---
kind: concept
slug: automated-fact-checking
title: Automated fact-checking
summary: >-
  Checking a claim against retrieved evidence and reporting the result with sources. Two published routes: an LLM with live web search that writes sourced replies to public posts, and a retrieve, select-rationale, label pipeline over fixed scientific abstracts. Author-reported results only. Trade-off: coverage and speed against evidence quality, cost and domain scope.
tags: [fact-checking, misinformation, retrieval, multimodal, social-media, scientific-literature]
papers: ["1806.07687", "2403.11169", "2004.14974", "2512.12413"]
related: [retrieval-augmented-generation, rationale-extraction, domain-adaptation, hallucination-mitigation, hallucination-detection]
status: draft
updated_at: 2026-11-01
---

## Overview

The problem is that a claim often needs checking against evidence the reader
cannot see, and a bare verdict is hard to trust. Crowd notes on X were, in the
authors' February 2023 snapshot of the 2403 paper, absent for 88% of suspicious
posts and without a highly rated note in the first hour for 93%
[arxiv:2403.11169 §Introduction]. Scientific claims raise a different
difficulty: checking them needs domain experts and a corpus of specialist
literature, and general fact-checking datasets are built on news and
Wikipedia [arxiv:2004.14974 §1] [arxiv:2004.14974 §2].

The shared idea is to let a model do the research step by step: find
candidate evidence, pick the passages that matter, and produce a verdict or a
reply that links them [arxiv:2403.11169 §Approach] [arxiv:2004.14974 §5]. The
shared invariant is that every factual statement in the output is traceable to
a source that the system actually retrieved. Both papers evaluate with
human judgements or expert labels rather than lexical overlap. The first
paper argues that ROUGE cannot judge whether a correction is factual,
complete and sourced [arxiv:2403.11169 §Introduction].

## Task framing

Before choosing a route, fix the task. A 2018 survey sorts fact-checking
work by three choices: the input (a claim sentence, a knowledge-base triple or
a whole document), the output (a binary label, an ordinal or multiclass label,
a score, or a label with supporting sentences) and the evidence type (none,
metadata, knowledge graph, text, previously checked claims or social-network
signals) [arxiv:1806.07687 §3.1] [arxiv:1806.07687 §3.2] [arxiv:1806.07687 §3.3].
Evidence is the axis that decides what a system can output and how it must be
evaluated [arxiv:1806.07687 §1]. The survey reports no experiments of its own,
so it gives no performance figures [arxiv:1806.07687 §4]. Its FEVER-style
pipeline (retrieve documents, select sentences, classify against them) has the
same shape as the scientific route below [arxiv:1806.07687 §5].

A caution from the same survey: a method that assumes the evidence is given,
such as an entailment model, does not solve the retrieval step on its own
[arxiv:1806.07687 §5]. The survey also notes that most surveyed datasets are
small or tied to one source, which limits training [arxiv:1806.07687 §4]. The
survey dates from 2018 and predates the LLM routes below, so its method
descriptions are historical and cannot say which later approaches were adopted.
Card: [Automated Fact Checking: Task formulations, methods and future directions](/papers/1806.07687).

## Variants

There are two published routes, plus the image path of the first, and a
measure of the human reader that sits beside them.

### Public posts: retrieval-grounded reply

The pipeline is: queries, a search per publisher tier, a time filter, a
relevance filter, evidence extraction, and a reply [arxiv:2403.11169
§Approach]. Publisher tiers (High, Medium, Low) come from Media Bias/Fact
Check factuality and bias ratings, and the tiers are searched in that order
[arxiv:2403.11169 §Experiments]. Only pages published before the post are
used, which simulates a correction written at posting time
[arxiv:2403.11169 §Experiments]. Author-reported: 12 expert fact-checkers
rated replies 8.1/10 overall, against 6.3 for high-helpfulness Community
Notes and 5.9 for GPT-4 without retrieval [arxiv:2403.11169 §Results].
Card: [Customized large language models can outperform Community Notes in
correcting misinformation](/papers/2403.11169).

#### Image posts: text description plus image matching

Each image is first turned into text: a caption, celebrity labels and OCR
are combined by the LLM into an "informative image description"
[arxiv:2403.11169 §Approach] [arxiv:2403.11169 §Experiments]. The search
then also uses reverse image search, and candidate pages are scored on text
and on image similarity [arxiv:2403.11169 §Approach]. Removing retrieval or
the full system scored 25% and 33% higher, in relative terms, than the
versions without retrieval and without vision [arxiv:2403.11169 §Results].

The released code differs from the paper on this path, notably in how the
image captions are revised and how pages are kept. See the card's Paper vs
code section before reusing any threshold.

### Scientific claims: retrieve, select rationale, label

The task takes a claim and a corpus of abstracts. The system returns the
abstracts that support or refute the claim, and for each one a label from
SUPPORTS, REFUTES or NOT_ENOUGH_INFO plus the sentences that justify it
[arxiv:2004.14974 §4]. The benchmark, SciFact, has 1,409 expert-written claims
checked against 5,183 abstracts, split 809 / 300 / 300 for train, dev and test
[arxiv:2004.14974 §3] [arxiv:2004.14974 tab. 2].

The baseline, VeriSci, runs three stages: TF-IDF retrieval of the top three
abstracts, RoBERTa-large selection of rationale sentences, and RoBERTa-large
label prediction from the selected sentences and the claim
[arxiv:2004.14974 §5] [arxiv:2004.14974 app. A.1]. Author-reported open-setting
test F1 is 39.5 for sentence selection with label and 46.5 for abstract label
with rationale [arxiv:2004.14974 tab. 4]. Pre-training on FEVER before SciFact
raised dev label accuracy from 75.7 to 81.9 [arxiv:2004.14974 tab. 3], and the
authors report relative gains of 47% and 28% over the FEVER-only model on those
two metrics [arxiv:2004.14974 §6.2]. The same table shows that an oracle
rationale selector reaches 63.2 on the sentence metric in the open setting, so
the authors locate much of the gap in the earlier stages [arxiv:2004.14974 tab. 4]
[arxiv:2004.14974 §6.2].

Card: [Fact or Fiction: Verifying Scientific Claims](/papers/2004.14974).

### Human readers: measuring verification habits

This is not an automated system. It is a 13-item, 5-point self-report scale of
"critical thinking in AI use", with subscales for verification, motivation to
understand AI and reflection on responsible AI [arxiv:2512.12413 tab. 1]. It
fits when the question is how well people check AI output, for example to
measure the human side of a deployment. Total Cronbach's α was .88 in the
authors' first validation sample (N = 376) [arxiv:2512.12413 §4.1.2].

The link to checking behaviour rests on one study. In Study 6 (N = 191),
participants judged 12 statements through a GPT-4o chatbot, and the scale
predicted their veracity accuracy with b = 6.91 and R² = .06, a small share
of variance, although the authors' discussion calls it a moderate effect
[arxiv:2512.12413 §7.2] [arxiv:2512.12413 §8].
In Study 3 the scale had no reliable link with a cognitive reflection test
[arxiv:2512.12413 §4.2.2]. All results are author-reported and self-reported.
Card: [Understanding Critical Thinking in Generative Artificial Intelligence
Use](/papers/2512.12413).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Claims must link to checkable sources | Retrieval-grounded reply | GPT-4 without retrieval had 49% page-not-found links; the full system's references were nearly all reachable and 96% relevant [arxiv:2403.11169 §Results] |
| Fewer moving parts, text posts | Text-only path | Omits the image steps: captioning, reverse image search and image matching [arxiv:2403.11169 §Approach] |
| Posts include images | Image path, with vision | Full system 33% higher than the no-vision version, relative, in the authors' ablation [arxiv:2403.11169 §Results] |
| Claim checked against a fixed, specialist corpus with labels | Retrieve, select rationale, label | Benchmark and baseline with per-abstract labels and rationales [arxiv:2004.14974 §4] [arxiv:2004.14974 §5] |
| Little labelled data in the target domain | Pre-train on a general fact-checking set first | Dev label accuracy 81.9 with FEVER then SciFact, against 75.7 on SciFact alone [arxiv:2004.14974 tab. 3] |
| Sparse or breaking evidence | Do not rely on the reply alone | The 2403 authors did not measure abstention or information voids [arxiv:2403.11169 §Discussion] |
| Non-English, video, other platforms or other domains | Evaluate first | Only English text and images on one platform were studied [arxiv:2403.11169 §Discussion]; SciFact is English scientific text [arxiv:2004.14974 §3] |
| Medical decisions | Do not use | The authors call the system a research prototype and say it must not be used for medical decisions [arxiv:2004.14974 §1] |
| Human review is available | Draft and review | Expert agreement was only fair for most criteria, weighted κ mostly 0.28 to 0.51 [arxiv:2403.11169 §Experiments]; SciFact annotator agreement was 0.75 for labels [arxiv:2004.14974 §3.3] |
| Need to measure how well people check AI output | Human verification scale, not a system | 13-item self-report scale; criterion evidence comes from one study and explains R² = .06 of accuracy [arxiv:2512.12413 §4.1.2] [arxiv:2512.12413 §7.2] |
| Choosing the task definition or label set before building | Use the survey's three axes | Input, output and evidence type are compared across prior work [arxiv:1806.07687 §3] |
| Claims that repeat or paraphrase earlier checked claims | Match to already-checked claims | The survey says this covers only repeated or paraphrased claims [arxiv:1806.07687 §5] |
| Cost per item matters | Measure your own | The authors estimate about 0.2 USD and about two minutes per post on an M1 CPU, with no cost reported for the baselines [arxiv:2403.11169 §Discussion] |

Reader effect: in a survey of 988 U.S. participants, the correct belief
that a post was misleading rose from 4.5 to 4.9, a 10% gain, as
reported by the authors [arxiv:2403.11169 §Results]. Source disclosure
(human or AI) had no significant effect on trust or sharing
[arxiv:2403.11169 §Results].

Before reusing the scientific numbers, note the spread of the evaluation.
Bootstrap standard deviations on the 300-claim test set are 2 to 4 F1 points
[arxiv:2004.14974 tab. 4], and the COVID-19 case study is one annotator judging
36 claims, 23 of them plausible [arxiv:2004.14974 §6.3].

## Implementation sketch

Language-agnostic pseudocode of the text-only core of the first route. Steps
4 and 5 are where the variants differ. The paper's stop rule (two refuting
pages) is used here, and the code's item count is noted in the card.

The scientific route follows the three stages in the Variants section. Its
pseudocode is on the [card](/papers/2004.14974), under Method.

```pseudocode
# Retrieval-grounded correction, text-only post
# Input:  post text, post time
# Output: reply with source links

def correct_post(post):
    # 1. LLM writes search queries; no query if the post is uninformative
    #    [arxiv:2403.11169 §Approach]
    queries = llm_queries(post)
    refuting_pages = 0
    evidence = []
    # 2. Search publisher tiers from most to least credible
    #    [arxiv:2403.11169 §Experiments]
    for tier in ["High", "Medium", "Low"]:
        pages = search_and_crawl(queries, tier)
        # 3. Keep only pages dated before the post, similar to its text
        #    [arxiv:2403.11169 §Experiments]
        pages = [p for p in pages if p.date < post.time
                 and text_similarity(post, p) >= threshold]
        for page in pages:
            # 4. LLM quotes explicit refutations and context passages
            #    [arxiv:2403.11169 §Approach]
            refutes, context = llm_extract_evidence(post, page)
            evidence += refutes + context
            if refutes:
                refuting_pages += 1
        # 5. Stop after two refuting pages, the paper's rule
        #    [arxiv:2403.11169 §Approach]
        if refuting_pages >= 2:
            break
    # 6. Reply with cited facts; ranking by tier is the paper's intent
    #    [arxiv:2403.11169 §Approach]
    return llm_reply(post, evidence)
```

## Contradictions

No conflicting numbers between the papers. The 2018 survey reports no results
of its own, only dataset sizes [arxiv:1806.07687 §4], and the 2512 paper is a
self-report scale, not a checker. The other two results measure different things: the first compares expert ratings of replies with
Community Notes and GPT-4 on public posts, and the second reports F1 on a
claim-abstract benchmark. The numbers should not be compared across the two
routes. Within the first paper, its text and its released code differ on the
stop rule, on how evidence is ordered, and on the image path. These
differences are listed on the [card](/papers/2403.11169), which should be
read before any reuse. The second paper's text and code also differ, on
retrieval details and on label training; see its Paper vs code section.

## Papers

- [Automated Fact Checking: Task formulations, methods and future directions](/papers/1806.07687): a survey that unifies task definitions by input, output and evidence type, and reviews datasets and methods; it reports no experiments of its own.
- [Customized large language models can outperform Community Notes in correcting misinformation](/papers/2403.11169): an LLM with tiered web retrieval and image matching writes sourced replies to posts, and is rated above crowd notes and plain GPT-4 by experts (author-reported).
- [Fact or Fiction: Verifying Scientific Claims](/papers/2004.14974): SciFact benchmark and the VeriSci baseline, which retrieves abstracts, selects rationale sentences and predicts a label for each (author-reported open-setting F1 of 39.5 and 46.5).
- [Understanding Critical Thinking in Generative Artificial Intelligence Use: Development, Validation, and Correlates of the Critical Thinking in AI Use Scale](/papers/2512.12413): a 13-item self-report scale of how people verify and reflect on AI output, validated over six studies; it is a human measure, not a fact-checking system (author-reported).