43%
trust news in general
The 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
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 analysis6bias dimensions
27bias patterns
3AI models in the ensemble
Bias analysis
Example article
Objectivity7.3
Example analysis · 2027 budget figures · 30 April 2026
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.
Black box
Generative AI with no duty to justify
Language models produce text without evidence. Measuring bias requires verifiable statements — not more black-box verdicts.
Klyptra design principle: verbatim evidence & audit trail
0
comparable German-language tools
Existing 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.
01
Submit
You submit the article text directly — by paste or PDF upload, optionally with a title and source. Klyptra analyzes exactly that text, not the outlet behind it.
- Paste text or PDF
- Title & source optional
- 200–50,000 characters
02
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.
- 3 models in parallel
- Median aggregation
- Verbatim evidence per score
03
Substantiate
Every analysis delivers the six dimensions with a score, one verbatim quote per rating, and markup highlights in the original text — as a shareable permalink.
- Verbatim evidence per dimension
- Markup overlay in the text
- Permalink (30-day TTL)
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?
Whose viewpoint shapes the opening, whose is treated as a reaction? Who is named, who remains unnamed?
Example
“Police clear camp” vs. “Activists cleared out” — same event, different subjects.
02
Word choice
Which words carry judgments?
Loaded language, diminutives, escalation vocabulary. Captured verbatim — Klyptra quotes the words in question literally.
Example
“water down” vs. “make more flexible” — both describe the same thing.
03
Source diversity
How many voices are heard?
Number of directly quoted sources, diversity of political positioning, ratio of primary to secondary sources.
Example
An article with three government and zero opposition voices is not balanced.
04
Fact / opinion
Are observation and judgment kept separate?
Does the text mark commentary as such? Is speculation presented as fact?
Example
“The reform will fail.” — a forecast, not a fact.
05
Completeness
What is left out?
Which relevant background, counter-positions or follow-up aspects go unmentioned? Recognizably one-sided fact selection and context gaps that distort the framing.
Example
Context gaps are often the clearest bias indicators.
06
Emotional balance
How emotionally charged is it?
Sensational language, exclamation marks, outrage markers. Low values signal heavy emotionalization.
Example
“Shock!”, “Scandal!” — typical tabloid markers.
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 — and are evaluated against established benchmarks.
Media Bias Taxonomy
2023Spinde, T., et al. — “Media Bias Taxonomy: A Systematic Literature Review of Computational Approaches to Identifying and Mitigating Media Bias.” ACM Computing Surveys (2023), arXiv:2312.16148.
Theoretical framework for the six bias dimensions. Klyptra's scale operationalizes the taxonomy.
BABE Dataset
2021Spinde, T., Plank, M., Krieger, J.-D., Ruas, T., Gipp, B., Aizawa, A. — “Neural Media Bias Detection Using Distant Supervision With BABE — Bias Annotations By Experts.” Findings of EMNLP 2021.
Training and evaluation dataset with expert-annotated bias labels (3,700 sentences). Calibrates the few-shot examples in Klyptra's prompts.
MBIB Benchmark
2023Wessel, M., Horych, T., Ruas, T., Aizawa, A., Gipp, B., Spinde, T. — “Introducing MBIB — the first Media Bias Identification Benchmark Task and Dataset Collection.” Proceedings of SIGIR 2023.
Comparability between bias-detection systems. Klyptra documents performance on MBIB subtasks.
Horych et al.
2025Horych, T., Mandl, C., Ruas, T., Greiner-Petter, A., Gipp, B., Aizawa, A., Spinde, T. — “The Promises and Pitfalls of LLM Annotations in Dataset Labeling: a Case Study on Media Bias Detection.” Findings of NAACL 2025.
Methodological foundation for handling LLM-based annotations — Klyptra's ensemble setup directly addresses the weaknesses documented in this work.
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 political classification — no left/right axis
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
- MBIB comparison values per quarter
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.