Tackling the Root of Misinformation by Teaching Laypeople about Logical Fallacies via Socratic Questioning and Critical Argumentation

Synthesis: Shi, Wang, Ni, Pal Chowdhury and Sachan ask whether an Large Language Models (LLMs) can teach laypeople to recognize logical fallacies rather than simply flag them, and build LFTutor, an Intelligent Tutoring system that steers every turn of a debate through intent-based pedagogical strategies. The framework decomposes a student's reply with the Toulmin model of argument structure, detects one of four student intents, selects one of four strategies (Responding, Evidence, Assumption, Refutation) ordered by Toulmin component priority, and routes the drafted reply through a verifier agent before it is sent. Because no established metric set existed for this task, the authors derived eight evaluation dimensions from a pilot study and from education and argumentation literature: Divergence, Stance Change, Repetition, Lack of Refutation, Lack of Evidence Inquiry, Strategy Fixation, Unexplained LF Terms and Passive Guidance. Automatic evaluation with GPT-4o over 1,000 dialogues per framework gives LFTutor an average of 84.5% of dialogues passing each metric, against 61.5% for a heuristics-enhanced baseline and 31.2% for a role-play baseline; with adversarial students the gap narrows to 60.7% versus 30.2%. In a human study, 20 participants rated LFTutor significantly better on eight of nine metrics. The contribution is double: a reusable way to measure Critical Thinking tutoring quality through dialogue-level evaluation, and evidence that structured dialogue control, not a better prompt, is what separates a fallacy tutor from a debating chatbot.

Key Findings

1. LFTutor outperforms both baselines. Over 1,000 GPT-4o dialogues per framework judged by QwQ-32B, it averaged 84.5% of dialogues passing each metric, against 61.5% for BASE W/ PROBLEMS and 31.2% for BASE.
2. The framework's controls, not its vocabulary, drive the gains. Four ablations are reported (removing the disagreement check, removing strategy verification and correction, reversing the priority order of the pedagogical strategies, and allowing multiple strategies per turn), and each lowers pass rates on a subset of the metrics relative to the full system, with strategy verification the single largest contributor.
3. Strategy order matters. Reversing the priority from Responding-Evidence-Assumption-Refutation to Refutation-Assumption-Evidence-Responding reduces pass rates, so the fixed ordering is doing work beyond the presence of the four strategies.
4. Ablation evidence is bounded by budget. The ablation study was run on a randomly sampled subset of the sentences rather than the full set, which the authors give as a cost limitation.
5. The disagreement bank protects guidance. Removing the Disagreement Check lowered Passive Guidance by 13.6 points (43.6 to 30.0) and raised Repetition by 13.7 points.
6. Adversarial students compress but do not erase the gain. LFTutor averaged 60.7% against 30.2%, with the widest margin on Unexplained LF Terms (86.2 vs 49.5).
7. Human raters favored LFTutor on eight of nine metrics. On a 1 to 5 Likert scale, Divergence rose from 1.65 to 3.30, Lack of Evidence Inquiry from 2.65 to 4.20, and Helpfulness from 1.65 to 4.15.
8. Repetition was the exception. Participants rated LFTutor 3.10 against 2.65 (p = 0.08), because real participants varied their tactics and gave the baseline fewer openings to repeat; Passive Guidance stayed weakest at 43.6% automatic.

How the framework works

LFTutor runs four steps each turn. The Disagreement Check consults a long-term memory bank of disagreement points: a genuinely new opposing claim is logged and handled, while repeated examples or assumptions cause the tutor to skip ahead and ask for fresh arguments. Intent Detection then classifies the reply into one or more of four intents mapped to Toulmin components: a request for clarification, missing evidence, missing assumptions, or a complete but flawed argument. Intent-based Strategy Selection picks one strategy from Responding, Evidence, Assumption and Refutation, following the priority of claim, grounds, warrant and rebuttal. Verified Strategy Execution drafts the reply and has a separate agent check that it respects the chosen strategy and stays on topic, with a failure triggering a rephrase. The design draws on Socratic questioning for open probing and on critical argumentation for targeted refutation, implemented as a Pedagogical Agent on GPT-4o with no fine-tuning.

What the automatic evaluation shows

Automatic evaluation used Elec2Deb20, a set of U.S. presidential debate sentences from 1960 to 2020, with student turns played by simulated students. Normal students defended the speaker's position with evidence; adversarial personas were built to trigger each known tutor weakness deliberately. Across eight metrics LFTutor led on almost every dimension, most sharply on Stance Change (87.9 against 48.1 for BASE W/ PROBLEMS) and Repetition (78.3 against 42.5). The heuristics baseline that was told to avoid known problems performed well early in a dialogue but lost focus as the conversation lengthened, which the authors read as evidence that warning a model about its flaws does not sustain quality. Under adversarial pressure the advantage narrowed to an average of 60.7%, and LFTutor's Divergence and Strategy Fixation fell because effort shifted from explaining fallacies to fending off irrelevant content.

What the human study found

Twenty participants fluent in English debated each tutor for five to ten rounds on two Elec2Deb20 sentences they had previously judged logically valid, then rated tutor behavior on a 1 to 5 Likert scale. LFTutor scored significantly higher on eight of nine metrics, highly significant for Divergence, Stance Change, Lack of Refutation, Lack of Evidence Inquiry and Passive Guidance (p = 0.00 in the paper's table), and significant for Strategy Fixation, Unexplained LF Terms and Helpfulness (p = 0.01 to 0.03). Qualitative themes matched the numbers: ten of twenty participants called LFTutor more focused, nine praised its ample guidance, thirteen described BASE as debating in a general setting without specificity, and six found it easily persuaded. The authors note that real participants argued with more varied and inconsistent tactics than the simulated students, including emotional appeals, which is why Repetition behaved differently across the two settings.

Where it falls short

Passive Guidance is the weak point: LFTutor passed it in only 43.6% of automatic dialogues. The authors attribute this to the absence of high-level, ordered dialogue goals that would let the tutor steer the conversation rather than react turn by turn, and propose combining such goals with the existing strategies. The framework also degrades when a student pushes it, with Divergence and Strategy Fixation falling under adversarial pressure because the tutor spends turns managing off-topic content. Cost bounds the evidence: the ablation study covered only 100 randomly sampled sentences because of budget limits, and automatic judging rests on one open reasoning model. The metric set itself omits dimensions the authors acknowledge matter, including emotional support, personalization and ease of understanding. Improvement over the baseline is real but uneven rather than uniform.

What this means for practice

• Model developers. Treat dialogue control as a pipeline, not a prompt. The ablation evidence points to per-turn intent detection plus an independent verification step as the load-bearing parts, so build and test those as separate stages rather than folding them into one system message.
• Researchers. Adopt the eight-dimension rubric when comparing fallacy or argumentation tutors, and validate any automatic judge against human labels on your own task before reporting judge-scored gains.
• Educators. Favor a tutor that commits to one pedagogical move per turn and holds its position under pressure; a model that concedes or mirrors the learner does not build the critical stance the exercise is meant to produce.
• Assessment designers. Score tutoring dialogues at the dialogue level against explicit pass or fail criteria, and expect guidance-style dimensions to need ordered goals before a tutor can score well on them.

Limitations

• The human study used 20 participants who were predominantly master's-level or doctoral students in technical fields, so the "laypeople" framing is broader than the sample.
• Automatic evaluation rests on a single LLM judge validated at F1 above 73% per metric against three annotators, checked on 100 sampled dialogues, so judge-scored gains inherit that error rate; the metric set also excludes emotional support, personalization and ease of understanding, which the authors name as gaps rather than as measured zeros.
• The ablation study used only 100 randomly sampled sentences and the normal student setting, so the component contributions carry wider uncertainty than the main results.
• All tutoring dialogues are generated by simulated GPT-4o students on debate sentences from one dataset covering U.S. presidential debates, and participants argued with more varied tactics, which changed the Repetition result between settings.

Connected Concepts

• Socratic Method — open probing of the student's reasoning is one of the two pedagogical foundations
• Critical Thinking — fallacy recognition is treated as a trainable critical-thinking outcome
• Intelligent Tutoring — LFTutor is positioned as an LLM tutor, not a fallacy detector
• Pedagogical Agent — the tutor is implemented as a multi-agent pipeline with a verifier
• Conversational AI — the intervention happens entirely in multi-turn debate dialogue
• Simulating Students — automatic evaluation depends on GPT-4o student personas, normal and adversarial
• AI Sycophancy — concession to persuasion is one of the failure modes the design targets
• Scaffolding — strategies are ordered to build the argument component by component
• Prompt Engineering — the baseline comparison isolates structured steering from prompt heuristics
• Assessment Validity — the metric set and judge validation are the paper's measurement contribution
• AI Ed Evaluation — the eight-dimension rubric is an evaluation framework for tutoring dialogues
• Critical Pedagogy — teaching laypeople to resist manipulative argument connects to critical pedagogy aims

Connected Articles

• Misconception Acquisition Dynamics in Large Language Models — what training data a model needs before it can hold or recognize a faulty reasoning pattern
• The Socratic trap: Benchmarking the capacity of large language models to generate strategic misconceptions in computer science education — how Socratic questioning interacts with strategic misconceptions
• Towards the Pedagogical Steering of Large Language Models for Tutoring: A Case Study with Modeling Productive Failure — the pedagogical steering idea LFTutor builds on
• Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most — a review of LLM tutoring agents and their pedagogical behavior
• Sycophancy is an Educational Safety Risk: Why LLM Tutors Need Sycophancy Benchmarks — sycophancy as an educational safety problem
• The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows — how models identify and misattribute logical fallacies
• A Bottom-Up Taxonomy of Student Discourse with a Socratic AI Physics Tutor — a taxonomy of Socratic tutoring moves in another subject
• Analyzing Undergraduate Problem-Solving in Physics Through Interaction With an AI Chatbot — an earlier Socratic chatbot evaluated with students
• Measuring Whether LLM Tutors Teach or Solve: A Diagnostic for Educational Impact — measuring whether tutors teach rather than answer
• The Missing Evaluation Axis: What 10,000 Student Submissions Reveal About AI Tutor Effectiveness — behavioral evaluation of AI tutor responses
• Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments — when scaffolding helps and when it should be withdrawn
• A light-touch AI literacy intervention helps protect against AI political persuasion — warnings and persuasion in political AI literacy

Citation

Shi, M., Wang, J., Ni, J., Pal Chowdhury, S., & Sachan, M. (2026). Tackling the Root of Misinformation by Teaching Laypeople about Logical Fallacies via Socratic Questioning and Critical Argumentation. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics.