FactNinja is a tool for analyzing visual content — posters, memes, social media screenshots, photos with text, propaganda graphics. When you upload an image, we don't perform one analysis, but a whole cascade of specialized views that complement each other. Each view answers a different question — what is written there, what is it trying to evoke in you, what claims does it contain, was it created by artificial intelligence.
This multilayered approach is intentional. A simple verdict like "truth" or "lie" would be misleading — propaganda is often a combination of true statements in a manipulative framework, or conversely, honest communication with clumsy formulations. Instead of a conclusion, we present you with a breakdown: individual layers that you can piece together into your own judgment. This text aims to explain how our tools work inside — what science underlies them, where their strengths lie, and where their limits are.
What Happens When You Upload an Image
The moment you click Analyze, the system launches a cascade of separate tasks in the background. Each addresses one aspect of the image and runs independently — the failure of one layer does not crash the others. For example, if text recognition technically fails on a very stylized meme, the main analysis and emotional breakdown continue. The entire cascade typically takes several tens of seconds to two minutes — depending on the complexity of the image and the load on AI services. While waiting, you see live indicators of what is happening.
Assistant Analyses — Perspectives, Not Truths
When you choose an assistant for analysis, you are not selecting different artificial intelligences — you are choosing a different role that the same AI plays. Each assistant has a different system prompt in its head, determining from what angle it views the image. The default FactNinja assistant is balanced and covers key points. The media literacy professor has a pedagogical tone and focuses on what is important for teaching critical thinking. Other assistants have their own specifics, which you will find in the app when selecting them.
The same image run through different assistants may have different accents — and that is intentional, not a mistake. A single poster can be legitimately viewed from multiple angles. Sometimes you want a pedagogical breakdown, other times a factual balance. The Visual Insights sidebar shows all previous perspectives together so you can compare them. Multiple perspectives are a strength in our methodology.
How AI "Sees" an Image
Under the hood, we use vision-capable language models — multimodal artificial intelligence from OpenAI, currently the GPT-5 and GPT-5-nano models. These models were trained on massive datasets of images with captions and can describe, categorize, and interpret visual content in natural language. When AI sees a poster with Lenin, it's not just about recognizing the face — the model understands Soviet constructivism typography, identifies rhetorical patterns, connects the image with historical context.
Importantly, AI does not see like you. It doesn't count pixels, it doesn't evaluate RGB values. It works with a learned representation — an abstract understanding of the visual world built from training data. This means it recognizes typical patterns, but sometimes sees what is not there (hallucinations), and sometimes misses nuances that a human observer would catch at first glance.
⚠ Key Principle: FactNinja is a helper, not a judge. Analysis is a starting point for your own thinking — not a verdict. Multiple perspectives, scientific foundations, and transparent methodology exist precisely so you can make the final judgment yourself.
Augmented Data — What AI Specifically Finds
Alongside the free text of the main analysis, AI also returns a structured output. These structured fields are key because they enable visualizations (donuts, maps, graphs), filtering, and further processing.
Key Topics are a list of what AI captured in the image — from specific objects to visual styles to abstract concepts. Verifiable Claims are specific statements that can be independently verified. Each has a status: true, partially true, false, misleading, or unverifiable. Detected Logical Fallacies are argumentative errors linked to our fallacy glossary. Lay Summary is a short, understandable explanation for someone without context.
Visual Insights — Emotional Radiology
Visual Insights is our flagship feature — and it deserves its own explanation because it is based on two scientific models with almost forty years of history and which are classics in the psychology of emotions. Access to Visual Insights is available from each analysis via the cyan Visual Insights button, which opens a separate page with the emotional profile of the material.
🎯 Key Principle: Visual Insights measures what the material wants to evoke, not what you feel. The same poster can evoke pride in one viewer and disgust in another — our analysis describes the author's intent, not the audience's reaction. This is a key difference for correctly reading the results.
Plutchik's Wheel of Emotions
Robert Plutchik, an American psychologist with a background in biology, published a seminal work in 1980 A general psychoevolutionary theory of emotion. He argued that emotions are not cultural constructs but evolutionarily old reactions serving survival. Joy pushes us towards what benefits reproduction and social bonds. Fear leads us away from danger. Anger gives us the energy to confront a threat. And that there are eight basic emotions from which all others are composed, similar to how all colors are composed of three primary pigments. He arranged these eight emotions into a circle — neighboring emotions are close, opposites are opposites.
In the app, for each of the eight emotions, we indicate intensity on a scale of 0 to 100, depending on how strongly the material evokes this emotion. In the Plutchik wheel graph, this is displayed as a "blooming flower" — the longer the petal of a certain color, the stronger the emotional appeal of that type. A heroized portrait of a leader with a flag typically shows high values of joy, trust, and anticipation. A poster warning of an enemy shows strong fear, disgust, and anger (as in the illustration above).
Dyads — Compound Emotions
Plutchik's theory goes further — combinations of two neighboring basic emotions create dyads, compound emotions with their own names. Joy together with trust forms love. Fear with surprise creates awe. Sadness with disgust gives regret. Anger with anticipation produces aggressiveness. The Dyads section in Visual Insights shows which of these combinations are strongest in the given material — and this often says more about the author's intent than individual basic emotions.
Russell's Model — Valence and Activation
The second scientific foundation is the work of James A. Russell, also published in 1980. Russell argued that emotions can be broken down into two orthogonal dimensions. The first dimension is valence — how pleasant or unpleasant the experience of the emotion is. The second is activation — how calm or excited the internal state it evokes.
Visual Insights displays in Russell's space one dot — the overall emotional "center of gravity" of the material. This helps to see where the material is leaning faster than a list of individual emotions. A poster with a heroized leader ends up typically in the "excitement, joy" quadrant. A poster with a devastated land and hungry children ends up in "anxiety, anger". Both models (Plutchik's circular map and Russell's 2D map) complement and verify each other.
Level of Manipulation
Alongside the emotional analysis, Visual Insights also evaluates the overall level of manipulation — a synthetic evaluation combining the strength of emotions, the number and type of rhetorical tricks used, and the overall intensity of the material's persuasive apparatus.
The key when reading the level of manipulation is: manipulation is not the same as a lie. The material can be factually true and still very manipulative. Conversely, the material can be LOW manipulative and contain misinformation. The level of manipulation measures the strength of the persuasive apparatus, not the truth value of the content.
OCR — We See Text in the Image
FactNinja can extract all readable text from an image. We use the GPT-4o-mini model for recognition, which is trained to see and read in challenging conditions: handwritten text, stylized typography, photographed text at an angle, low-resolution mobile screenshots. It works surprisingly well even where traditional OCR tools fail.
Verbatim Extraction
Our OCR has a key principle: we transcribe exactly what is visible. Without correcting typos, without adding missing context. If a letter is missing on the poster, it is also missing in our OCR output. This approach is intentional — OCR for us is not a literary transcription, it is a forensic trace. For journalists, historians, and fact-checkers, the difference between "mobilization" and "mobilisation" is a key dating clue.
Multilingual OCR and Automatic Translation
We support OCR in practically all languages — recognizing Latin, Cyrillic, Chinese characters, Japanese hiragana and katakana, Korean hangul, Arabic, Hebrew, Devanagari, Greek, Thai, and others. Detection of the actual source language is based on Unicode ranges, so it works reliably even with mixed texts.
If the text is in a different language than your interface, we display both: original on the left, translation on the right. We keep the original for forensic value, the translation provides understanding of the content.
AI Detection — Was It Created by Artificial Intelligence?
With the growing availability of generative artificial intelligence, recognizing AI-generated content is an important context. For propaganda analysis, it is crucial — a photograph of a real event has different evidential value than an AI-generated scene, even if both look identical.
We use an external service Sightengine for AI content detection, which returns the probability of AI generation in percentages. Sightengine is trained on millions of AI-generated and authentic images and detects typical artifacts of generative models.
It's important to know what AI detection cannot do. Photo-realistic generations of the latest models sometimes slip through with low probability. Edited real photos (Photoshop, deepfakes based on real footage) may not be captured by detection. And overall, a 70% AI probability does not mean "definitely AI" — it means "AI signal is strong, but human judgment is needed". The score is a tool to focus your attention, not an automatic verdict.
Multilingualism — Analysis Lives in All Languages
FactNinja supports six languages: Czech, English, Slovak, Polish, German, and Spanish. Analysis texts and comments are automatically translated between these languages in the background. When an author writes an analysis in Czech, the system immediately translates it into the remaining five languages. A page visitor sees the analysis in the language set in their browser.
This multilingualism means that a Czech analysis of a Russian propaganda poster is automatically available to English, Slovak, Polish, German, and Spanish audiences — without the author's intervention. Your text reaches an international audience on its own.
Limits of the Helper — What FactNinja Doesn't Do
For fair expectations, it is necessary to explicitly state what you cannot expect from FactNinja. Blind trust in AI tools leads to bad decisions. Conversely, calibrated expectations make FactNinja a powerful helper.
We Do Not Judge "Truth" as a Whole
The analysis does not end with a verdict like "This poster is true" or "This poster is a lie". A poster may contain true claims in manipulative rhetoric, false claims with a neutral tone, or be symbolic, where the question of truth doesn't even make sense. Visual Insights measures emotional appeal, augmented data identifies specific claims, but the overall verdict always belongs to you.
We Don't Catch Everything
AI makes mistakes. Sometimes it sees what is not there, sometimes it misses what is there. Your own contextual judgment, cross-verification with other sources, and consultation with experts in critical cases are always necessary. FactNinja speeds up and structures your thinking, but it does not replace it.
We Are Not a Substitute for Journalistic Work
Analysis of a propaganda image does not end with FactNinja. Reverse image search (Google Lens, Yandex, Bing, TinEye) finds where the image comes from. Archival databases like the Internet Archive show when the material first appeared. Consultations with historians and regional experts supplement what AI cannot do. FactNinja is a starting point. The rest of the work is human.
Scientific Foundations
For a deeper dive into the theories on which FactNinja is based, we recommend the following literature.
Emotions and Propaganda
Robert Plutchik published in 1980 the work "A general psychoevolutionary theory of emotion" in the book Emotion: Theory, research, and experience. This work is the foundation of the entire Plutchik model of eight emotions that we use. James A. Russell published in the same year 1980 the work "A circumplex model of affect", on which our 2D map of valence and activation is based. Paul Ekman published in 1992 "An argument for basic emotions" — an argument for the existence of universal basic emotions across cultures.
Rhetorical and Manipulative Techniques
Robert B. Cialdini in the book Influence: The Psychology of Persuasion (1984) described six key principles of persuasion — reciprocity, commitment, social proof, authority, liking, scarcity. Steven A. McCornack published in 1992 "Information manipulation theory" — a theory on how truth can be manipulated through omission, ambiguity, and relevance. Douglas Walton in the book Informal Logic: A Pragmatic Approach (2008) systematically addressed argumentative fallacies.
Visual Propaganda and Media
The classic work of Jacques Ellul Propaganda: The Formation of Men's Attitudes (1965) laid the foundation for modern propaganda studies. For a more contemporary overview, we recommend Propaganda and Persuasion (2018) by Jowett and O'Donnell.
How to Read Analysis Critically — Three Rules
In conclusion, three simple rules that will help you interpret FactNinja results meaningfully.
1. Visual Insights Describe Intent, Not Truth
High values of fear and low trust in Plutchik's wheel mean that the material is a scare appeal. But such an appeal can be justified (warning of a real threat) or abused (manufactured panic). Visual Insights won't tell you which case it is — the context you bring does.
2. Manipulation Is Not the Same as a Lie
A VERY HIGH manipulation rating means that the material strongly shapes opinion. It can be a lie, but it can also be a very well-done campaign for a good cause. By "manipulation" we don't mean "a bad thing" — we mean a strong persuasive apparatus. A vaccination campaign ad can be VERY HIGH manipulative and yet socially beneficial.
3. More Views Are Strength
One material, multiple assistant analyses, Visual Insights, OCR translation, reverse image search — each of these layers adds a piece of the puzzle. None of them is the final word. The final word is you. Our job is to provide you with tools, transparently describe the methodology, and provide scientifically supported data. Your job is to make sense of it.
— The FactNinja Team