quickly identify checkable claims during live political events.
- Enriching political video content with on-screen fact-check overlays for streaming audiences.
- Researching and prototyping scalable AI-assisted journalism workflows for live accountability reporting.
Pros
- First-of-Its-Kind Real-Time Coverage: Squash brings automated fact-checking into live broadcast environments, a capability that no mainstream tool offered at the time of its development.
- Human-AI Collaboration: The editor-in-the-loop model ensures that AI speed is balanced with human editorial judgment, reducing embarrassing or misleading false positives.
- Built on Open Fact-Checking Infrastructure: Integrates with the ClaimReview database and ClaimBuster, leveraging the broader fact-checking ecosystem rather than building in isolation.
Cons
- High False Positive Rate: During testing, a significant portion of displayed fact-checks were not relevant to what the speaker actually said, undermining viewer trust.
- Requires Pre-Existing Fact-Check Database: Squash can only surface fact-checks that have already been published and tagged — it cannot generate new fact-checks on the fly.
- Experimental and Research-Stage: Squash is not a production-ready commercial product; it is an ongoing research project with limited public availability and active development needs.
Frequently Asked Questions
How does Squash detect what a politician is claiming?
Squash listens to live audio, converts it to text via speech recognition, and then uses the ClaimBuster algorithm to identify sentences that contain verifiable factual claims, ignoring filler or opinion statements.
Where do the fact-checks come from?
Squash searches a database of previously published fact-checks structured using the ClaimReview schema — a standard used by major fact-checking organizations worldwide.
Is Squash available to the public?
Squash is an experimental research product from the Duke Reporters' Lab. Public availability is limited; it has primarily been tested internally during major political events like the State of the Union address.
How does the human editor interface work?
When Squash identifies a potential fact-check match, instead of automatically displaying it, it sends the recommendation to a human editor who reviews it and decides whether to allow it to appear on screen.
What is the Caucus game mentioned in relation to Squash?
Caucus is a mobile tagging game created by the Reporters' Lab team that allows contributors to assign subject tags to fact-checks. These tags help Squash perform more accurate topic-based NLP matching rather than relying solely on keyword overlap.
