CHANTAL PISARZOWSKI

®2024

Parallaxe

Parallaxe

AI Mediation in political discussions

AI Mediation in political discussions

project type

AI Mediation

Year

2025

Role

Project Lead

technology

Multi-Agent AI, NLP, Real-Time Fact Verification

Parallaxe is an AI-mediated dialogue system developed in cooperation with Fraunhofer HHI that brings people with opposing political views into conversation — through neutral rephrasing and evidence-gated fact-checking in real time.

PROJECT DESCRIPTION

Parallaxe

Parallaxe is an AI-mediated dialogue system that connects people with opposing political views and fact-checks their claims in real time. A Mediation Layer guides the conversation: it rephrases charged language into neutral phrasing, translates across languages, and surfaces common ground. A Verification Layer triggers multi-agent fact-checking whenever a verifiable claim is made — issuing corrections only when evidence is sufficient, and communicating uncertainty when it is not.

Showreel

Showreel

Showreel

CONCEPT

Why it matters

Democratic societies are struggling to maintain a shared factual basis: algorithms amplify filter bubbles, disinformation spreads rapidly, exchange across ideological boundaries is increasingly rare. At the same time, fact-checking itself can become the problem — when sources are chosen subjectively, when opinions are treated as facts, or when the system feigns certainty it doesn't have. parallaxe addresses both sides: it brings people back into conversation while ensuring that verification follows strict evidence standards. No verdict without an evidence chain, no correction without proof. The goal is not to tell people what to think, but to rebuild the conditions under which productive disagreement becomes possible again.

Democratic societies are struggling to maintain a shared factual basis: algorithms amplify filter bubbles, disinformation spreads rapidly, exchange across ideological boundaries is increasingly rare. At the same time, fact-checking itself can become the problem — when sources are chosen subjectively, when opinions are treated as facts, or when the system feigns certainty it doesn't have. parallaxe addresses both sides: it brings people back into conversation while ensuring that verification follows strict evidence standards. No verdict without an evidence chain, no correction without proof. The goal is not to tell people what to think, but to rebuild the conditions under which productive disagreement becomes possible again.

PROJECT DESCRIPTION

Technical Architecture

Parallaxe is a hybrid system built on two layers. The Mediation Layer guides dialogue, rephrases charged language, asks contextual follow-ups, translates between languages, and surfaces common ground — without forcing agreement. The Verification Layer uses a multi-agent pipeline: a Classifier Agent categorizes each claim, an Authority Registry routes it to the appropriate source type, Verifier Agents check against primary sources, and a Judge Agent issues a verdict or abstains when evidence is insufficient. Core principles: Claim-Type → Source-Type Routing reduces reliance on subjective trust lists. Calibrated Abstention communicates uncertainty rather than issuing unsupported verdicts. Verifiable claims are checked; value judgments receive context instead of binary verdicts.

Parallaxe is a hybrid system built on two layers. The Mediation Layer guides dialogue, rephrases charged language, asks contextual follow-ups, translates between languages, and surfaces common ground — without forcing agreement. The Verification Layer uses a multi-agent pipeline: a Classifier Agent categorizes each claim, an Authority Registry routes it to the appropriate source type, Verifier Agents check against primary sources, and a Judge Agent issues a verdict or abstains when evidence is insufficient. Core principles: Claim-Type → Source-Type Routing reduces reliance on subjective trust lists. Calibrated Abstention communicates uncertainty rather than issuing unsupported verdicts. Verifiable claims are checked; value judgments receive context instead of binary verdicts.

Parallaxe

Achievements

Exhibitions

Berlin Science Week

Parallaxe

Parallaxe

Parallaxe

Credits

Prof. Andreas Ingerl

Prof. Andreas Ingerl

Marcel Bückner

Marcel Bückner

Thomas Koch (Fraunhofer HHI)

Thomas Koch (Fraunhofer HHI)

Supervisior
Supervisior
Supervisior
Supervisior
Supervisior
Supervisior
Cooperation Partner
Cooperation Partner
Cooperation Partner