Show Me the Work: Fact-Checkers' Requirements for Explainable Automated Fact-Checking

1 Introduction

• Generative AI adds to the volume and sophistication of online disinformation, while fact-checking projects face insufficient funding and political pressures.A recent report indicates that the number of fact-checking projects worldwide plateaued in recent years and decreased in 2024.Source: §1 · p. 1
• Fact-checking requires expert knowledge and source judgement, while professionals remain sceptical of automation and its usefulness without explanations [62] [85].Prior work suggests explanations and rationale could alleviate this scepticism.Source: §1 · p. 1
• The study asks what fact-checkers consider when explaining decisions, where automated explanations are useful, and how they can meet fact-checkers’ needs.Its contributions include accounts of claim selection, evidence assessment, verdict decisions, tool use, and explanation requirements across the process.Source: §1 · p. 2

2.1 The structure and practice of fact-checking

• Fact-checkers work across claim selection, evidence search, verdict assignment, and communication, while the reasoning behind these decisions remains unclear [62].The study focuses on independent external fact-checkers, although its authors anticipate relevance to internal fact-checkers whose tasks are typically a subset.Source: §2.1 · p. 2

2.2 The state of automated fact-checking

• Automated fact-checking spans claim detection, evidence retrieval, veracity prediction, and explanation generation, but retrieval is often coarse-grained and explanation methods have practical limitations [16] [12] [68] [32] [52] [54] [14] [31] [59].A journalist evaluation found 59% accurate claim classifications and 58% relevant retrieved evidence sentences; a content analysis found 31% of automation-focused papers included retrieval, while 81% relied solely on classification.Source: §2.2 · pp. 2–3

2.3 Explainable AI for fact-checking

• Existing fact-checking explanation criteria were proposed by researchers rather than grounded in the needs of intended users such as fact-checkers [67] [74].Empirical evaluation has largely involved laypeople in a limited selection of Western countries; one study found explanations increased over-reliance on incorrect AI predictions [75].Source: §2.3 · p. 3

3 Method

• The interview guide elicited recent-claim reasoning, tool use at each process step, and explanation needs, with follow-ups adapted to participants and developing themes.The study was a collaboration between NLP and HCI researchers.Source: §3 · pp. 3–4

3.1 Participant recruitment

• Ten professionals were recruited through fact-checking and journalism networks, social media, and an international conference; the sample included participants from five continents.The researchers report reaching theoretical saturation after eight interviews, with two further interviews confirming it.Source: §3.1 · p. 4

3.2 Data collection

• All interviews were conducted remotely via Zoom and lasted approximately 1 hour; with participant permission, they were recorded and transcribed, and participants were compensated with a USD $50 online gift voucher.Source: §3.2 · p. 4

3.3 Data analysis

• Transcript analysis used iterative inductive open, axial, and selective coding, with themes refined collaboratively and mapped to the research questions.The researchers developed a separate theme for fact-checkers’ demand for explanations of AI uncertainty.Source: §3.3 · pp. 4–5, Table 2 · p. 5

3.4 Methodological limitations

• The authors identify a small sample, English-only interviews that may have limited participation, and the subjectivity inherent in self-reported data as methodological limitations.The sample was geographically diverse and included many non-native English speakers.Source: §3.4 · p. 5

4.1.2 Automated tools for claim detection

• Claim-detection tools expedited monitoring and surfaced checkable claims, but cost and reader replicability led some fact-checkers to prefer free, open-source tools.Participants generally paid little attention to how the tools worked; P9 developed a folk theory from patterns in flagged claims.Source: §4.1.2 · pp. 5–6, §4.1.3 · p. 6

4.1.3 Explanation needs for automated claim detection

• For low-stakes claim detection, participants often accepted limited explanations, but selection rationales, bias risks, tool coordination, and language deficiencies remained concerns.Folk theories helped some participants decide when and how to use tools and assess their suggestions.Source: §4.1.3 · p. 6

4.2.1 Decision-making in evidence retrieval

• Fact-checkers prioritized primary, raw evidence and assessed source credibility, relevance, methodology, time frame, and modification, while automated systems often rely on secondary sources and assess documents individually.This mismatch may impede adoption and make it harder for automated systems to retrieve key primary evidence for a verdict.Source: §4.2.1 · pp. 6–7

4.2.2 Automated tools for evidence retrieval

• Fact-checkers used tools to identify sources and extract or retrieve evidence, but reported wrong chatbot answers and language limitations; locally developed language tools could work well.They remained wary of relying on automated outputs without checking evidence relevance and source reliability.Source: §4.2.2 · p. 7

4.2.3 Explanation needs for automated evidence retrieval

• Fact-checkers wanted to verify that cited evidence supported an output and to understand why sources or specific text or image parts were deemed relevant; systems should also identify and explain missing information.When evidence was inaccessible, fact-checkers sought expert input, a gap automated systems would need to flag for follow-up.Source: §4.2.3 · p. 8

4.3.1 Decision-making in veracity prediction

• Verdicts rested on evidence but were shaped by collaborative review and could benefit from nuanced wording, whose appropriate complexity depends on claim, context, and audience.Participants noted that categorical labels are clear but can flatten nuance, while longer verdicts can specify which parts of a claim are true or false.Source: §4.3.1 · p. 8

4.3.2 Automated tools for veracity prediction

• Fact-checkers doubted automated veracity prediction for complex, culturally situated claims, while systems prioritized prediction over documenting how a verdict was reached.Concerns included hallucination, bias, and shortcut learning rather than useful features.Source: §4.3.2 · p. 9

4.3.3 Using AI tools for detecting AI manipulation

• Fact-checkers used AI-manipulation detectors and triangulated outputs, but mismatched confidence and opaque reasoning undermined trust as manipulated media improved.Some found confidence scores reassuring, yet participants wanted explainable tools they could trust and use without misleading readers.Source: §4.3.3 · p. 9

4.3.3 Using AI tools for detecting AI manipulation.

• Fact-checkers increasingly rely on AI tools yet find them hard to understand and unreliable, and seek explanations of why and how systems reach verdicts and who created them.Understanding when tools should or should not be relied upon remains important even as their accuracy may improve.Source: §4.3.3 · p. 9, §4.3.4 · p. 9

4.3.4 Explanation needs for automated veracity prediction.

• Explanations should show how evidence supports a verdict and how the model searched and selected evidence, using human fact-checking terms rather than overly technical accounts.Participants also wanted uncertainty explained through source reliability and evidence limitations, not just a percentage.Source: §4.3.4 · pp. 9–10
• Training-data origins, volume, and languages help fact-checkers identify bias and information gaps and judge when tools may fail, including across non-English contexts and African accents.Participants viewed disclosure of training data as important for calibrating when to rely on a model.Source: §4.3.4 · p. 10

4.4.1 Decision-making in explaining and communicating fact-checks.

• Fact-checking explanations should enable readers to reproduce the process by identifying evidence sources, how they were checked, and why they were selected.Participants described replicability as central to “showing the work,” beyond simply giving reasons for a verdict.Source: §4.4.1 · p. 11
• Verdict type and claim complexity shape explanations: partially true or unverifiable claims need context, while complex stories call for clear structure despite research often being nonlinear.Fact-checkers reported laying out evidence for unverifiable claims and organizing complex explanations into a logical path for audiences with unknown background knowledge.Source: §4.4.1 · pp. 11–12

4.4.2 Automated tools for communicating fact-checks.

• Fact-checkers use LLM chatbots to edit, summarize, and disseminate work, but raised concerns about journalists’ work being used for training without permission and stigma around tool use.They also used ChatGPT to cross-check arguments and draft scripts, while emphasizing that sources still need verification.Source: §4.4.2 · p. 12

4.4.3 Explanation needs for fact-checkers.

• Fact-checkers need to verify model decisions using specific, checkable evidence; numerical confidence scores alone can be unhelpful and do not measure evidence quality.Local explanations can also make decisions easier to assess for journalists with limited technical expertise.Source: §4.4.3 · pp. 12–13
• Fact-checkers differed on whether explanations should prioritize easy-to-understand narratives or faithful accounts of model decisions, so both qualities must be considered across audiences and contexts.Some worried that a large gap between an explanation and the model’s actual process could make it unreliable.Source: §4.4.3 · p. 13

5 Discussion

• The discussion identifies critical gaps between automated tools and fact-checkers’ practices, framing explanation needs as part of a broader mismatch in goals and capabilities.It calls for interdisciplinary and human-centred work to address that mismatch.Source: §5 · p. 13, §5.1 · p. 13, §5.2 · p. 14

5.1 Fact-checker explanation needs versus automated fact-checking capabilities

• Claim detection needs relatively few explanations, evidence retrieval needs reliable relevance and credibility checks, and veracity prediction calls for detailed links between evidence and verdicts.Existing evidence highlights and saliency maps raise unresolved concerns about faithfulness and legitimacy; feature scores and evidence summaries for veracity prediction also face stability and reliability concerns.Source: §5.1 · p. 13
• Simple percentage confidence scores do not help fact-checkers relate uncertainty to evidence quality or source reliability, and LLMs may overestimate their confidence.Explanations also need to reflect contextual nuance, including whether facts are static, temporal, or dynamic.Source: §5.1 · pp. 13–14
• Fact-checkers want replicable explanations that trace model processes, connect specific evidence to verdicts, and expose uncertainty and information gaps.Current multi-hop and chain-of-thought approaches still fall short of the complex logic required for real-world claims.Source: §5.1 · p. 14

5.2.1 The importance of primary sources.

• Automated systems often rely on secondary sources, while fact-checkers value primary evidence and require tools to assess evidence quality, relevance, and reliability.Secondary sources may be biased or only indirectly relevant, while primary evidence can be messy, unstructured, missing, or incomplete.Source: §5.2.1 · p. 14

5.2.2 Beyond binary verdicts.

• Categorical verdicts can flatten nuanced claims, while fact-checkers regard documenting the pathway and inferences to a verdict as central to credibility.Current systems often prioritize the verdict through token highlighting, rules, or evidence summaries; they should explain their processes so fact-checkers can verify them.Source: §5.2.2 · p. 14, §5.2.3 · pp. 14–15

5.3.1 Future work in explainable automated fact-checking.

• Future HCI work can examine explanations for claim filtering, information extraction, and source evaluation, while automated veracity prediction must better align with fact-checker reasoning.Neuro-symbolic approaches have followed human logic only in highly controlled settings with artificially created datasets, and useful ways to present uncertainty sources and resolution remain unexplored.Source: §5.3.1 · p. 15

5.3.2 Ethical considerations and broader implications of automated fact-checking.

• Reliance on automated fact-checking may increase public scepticism and erode trust, while algorithmic bias and weak local grounding risk reproducing inequalities and silencing minority voices.Local knowledge and communal practices are key to successful claim verification, yet system development often relies on researcher imaginaries rather than actual fact-checker practice.Source: §5.3.2 · p. 15

6 Conclusion

• Current automated fact-checking tools have critical gaps between their capabilities and fact-checkers’ goals and needs.The paper outlines information requirements at each process step as a route to making tools more useful.Source: §6 · p. 15
• Effective tools require design grounded in fact-checkers’ decision processes and explanation needs, alongside meaningful engagement and interdisciplinary research.The conclusion calls for systems that support and empower fact-checkers rather than fully automated fact-checking.Source: §6 · p. 15
• Fully automated fact-checking systems may be easier to subvert as political conditions shift, given that fact-checking practice is targeted for ideological and political reasons.The conclusion frames support for fact-checkers as preferable to seeking full automation.Source: §6 · p. 15

A Pre-Interview Questionnaire

• Participants are asked which organisation they work for, their role, and how long they have worked as fact-checkers.The questionnaire separately asks how long they have worked for their current organisation and in their current role.Source: A Pre-Interview Questionnaire · p. 19
• The questionnaire asks whether participants have used AI tools in fact-checking and which tools they have used..Source: A Pre-Interview Questionnaire · p. 19
• Participants are asked their current country and gender..Source: A Pre-Interview Questionnaire · p. 19

1. Introductory questions

• Participants are asked to describe their job and the work they do..Source: §1 · p. 19
• Participants are asked to describe a claim they checked recently and the process and steps involved..Source: §1 · p. 19
• A backup prompt asks what specific goals participants have at each step of the process..Source: §1 · p. 19

• Interviewees are asked how challenging fact-checking claims is and what the process involves.Source: §1 · p. 20
• The prompts frame claim difficulty as a topic for describing both the work involved and factors that may make it harder.Source: §1 · p. 20, C Example themes and codes · p. 20
• A backup prompt asks whether any factors make some claims particularly challenging to fact-check.Source: C Example themes and codes · p. 20

• Interviewees are asked to describe how they searched for and selected evidence for a claim.Source: §2 · p. 20
• They are asked how they decided which veracity label to assign to a claim.Source: §2 · p. 20
• They are asked which technological tools they use to verify claims and what advantages and disadvantages those tools have.Source: §2 · p. 20

• Interviewees are asked how they would explain a claim's assigned label and what challenges arise in explaining the decision.Source: §3 · p. 20
• They are asked whether claim-specific factors and audience knowledge affect what information they present, including differences between editors and readers.Source: §3 · p. 20
• They are asked what transparency means in fact-checking, how their organisation achieves it, and whether tools help produce decision explanations.Source: §3 · p. 20

• Interviewees are asked whether they know or have used AI-support tools, what the tools do, and their advantages and disadvantages.Source: §4 · p. 20
• They are asked which fact-checking steps AI tools could support and what information would help them use and check such assistance.Source: §4 · p. 20
• They are asked whether explanations of a system's general workings would be useful and what criteria define a useful explainable fact-checking system.Source: §4 · p. 20

5. Closing questions

• Participants were asked whether anything important or relevant remained undiscussed.A follow-up prompt also invited them to share anything else.Source: §5 · p. 20
• Table 3 is captioned as giving examples of themes, subthemes, and codes developed from transcript analysis.Source: Table 3 · p. 21