Closed beta · Access on request · Research-based methodology

Media bias. Research-based analysis.

Klyptra analyzes German news articles across six research-backed bias dimensions — with verbatim evidence, a multi-model ensemble, and complete audit trails. No left-right verdict. Just language patterns.

Or look at an example analysis
6
bias dimensions
27
bias patterns
3
AI models in the ensemble

The problem

Trust is falling. Distortion is growing. Tools are missing.

Media bias can be studied systematically — but only if the method itself is transparent and verifiable. That is exactly where Klyptra comes in.

43%trust news in generalThe lowest value since 2015 — trust in Germany has fallen continuously since the survey began. Trust in one's own chosen sources stands at 53%.Reuters Institute Digital News Report 2024 — findings for Germany
Black boxGenerative AI with no duty to justifyLanguage models produce text without evidence. Measuring bias requires verifiable statements — not more black-box verdicts.Klyptra design principle: verbatim evidence & audit trail
0comparable German-language toolsExisting tools (AllSides, Ad Fontes Media) are US-focused. For German sources, the German language and German discourse, there has been no systematic coverage so far.Market research, as of April 2026

How Klyptra works

Three steps. No black box.

Every step is documented and traceable. Every score can be traced back to the individual piece of evidence in the original text.

Analyze

An ensemble of three independent language models — from three pipelines (OpenAI/USA, Mistral/EU, DeepSeek/China) — scores each text across six dimensions, on a 0-to-10 scale with verbatim evidence.

USA
OpenAI
EU
Mistral
China
DeepSeek

Objectivity scale

Every text is placed on a continuum from 0 (heavily slanted) to 10 (maximally factual).

0510

Media Bias Taxonomy · 2023

Theoretical framework for the six bias dimensions. Klyptra's scale operationalizes the taxonomy. The system detects 27 specific bias patterns — from framing to selective fact selection.

FramingWord choiceOmission+ 24 more

The six dimensions

Bias isn't a single value. It's six.

Klyptra never reduces an article to one number. The six dimensions operationalize the Media Bias Taxonomy (Spinde et al. 2023) and can be evaluated individually — an article can be weak on framing yet strong on source diversity.

01

Framing

Which perspective is treated as the norm?

“Police clear camp” vs. “Activists cleared out” — same event, different subjects.

02

Word choice

Which words carry judgments?

“water down” vs. “make more flexible” — both describe the same thing.

03

Source diversity

How many voices are heard?

An article with three government and zero opposition voices is not balanced.

04

Fact / opinion

Are observation and judgment kept separate?

“The reform will fail.” — a forecast, not a fact.

05

Completeness

What is left out?

Context gaps are often the clearest bias indicators.

06

Emotional balance

How emotionally charged is it?

“Shock!”, “Scandal!” — typical tabloid markers.

Who it's for

Three audiences, one tool.

Klyptra isn't equally useful to everyone — but for each of these groups it solves a clearly defined problem.

Citizens

Anyone who reads the news should know what framing it carries — without having to complete a linguistics degree.

  • Analyze individual articles — six bias dimensions with evidence
  • Result as a shareable permalink
  • No rating of newsrooms — any discernible political leaning is flagged as interpretation and does not feed into the objectivity score

Newsrooms

Editorial teams can check individual texts for framing and balance before publication.

  • Analyze your own articles across six dimensions
  • Verbatim evidence of which phrasings carry the bias
  • Evidence chains available for internal discussion

Research

Reproducible methodology, documented model versions and JSON exports for empirical media research.

  • JSON exports of all analyses
  • Documented prompts and model versions
  • Full model provenance for every analysis

Scientific basis

Not a home-made scale.

Klyptra doesn't invent its own theory of bias. Methodology and scale rest on published, peer-reviewed research. An evaluation against established benchmarks is in preparation and has not yet been carried out.

  • Media Bias Taxonomy

    2023

    Theoretical framework for the six bias dimensions. Klyptra's scale operationalizes the taxonomy.

  • BABE Dataset

    2021

    Training and evaluation dataset with expert-annotated bias labels (3,700 sentences). Calibrates the few-shot examples in Klyptra's prompts.

  • MBIB Benchmark

    2023

    A reference frame for comparability between bias-detection systems. Used as a yardstick for a planned evaluation — Klyptra has not yet published any MBIB figures.

  • Horych et al.

    2025

    Methodological foundation for handling LLM-based annotations — Klyptra's ensemble setup directly addresses the weaknesses documented in this work.

Beta access

The closed beta is live — access on request.

Enter your email address. We review every request manually and, on approval, send you a personal access key — research and newsrooms first, then individuals.

No spam. No sharing. Your email is used only to process your beta request and is deleted on request.

By submitting, you accept our privacy policy.