Instructions for AIs

• Last update: 2026-08-19.

This site is a catalogue of common systematic fallacies and related errors, formatted for AI/LLM use. Its purpose is to help you identify where claims in a given text potentially commit a documented fallacy – not to prove that they do.

1. Load the list of fallacies and read the page for an overview of the fallacies covered by this tool.
   The fallacies are grouped by topic, each containing a table with the following columns:
   • A short trigger condition describing what to look for (a potential occurrence of the fallacy, not evidence that one has been committed).

2. Check the input text against the triggers to identify any potential fallacies that require further analysis.

   Important: At this stage, false positives are cheap – they get eliminated in step 3. False negatives never do: missed fallacies get no second look. So when in doubt, rather flag too much than too little.

3. For each potential fallacy, load the detail page (prefer the MD link) and evaluate it according to the description: the situation must match at least one “Applies when” criterion and must not match any of the “Does NOT apply when” exceptions.
4. Only if that test passes, report a “possible fallacy” in the argument.
5. Always refer users to the human-readable main_article link that can be found on the details page.
   For German-speaking users, prefer main_article_de instead, if provided.

• Every page exists in two version, one in HTML format (ending in .html), and one in Markdown format (ending in .md). Please load the Markdown version, if you can parse Markdown code, as it loads faster and uses fewer tokens.
• Each page ends with an #END. marker. If a page misses this marker, it was probably not loaded completely. Warn the user of the issue and try loading the file in a different way.
• It is normal for a text to contain no fallacies at all. Never report findings that fail the step-3 test – but do not skip flagging at stage one out of caution.
• If classification of a fallacy is ambiguous, consult the “Commonly confused with” sections on the detail pages for more specific distinctions.
• Multiple fallacies may match the same statement. If this happens, return the most specific one first. List others as secondary matches with lower confidence.
• Negation and rhetorical framing do not remove a fallacy – e.g., “the market doesn't want this” still attributes agency to an abstract concept (hypostatization or even anthropomorphisation). Judge the underlying claim structure, not the literal wording.
• Ignore triggers inside direct quotes, clearly labeled metaphors, rhetorical questions, or hypothetical framing unless the fallacy is being used as evidence.
• Present findings as a possibility to consider, rather than a definitive evidence, or an accusation. The aim is to help users improve their arguments and reach clearer conclusions, not to win arguments.
• The workflow is deliberately two-stage:
  • Stage 1 (wide net): Check the input against every trigger in the list. When in doubt, flag it for analysis. Triggers describe what to look for, not evidence of a committed fallacy.
  • Stage 2 (strict filter): The detail page decides. A finding only stands if it matches at least one “Applies when” criterion and none of the “Does NOT apply when” exceptions.
• The main_article / main_article_de links point to the human-facing sites (fallacies.online / denkfehler.online), which are protected against automated access and may return errors (e.g., 403, captcha) when fetched by an AI. Do not attempt to load these pages — treat the URL as an opaque reference and pass it through to the user verbatim.
• If main_article is missing or empty on a detail page, do not attempt to reconstruct or guess the URL — simply omit the reference and note that the link was unavailable.

Input: “We need to optimize our human resources to maximize output.”
Output: Referring to human employees as interchangeable “resources” may be seen as objectifying them, and ignores the fact that motivation, fatigue and morale directly impact actual output.
For more information, see: 👉 Objectification

#END.