Ukrainian developer Serhiy Kornienko created Bullshit Detector, an open-source add-on for AI assistants that checks facts in videos, articles, documents, and posts.
Ukrainian developer Serhiy Kornienko created Bullshit Detector, an open-source add-on for AI assistants that checks facts in videos, articles, documents, and posts.
The developer explained the tool on Reddit. A user can give the AI a piece of material and ask it to process it. The tool will convert the material into text, find testable claims, and search for evidence for each one.
Then the user will receive a report. The AI assistant will tell you which statements were confirmed, which were not, and which could not be verified. Each conclusion will be accompanied by references to sources, and the material will also receive an overall score from 0 to 10. «The main rule is that there is no verdict without a source. The model has no right to say „this is true“ off the top of its head. Only confirmed facts,» explains Serhiy.
Bullshit Detector consists of «skills» — files with instructions and scripts that can be added to compatible AI assistants, such as Claude Code, OpenAI Codex, Cursor. So the add-on itself is free, but for it to work you need your own account in the corresponding AI service. One user jokes: «I wouldn’t mind if this was packaged into a service that I could dump on my mom so she could check her information sources. But now I’m like a fact-checker, it’s hard.»
Serhiy notes that he sees a certain convenience in the fact that his development is an addition to AI agents, rather than a separate service, because for a separate service he would need to host a server and pay for it. «I don’t have the budget yet to make a free fact checker with my keys, and I don’t want to make some kind of freemium with quotas either,» the developer explains.
According to Serhiy, the add-on is able to distinguish between the fact of the statement itself and the truthfulness of what was said. For example, an AI add-on may find that the president made a certain statement, but this does not automatically make the statement true. Therefore, the AI agent will look for confirmation of the president’s words in other sources.
The tool evaluates sources on 5 levels. Primary sources, such as documents, court decisions, and reports, have the greatest weight. This is followed by media, industry publications, and corporate blogs. Anonymous posts and forums are at the lowest level.
However, Bullshit Detector is not an infallible tool. Its result also depends on which pages the AI finds and whether it can access them. So the AI assistant may not see an article on sites such as Reuters, Financial Times, The Guardian, because these and some other publications block automatic reading of their materials. The add-on is also better at checking individual facts than the logic of the entire material. Therefore, when the material uses correct numbers, but the author draws a false conclusion from them, it becomes more difficult for the AI assistant to notice such manipulation.
There is no independent testing of Bullshit Detector yet. The author himself continues to refine the project, and asks users to report what works poorly or what is missing. The project is distributed under the MIT license. Its code is available on GitHub.
