Automated fact-checking: progress and challenges
FEVER dataset and shared task. Vlachos recaps the creation of FEVER (Fact Extraction and VERification), a milestone 185K-claim benchmark where claims are verified against Wikipedia sentences. The dataset required careful annotation (50 native speakers, claim mutation to avoid bias, separate constructors and verifiers). Inter-annotator agreement was decent (0.68 Fleiss' kappa), and 91% of claims passed final author verification. Baselines achieved ~31% accuracy with evidence retrieval alone, rising to 50–51% when ignoring evidence—showing that retrieving the correct evidence genuinely matters. Over three years, FEVER-track performance climbed from 64% to 75%, demonstrating steady progress but remaining unsolved (one-quarter of instances still misclassified).
Dataset biases and bias correction. Vlachos emphasizes a fundamental problem: the way data is constructed introduces exploitable biases. A common pattern is "hypothesis-only bias"—models learn that negations in a claim correlate with refutation verdicts, so they predict refuted regardless of evidence. Two mitigation strategies: (1) fine-tune with balanced data lacking the bias (akin to "inoculation"), but risk catastrophic forgetting of original task performance; (2) downweight biased instances during training. Vlachos describes using Elastic Weight Consolidation (EWC), a parameter-importance method that penalizes changing important parameters from the original task while allowing less-important parameters to adapt. Results show EWC achieves better trade-off: maintaining ~85% accuracy on the original FEVER while improving on bias data vs. ~74% if fine-tuned without regularization.
FeverTab and multimodal verification. New shared task extending FEVER to use both Wikipedia tables and text; released in July (from talk context, ~2019) with workshop at EMNLP aiming for hybrid event in Punta Cana. Participants can submit approaches for both structured and unstructured evidence retrieval.
Does fact-checking work? The social dimension
Limits of automated verdict labels. Vlachos points out that providing a verdict without engagement ("true/false") doesn't persuade audiences. Voters shown fact-checked speeches of politicians didn't change voting preferences. Backfire-effect research (suggesting corrections entrench false beliefs via conspiracy narratives) was not reproduced in larger studies, but the core finding holds: fact-checking effectiveness depends on trust in the fact-checker. Trust correlates with shared goals and values. On Brexit, publics trusted scientists on many topics but distrusted them on the EU question because they knew scientists' institutional interests favored EU membership.
Fact-checking as a social process. This leads Vlachos to a key insight: misinformation is fundamentally a social phenomenon (exchange of information between people); therefore fact-checking must also be social, embedded in conversation. While social media enables seeing outside our bubble, it also increases polarization—people grow more convinced of their own views despite exposure to alternatives. The question becomes: what conditions enable constructive conversation? Wikipedia's success as an online community suggests conversation can work, but replicating it (e.g., Wiki Tribune) has failed, so we must study what makes Wikipedia function.
Reasoning in groups: psychological foundations
Dual-system reasoning and the card-selection task. Kahneman's dual-system theory posits fast (biased) and slow (rational) reasoning; System 2 is underused. Classic card-selection task (a logical puzzle) sees 10–20% individual success but 80% small-group success—large difference attributable to discussion. Mercier and Sperber's "reasoning as argument" theory proposes that reasoning evolved for communication, not solitary thought—we argue to justify ourselves and are better at judging others' arguments than our own.
Deliberative Bots project. Vlachos and team ran studies with mechanical Turk using the card-selection task. Groups of ~3 people submitted individual answers, then deliberated, then resubmitted. Initial success rate: 12% (or 59% if partial credit for each card). After deliberation: 33% and 70%, respectively. Strikingly, 14% of groups with correct final answers had no participant initially choose it—deliberation created the right answer from all-wrong starting points. This validates group reasoning theory.
Dimensions of deliberation. Analyzing transcripts, coders annotated: moderation, reasoning (especially questioning why), and solution-challenging. The key hypothesis: probing for reasoning predicts success. Findings: probing for reasoning correlates positively but weakly with better outcomes; longer conversation correlates even more weakly; group consensus shows the strongest correlation—even though the task didn't require or incentivize agreement. This aligns with prior research: small groups making decisions together outperform wisdom-of-crowds as long as communication is effective.
From game to Wikipedia: studying real-world disagreement
WikiDisputesData. Vlachos collected 7,425 disagreements from Wikipedia talk pages where editors flag disputes via dispute tags. For each dispute, he gathered conversation history and edit summaries (metadata about changes). Some disputes resolve on their own; others escalate to dispute resolution boards where mediators intervene—treated as a proxy for non-constructive disagreement.
Predicting escalation. Models predict whether a dispute will escalate using features like politeness (gradient capturing how politeness changes over time, not just average), collaboration, toxicity, and sentiment. Neural networks outperform bag-of-words and feature-based methods. A hierarchical attention network modeling utterances → discussion structure helps. Example: a conversation starting angrily (high negativity in early utterance) initially scores high-escalation risk, but as tone improves (politeness increases), the model's prediction flips to non-escalation, and eventually it correctly predicts no escalation. The model calibrates confidence as it sees more utterances, showing healthy epistemic behavior.
Shadow Map project. Joint collaboration with Oxford/Toshiba starting January 2021 to develop chatbots helping users engage with opposing arguments on climate change.
Interactive fact-checking: From classification to conversation
The limitation of classification framing. Vlachos critiques viewing fact-checking as a supervised classification task (predict true/false). This frames it like sentiment analysis or spam filtering—isolated inputs, static labels. But fact-checking is inherently dialogical: understanding why someone holds a belief, presenting relevant information, respecting their autonomy.
Intellectual Turing test. The measure of whether an AI can defend a viewpoint as compellingly as its human proponent. This requires understanding a position's logical structure, its evidence base, and value priorities. A fact-checker must articulate why someone believes something before effectively countering it.
Dialogue-based fact-checking design. Rather than "this claim is false," propose conversational agents that:
- Present relevant information or sources the user may trust
- Ask probing questions ("Have you considered…?")
- Expose users to alternative perspectives without attacking them
- Encourage thinking through implications of beliefs
This differs from recommendation systems (optimizing engagement) and task-oriented bots (providing answers). The goal is expanding understanding while respecting autonomy.
Measurement and challenges. Testing dialogue systems is harder than benchmarking classifiers. Requires post-intervention measurement of whether users' views changed, whether they better understand opposing viewpoints, and whether changes persist. Platform incentives misalign: platforms optimize engagement; a dialogue system focused on thoughtful understanding might not. Institutional embedding requires buy-in from platforms, fact-checking orgs, or media outlets—not just algorithms.
Notes
This talk is valuable for positioning fact-checking at the intersection of technology and social science. Vlachos moves beyond the narrow ML framing (beat the benchmark) to ask: does this actually help? His honest answer—automated fact-checking alone doesn't change minds; trust matters; conversation matters—refocuses the field on mechanisms of belief change and the conditions under which people revise views.
The psychological section (card-selection task, deliberative bots, Wikipedia disputes) provides empirical grounding for why dialogue works: groups reason better together, but only if discussion is substantive and respectful. This motivates the pivot toward interactive systems over pure automation.
A limitation: the dialogue-based vision remains somewhat aspirational. Vlachos acknowledges that building such systems is nascent research and deployment faces institutional barriers (platform alignment, measuring effectiveness, avoiding manipulation). The timeline suggested is "give me a few years," which seems honest given the technical and social complexity.
The talk coheres around a core thesis: fact-checking is a problem of social reasoning, not just information retrieval. Therefore, solutions should facilitate group deliberation and perspective-taking, not replace human judgment with automated verdict systems.