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.
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.
Objectivity scale
Every text is placed on a continuum from 0 (heavily slanted) to 10 (maximally factual).
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.
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.
Framing
Which perspective is treated as the norm?
“Police clear camp” vs. “Activists cleared out” — same event, different subjects.
Word choice
Which words carry judgments?
“water down” vs. “make more flexible” — both describe the same thing.
Source diversity
How many voices are heard?
An article with three government and zero opposition voices is not balanced.
Fact / opinion
Are observation and judgment kept separate?
“The reform will fail.” — a forecast, not a fact.
Completeness
What is left out?
Context gaps are often the clearest bias indicators.
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
2023Theoretical framework for the six bias dimensions. Klyptra's scale operationalizes the taxonomy.
BABE Dataset
2021Training and evaluation dataset with expert-annotated bias labels (3,700 sentences). Calibrates the few-shot examples in Klyptra's prompts.
MBIB Benchmark
2023A 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.
2025Methodological 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.