Logical Fallacies
A practical logical-fallacies reference with clear explanations, usable examples, and teaching tools.
Theory article
In the age of AI, one of the best ways to teach logical fallacies is not to have students passively ask a chatbot for answers, but to have them collaboratively design a Gemini Gem or other pre-prompted agent that must identify, score, explain, rebut, and repair bad reasoning. When students build the analytical frame together, they make their own standards explicit and become much better at seeing both the power and the limits of AI diagnosis.
The goal is to help teachers turn AI from a shortcut into a visible thinking tool. The class uses the model as a provisional analyst whose output must be judged, revised, and improved by students rather than simply accepted.
It avoids the shallow classroom pattern where students paste in a passage, collect labels, and call that critical thinking. The human work here lies in building the prompt, auditing the output, tightening the distinctions, and deciding when the model is overreaching.
The learning payoff comes from making the criteria public and revisable.
The moment students try to write the Gem's instructions, they are forced to decide what counts as good evidence for a fallacy label, what counts as a false positive, and how much quotation is needed to make the diagnosis fair.
A strong agent prompt has to say how to choose among near neighbors. That means students must clarify the differences among Straw man,
Red herring, Ad hominem, False equivalence, and other often-confused entries rather than relying on vague familiarity.
When the agent overlabels, misses a caveat, or confuses two fallacies, the class has concrete material to debug. The failure is no longer hidden inside a teacher lecture; it becomes a visible reasoning event the whole room can inspect.
A good agent does not stop with naming. It also explains the misstep, answers it in plain language, and suggests a stronger rewrite. That keeps the exercise pointed toward better reasoning rather than mere accusation.
This works as a recurring exercise inside a critical thinking course.
Best general sequence: choose a text → build the prompt → run the agent → audit the output → revise the prompt → score and respond.
Use a short editorial, op-ed paragraph, debate exchange, or opening statement. The text should be rich enough to contain real argumentative moves, but short enough that the class can still inspect each quoted line carefully.
Draft the Gem's role, output format, and scoring rules on the board. Ask students what the agent must be prevented from doing, especially in cases where weak or ambiguous reasoning might tempt it into overdiagnosis.
Have the class watch one common output rather than scattering immediately into private runs. That shared output gives everyone the same object to critique.
For each fallacy the agent flags, ask: is the quote sufficient, is the label the most precise one, what nearby labels should be ruled out, and what caveat should be stated before we accept the diagnosis?
Tighten the instructions in response to the model's errors. Over time, the class learns that prompt-writing is really criteria-writing in disguise.
The final class product should not be "what the model said." It should be a human-vetted set of fallacy diagnoses, scores, rebuttals, and repairs.
A useful Gem has a narrow mission and a disciplined output.
The agent should not carpet-bomb a passage with labels. Ask it to identify only the clearest two or three fallacies unless the user specifically requests a full sweep.
Require a short, sufficient quotation for each finding. This keeps the agent from floating free of the text and makes the reasoning misstep inspectable.
A good answer says exactly how the passage moves from evidence to conclusion, where the drift happens, and why that move is too fast, too broad, too selective, or too distracted.
The agent should assign clear scores and then do something constructive: rebut the misstep in plain language, or repair the claim so it says only what the argument has earned.
The best prompts are modular: role, method, format, scoring, and response rules.
Tell the model to act like a careful classroom analyst rather than a prosecuting attorney. This single component sharply reduces overlabeling and rhetorical heat.
You are a careful logical-fallacy analyst for a critical thinking class. Your job is not to hunt for labels aggressively, but to identify only the most justified fallacies in the passage. Prefer specificity over broad accusation. If a suspected fallacy is weak, borderline, or plausibly explained another way, say so clearly. Do not moralize. Diagnose the reasoning, quote the relevant wording, explain the dynamics, and propose a fairer response or repair.
Strong outputs come from strong formatting constraints. If you want organization, specify headings, quote rules, score labels, and link fields directly.
Format every answer with this structure: ◉ Source Summary ➘ One or two sentences summarizing the article, speech, or debate passage. ◉ Fallacy Findings For each fallacy found, use this structure: ➘ Fallacy Name: [most specific label] ➘ Confidence Score (0-4): [number] ➘ Distortion Score (0-4): [number] ➘ Salient Quote: "[quote only the key lines needed to see the misstep]" ➘ Why It Fits: [3-5 sentences explaining the reasoning failure] ➘ Caveat: [state what would make this label too strong or misapplied] ➘ LogFall Link: https://logfall.com/fallacies/[slug]/ ➘ Response: [answer the misstep in plain language] ➘ Repair: [rewrite the claim in a stronger, fairer form] ◉ Overall Pattern ➘ Briefly describe what kind of reasoning drift the piece shows overall.
Separate confidence from importance. A model may be very confident that a minor fallacy is present, or only moderately confident that a major one shapes the passage.
Scoring rubric: 0 = not present or too weak to justify 1 = faint or merely possible 2 = present but modest 3 = strong and clear 4 = central and unmistakable Use two separate scores: ➘ Confidence Score: how sure you are that the label fits ➘ Distortion Score: how much that fallacy affects the passage's overall reasoning
The prompt should say how many fallacies to report, how much to quote, and when to withhold a label. Those guardrails often matter more than the fancy wording at the top.
Passage handling rules: 1. Quote only the lines needed to understand the fallacy. 2. Identify no more than the 3 strongest fallacies unless the user requests a fuller sweep. 3. Compare close alternatives before settling on the final label. 4. Prefer one precise label over several overlapping ones. 5. When no clear fallacy appears, say that directly instead of forcing a diagnosis.
Keep the numbers simple enough for students to compare and debate.
This score answers: how well does the passage justify the label? Students can argue over whether the quotation really supports the diagnosis or whether the evidence is too thin.
This score answers: how much does that fallacy matter to the overall argument? Some fallacies are present but peripheral; others shape the whole reasoning structure.
If you want a richer tool, add one more score for repairability: how easy is it to salvage the argument without giving up its core point?
The class should always be free to lower a score, reject a label, or replace the model's chosen fallacy with a better one. The point is calibration, not obedience.
Grade the students' reasoning about the AI, not merely the AI's output.
Did the group write clear instructions, good output constraints, and fair caveat rules? A vague prompt usually reveals a vague understanding.
Did students catch false positives, weak quotations, and sloppy category choices? The audit is often more pedagogically valuable than the first run itself.
Did the group answer the reasoning mistake clearly and then offer a stronger formulation? This is where the exercise becomes constructive rather than merely classificatory.
Can students say where the agent helped, where it overreached, and what that reveals about both AI and fallacy instruction? That meta-level reflection is part of the lesson.
AI can sharpen the unit, but only if the classroom remains intellectually in charge.
A Gem is a scaffold, not an oracle. Students should always be asked to justify the final judgment independently of the model's wording.
If students think more labels mean a better answer, the exercise will quickly turn into fallacy inflation. Reward precision, restraint, and clean comparison instead.
Some problems are really about bad statistics, weak causal design, or cognitive bias rather than classic fallacy forms. The agent should be taught to say that when needed.
The prompt is the curriculum in compressed form. When students can inspect and revise it, they are learning the criteria themselves rather than merely consuming an answer.
Takeaway
The strongest classroom use of AI is not passive extraction but collaborative calibration. Students learn logical fallacies more deeply when they must tell the agent what to look for, how to quote, how to compare close labels, how to score responsibly, and how to answer a fallacy in plain language once it is found.
Sources that ground the article or push the discussion further.
Tips for Creating Custom Gems (Gemini Apps Help) — Google's own guidance on writing clearer, more detailed Gem instructions.
Use Gems in Gemini Apps (Gemini Apps Help) — Overview of how Gems operate as repeatable custom instruction agents.
Logical Fallacy Detection (Findings of ACL EMNLP 2022) — Foundational NLP paper on logical fallacy detection as a model task.
Robust and Explainable Identification of Logical Fallacies in Natural Language Arguments (arXiv, 2022) — Useful on explainable, staged fallacy identification in natural-language arguments.