Mithazudnak.hu is an AI-driven fact-checking website developed by programmer Németh Marcell in roughly one and a half weeks. While the build was rapid, Marcell credits decades of programming experience and a long-standing interest in politics for making the project feasible.
Marcell has been programming since he was 16. He says that the arrival of modern AI made possible the quick implementation of ideas that earlier had stalled due to lack of time.
Data flow and the three-model pipeline
The site gathers posts published by public figures; videos and images are transcribed and everything is stored in a database. Processing then runs through three distinct AI models:
- the first model breaks the material into individual claims;
- the second model searches for and organizes sources into a knowledge base, compares them to the claims, issues judgments and indexes the items;
- the third model independently verifies results using the same dataset and issues an additional judgment.
Before publication an editorial approval step is required: currently Marcell manually filters which posts should be analyzed and authorizes content before it goes live.
What the site displays
Each analyzed post typically shows several pieces of information:
- a verdict indicating whether a claim is true or false;
- a "Csúsztatási index" (Slanting index) score on a 100-point scale (this index is non-linear and not based on a scientific standard; the AI assigns points based on the aggregate of findings);
- relevant checked claims associated with the post;
- notes on any double standards applied by the post, if relevant;
- a collection of sources and opened links used in the check;
- a summarizing opinion about the post.
The models do not treat every news outlet as an automatic factual source; even outlets considered credible serve as guidance and the system aims to locate original primary sources. For example, a model may accept a Telex article that cites a parliamentary speech, but if the original parliamentary record is available, that is treated as the preferred source.
Methodological limits and source-selection risks
Teczár Szilárd, editor-in-chief of the Lakmusz fact-checking site, says Mithazudnak usually finds sources that appear reliable, but sometimes uses processed versions of data instead of the primary source. This can cause problems if the original dataset is later updated while the processed version is not, leading the system to base a judgment on outdated information.
Regarding the Csúsztatási index, Teczár emphasizes that because it does not follow a pre-established scientific methodology, its scores should not be treated as absolute measurements. He warns that the public may too readily accept an AI-produced judgment as definitive.
What can’t be scored clearly?
Not every post lends itself to a clear true/false classification. Opinionated posts are a typical example: they express feelings rather than factual claims. Such posts may not receive a concrete verdict, but the site provides objective background research so readers can understand the context.
Volume effects and perceived political bias
Marcell says he is not affiliated with any political party, but he was highly active in public discourse during the campaign period. Since the elections, opposition actors have been posting far more frequently than the governing party, and higher posting volume produces more material for analysis. As a result, more evaluations appear on the site about opposition posts simply because there is more to analyze.
This quantitative imbalance can create the impression among some visitors that the site favors the governing side, though Marcell attributes such perceptions mainly to methodological constraints and differences in posting volume rather than intentional bias.
Human fact-checkers vs. AI
Teczár notes that human fact-checkers would often draw lines differently than an AI, and human work brings added value by contacting involved parties for comment or seeking additional data — tasks that AI does not currently perform. Such human-driven follow-ups can be crucial to validate or clarify claims.
Technical setup and costs
The platform and its infrastructure were built with significant AI support. Marcell developed his own programmer agent, Kodein.ai, which he made publicly available and says offers more transparency than some commercial alternatives.
AI usage incurs token-based costs. Marcell subscribes to both Claude and OpenAI models; their subscriptions cost around $200 per month each. The article notes that although the project does not have a continuous 1.1 million forint cost, operating models at scale can become expensive depending on token consumption.
Currently the process is not fully automated: Marcell manually filters posts for analysis, and a complete run of all system analyses for a single post can take 30–40 minutes. Automation may be possible later if enough labeled examples are collected during manual filtering to train the filters.
Future plans
Marcell does not present immediate commercial ambitions for the site; he plans to use it primarily as a professional reference in his career. He intends to shift emphasis toward presenting the objective arguments and positions behind posts rather than focusing mainly on scoring the falsity of claims. According to him, the technical infrastructure already supports this change; what remains is how content will be presented.
Takeaways
Supporters highlight Mithazudnak.hu's speed, capacity to handle large volumes and extensive link collections in its checks. At the same time, experts warn about methodological limits: source selection risks, the non-scientific nature of the Csúsztatási index and the lack of routine human follow-up can all skew judgments. Teczár concludes that while initiatives promoting verifiable public claims are welcome, the public should be cautious about treating AI-backed verdicts as unquestionable.
